From bb57f30c2de46cfca5419ad01738a41705f96cc3 Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Fri, 14 Oct 2022 10:56:41 +0200 Subject: [PATCH 01/82] init --- README.md | 1 + aesthetic_embeddings/insert_embs_here.txt | 0 modules/processing.py | 17 +++- modules/sd_hijack.py | 80 ++++++++++++++++++- modules/shared.py | 5 ++ modules/textual_inversion/dataset.py | 2 +- .../textual_inversion/textual_inversion.py | 35 +++++--- modules/txt2img.py | 11 ++- modules/ui.py | 59 +++++++++----- 9 files changed, 172 insertions(+), 38 deletions(-) create mode 100644 aesthetic_embeddings/insert_embs_here.txt diff --git a/README.md b/README.md index 859a91b6f..7b8d018b5 100644 --- a/README.md +++ b/README.md @@ -70,6 +70,7 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web - No token limit for prompts (original stable diffusion lets you use up to 75 tokens) - DeepDanbooru integration, creates danbooru style tags for anime prompts (add --deepdanbooru to commandline args) - [xformers](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Xformers), major speed increase for select cards: (add --xformers to commandline args) +- Aesthetic, a way to generate images with a specific aesthetic by using clip images embds (implementation of https://github.com/vicgalle/stable-diffusion-aesthetic-gradients) ## Installation and Running Make sure the required [dependencies](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Dependencies) are met and follow the instructions available for both [NVidia](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs) (recommended) and [AMD](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs) GPUs. diff --git a/aesthetic_embeddings/insert_embs_here.txt b/aesthetic_embeddings/insert_embs_here.txt new file mode 100644 index 000000000..e69de29bb diff --git a/modules/processing.py b/modules/processing.py index d5172f008..9a0337591 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -316,11 +316,16 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration return f"{all_prompts[index]}{negative_prompt_text}\n{generation_params_text}".strip() -def process_images(p: StableDiffusionProcessing) -> Processed: +def process_images(p: StableDiffusionProcessing, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, + aesthetic_imgs=None,aesthetic_slerp=False) -> Processed: """this is the main loop that both txt2img and img2img use; it calls func_init once inside all the scopes and func_sample once per batch""" + aesthetic_lr = float(aesthetic_lr) + aesthetic_weight = float(aesthetic_weight) + aesthetic_steps = int(aesthetic_steps) + if type(p.prompt) == list: - assert(len(p.prompt) > 0) + assert (len(p.prompt) > 0) else: assert p.prompt is not None @@ -394,7 +399,13 @@ def process_images(p: StableDiffusionProcessing) -> Processed: #uc = p.sd_model.get_learned_conditioning(len(prompts) * [p.negative_prompt]) #c = p.sd_model.get_learned_conditioning(prompts) with devices.autocast(): - uc = prompt_parser.get_learned_conditioning(shared.sd_model, len(prompts) * [p.negative_prompt], p.steps) + if hasattr(shared.sd_model.cond_stage_model, "set_aesthetic_params"): + shared.sd_model.cond_stage_model.set_aesthetic_params(0, 0, 0) + uc = prompt_parser.get_learned_conditioning(shared.sd_model, len(prompts) * [p.negative_prompt], + p.steps) + if hasattr(shared.sd_model.cond_stage_model, "set_aesthetic_params"): + shared.sd_model.cond_stage_model.set_aesthetic_params(aesthetic_lr, aesthetic_weight, + aesthetic_steps, aesthetic_imgs,aesthetic_slerp) c = prompt_parser.get_multicond_learned_conditioning(shared.sd_model, prompts, p.steps) if len(model_hijack.comments) > 0: diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py index c81722a0a..6d5196fea 100644 --- a/modules/sd_hijack.py +++ b/modules/sd_hijack.py @@ -9,11 +9,14 @@ from torch.nn.functional import silu import modules.textual_inversion.textual_inversion from modules import prompt_parser, devices, sd_hijack_optimizations, shared -from modules.shared import opts, device, cmd_opts +from modules.shared import opts, device, cmd_opts, aesthetic_embeddings from modules.sd_hijack_optimizations import invokeAI_mps_available import ldm.modules.attention import ldm.modules.diffusionmodules.model +from transformers import CLIPVisionModel, CLIPModel +import torch.optim as optim +import copy attention_CrossAttention_forward = ldm.modules.attention.CrossAttention.forward diffusionmodules_model_nonlinearity = ldm.modules.diffusionmodules.model.nonlinearity @@ -109,13 +112,29 @@ class StableDiffusionModelHijack: _, remade_batch_tokens, _, _, _, token_count = self.clip.process_text([text]) return remade_batch_tokens[0], token_count, get_target_prompt_token_count(token_count) +def slerp(low, high, val): + low_norm = low/torch.norm(low, dim=1, keepdim=True) + high_norm = high/torch.norm(high, dim=1, keepdim=True) + omega = torch.acos((low_norm*high_norm).sum(1)) + so = torch.sin(omega) + res = (torch.sin((1.0-val)*omega)/so).unsqueeze(1)*low + (torch.sin(val*omega)/so).unsqueeze(1) * high + return res class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): def __init__(self, wrapped, hijack): super().__init__() self.wrapped = wrapped + self.clipModel = CLIPModel.from_pretrained( + self.wrapped.transformer.name_or_path + ) + del self.clipModel.vision_model self.hijack: StableDiffusionModelHijack = hijack self.tokenizer = wrapped.tokenizer + # self.vision = CLIPVisionModel.from_pretrained(self.wrapped.transformer.name_or_path).eval() + self.image_embs_name = None + self.image_embs = None + self.load_image_embs(None) + self.token_mults = {} self.comma_token = [v for k, v in self.tokenizer.get_vocab().items() if k == ','][0] @@ -136,6 +155,23 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): if mult != 1.0: self.token_mults[ident] = mult + def set_aesthetic_params(self, aesthetic_lr, aesthetic_weight, aesthetic_steps, image_embs_name=None, + aesthetic_slerp=True): + self.slerp = aesthetic_slerp + self.aesthetic_lr = aesthetic_lr + self.aesthetic_weight = aesthetic_weight + self.aesthetic_steps = aesthetic_steps + self.load_image_embs(image_embs_name) + + def load_image_embs(self, image_embs_name): + if image_embs_name is None or len(image_embs_name) == 0: + image_embs_name = None + if image_embs_name is not None and self.image_embs_name != image_embs_name: + self.image_embs_name = image_embs_name + self.image_embs = torch.load(aesthetic_embeddings[self.image_embs_name], map_location=device) + self.image_embs /= self.image_embs.norm(dim=-1, keepdim=True) + self.image_embs.requires_grad_(False) + def tokenize_line(self, line, used_custom_terms, hijack_comments): id_end = self.wrapped.tokenizer.eos_token_id @@ -333,7 +369,47 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): z1 = self.process_tokens(tokens, multipliers) z = z1 if z is None else torch.cat((z, z1), axis=-2) - + + if len(text[ + 0]) != 0 and self.aesthetic_steps != 0 and self.aesthetic_lr != 0 and self.aesthetic_weight != 0 and self.image_embs_name != None: + if not opts.use_old_emphasis_implementation: + remade_batch_tokens = [ + [self.wrapped.tokenizer.bos_token_id] + x[:75] + [self.wrapped.tokenizer.eos_token_id] for x in + remade_batch_tokens] + + tokens = torch.asarray(remade_batch_tokens).to(device) + with torch.enable_grad(): + model = copy.deepcopy(self.clipModel).to(device) + model.requires_grad_(True) + + # We optimize the model to maximize the similarity + optimizer = optim.Adam( + model.text_model.parameters(), lr=self.aesthetic_lr + ) + + for i in range(self.aesthetic_steps): + text_embs = model.get_text_features(input_ids=tokens) + text_embs = text_embs / text_embs.norm(dim=-1, keepdim=True) + sim = text_embs @ self.image_embs.T + loss = -sim + optimizer.zero_grad() + loss.mean().backward() + optimizer.step() + + zn = model.text_model(input_ids=tokens, output_hidden_states=-opts.CLIP_stop_at_last_layers) + if opts.CLIP_stop_at_last_layers > 1: + zn = zn.hidden_states[-opts.CLIP_stop_at_last_layers] + zn = model.text_model.final_layer_norm(zn) + else: + zn = zn.last_hidden_state + model.cpu() + del model + + if self.slerp: + z = slerp(z, zn, self.aesthetic_weight) + else: + z = z * (1 - self.aesthetic_weight) + zn * self.aesthetic_weight + remade_batch_tokens = rem_tokens batch_multipliers = rem_multipliers i += 1 diff --git a/modules/shared.py b/modules/shared.py index 5901e6056..cf13a10d9 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -30,6 +30,8 @@ parser.add_argument("--no-half-vae", action='store_true', help="do not switch th parser.add_argument("--no-progressbar-hiding", action='store_true', help="do not hide progressbar in gradio UI (we hide it because it slows down ML if you have hardware acceleration in browser)") parser.add_argument("--max-batch-count", type=int, default=16, help="maximum batch count value for the UI") parser.add_argument("--embeddings-dir", type=str, default=os.path.join(script_path, 'embeddings'), help="embeddings directory for textual inversion (default: embeddings)") +parser.add_argument("--aesthetic_embeddings-dir", type=str, default=os.path.join(script_path, 'aesthetic_embeddings'), + help="aesthetic_embeddings directory(default: aesthetic_embeddings)") parser.add_argument("--hypernetwork-dir", type=str, default=os.path.join(models_path, 'hypernetworks'), help="hypernetwork directory") parser.add_argument("--allow-code", action='store_true', help="allow custom script execution from webui") parser.add_argument("--medvram", action='store_true', help="enable stable diffusion model optimizations for sacrificing a little speed for low VRM usage") @@ -90,6 +92,9 @@ os.makedirs(cmd_opts.hypernetwork_dir, exist_ok=True) hypernetworks = hypernetwork.list_hypernetworks(cmd_opts.hypernetwork_dir) loaded_hypernetwork = None +aesthetic_embeddings = {f.replace(".pt",""): os.path.join(cmd_opts.aesthetic_embeddings_dir, f) for f in + os.listdir(cmd_opts.aesthetic_embeddings_dir) if f.endswith(".pt")} + def reload_hypernetworks(): global hypernetworks diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 67e90afeb..59b2b0219 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -48,7 +48,7 @@ class PersonalizedBase(Dataset): print("Preparing dataset...") for path in tqdm.tqdm(self.image_paths): try: - image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.BICUBIC) + image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.Resampling.BICUBIC) except Exception: continue diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index fa0e33a2a..b12a8e6d8 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -172,7 +172,15 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn -def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_image_prompt): +def batched(dataset, total, n=1): + for ndx in range(0, total, n): + yield [dataset.__getitem__(i) for i in range(ndx, min(ndx + n, total))] + + +def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, + create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, + preview_image_prompt, batch_size=1, + gradient_accumulation=1): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -204,7 +212,11 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..." with torch.autocast("cuda"): - ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file) + ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, + height=training_height, + repeats=shared.opts.training_image_repeats_per_epoch, + placeholder_token=embedding_name, model=shared.sd_model, + device=devices.device, template_file=template_file) hijack = sd_hijack.model_hijack @@ -223,7 +235,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) - pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) + pbar = tqdm.tqdm(enumerate(batched(ds, steps - ititial_step, batch_size)), total=steps - ititial_step) for i, entry in pbar: embedding.step = i + ititial_step @@ -235,17 +247,20 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini break with torch.autocast("cuda"): - c = cond_model([entry.cond_text]) + c = cond_model([e.cond_text for e in entry]) + + x = torch.stack([e.latent for e in entry]).to(devices.device) + loss = shared.sd_model(x, c)[0] - x = entry.latent.to(devices.device) - loss = shared.sd_model(x.unsqueeze(0), c)[0] del x losses[embedding.step % losses.shape[0]] = loss.item() - optimizer.zero_grad() loss.backward() - optimizer.step() + if ((i + 1) % gradient_accumulation == 0) or (i + 1 == steps - ititial_step): + optimizer.step() + optimizer.zero_grad() + epoch_num = embedding.step // len(ds) epoch_step = embedding.step - (epoch_num * len(ds)) + 1 @@ -259,7 +274,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini if embedding.step > 0 and images_dir is not None and embedding.step % create_image_every == 0: last_saved_image = os.path.join(images_dir, f'{embedding_name}-{embedding.step}.png') - preview_text = entry.cond_text if preview_image_prompt == "" else preview_image_prompt + preview_text = entry[0].cond_text if preview_image_prompt == "" else preview_image_prompt p = processing.StableDiffusionProcessingTxt2Img( sd_model=shared.sd_model, @@ -305,7 +320,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini

Loss: {losses.mean():.7f}
Step: {embedding.step}
-Last prompt: {html.escape(entry.cond_text)}
+Last prompt: {html.escape(entry[-1].cond_text)}
Last saved embedding: {html.escape(last_saved_file)}
Last saved image: {html.escape(last_saved_image)}

diff --git a/modules/txt2img.py b/modules/txt2img.py index e985242b3..783420243 100644 --- a/modules/txt2img.py +++ b/modules/txt2img.py @@ -6,7 +6,14 @@ import modules.processing as processing from modules.ui import plaintext_to_html -def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: str, steps: int, sampler_index: int, restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, seed: int, subseed: int, subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool, height: int, width: int, enable_hr: bool, scale_latent: bool, denoising_strength: float, *args): +def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: str, steps: int, sampler_index: int, + restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, seed: int, subseed: int, + subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool, + height: int, width: int, enable_hr: bool, scale_latent: bool, denoising_strength: float, + aesthetic_lr=0, + aesthetic_weight=0, aesthetic_steps=0, + aesthetic_imgs=None, + aesthetic_slerp=False, *args): p = StableDiffusionProcessingTxt2Img( sd_model=shared.sd_model, outpath_samples=opts.outdir_samples or opts.outdir_txt2img_samples, @@ -40,7 +47,7 @@ def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: processed = modules.scripts.scripts_txt2img.run(p, *args) if processed is None: - processed = process_images(p) + processed = process_images(p, aesthetic_lr, aesthetic_weight, aesthetic_steps, aesthetic_imgs, aesthetic_slerp) shared.total_tqdm.clear() diff --git a/modules/ui.py b/modules/ui.py index 220fb80b5..d961d126e 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -24,7 +24,8 @@ import gradio.routes from modules import sd_hijack from modules.paths import script_path -from modules.shared import opts, cmd_opts +from modules.shared import opts, cmd_opts,aesthetic_embeddings + if cmd_opts.deepdanbooru: from modules.deepbooru import get_deepbooru_tags import modules.shared as shared @@ -534,6 +535,14 @@ def create_ui(wrap_gradio_gpu_call): width = gr.Slider(minimum=64, maximum=2048, step=64, label="Width", value=512) height = gr.Slider(minimum=64, maximum=2048, step=64, label="Height", value=512) + with gr.Group(): + aesthetic_lr = gr.Textbox(label='Learning rate', placeholder="Learning rate", value="0.005") + aesthetic_weight = gr.Slider(minimum=0, maximum=1, step=0.01, label="Aesthetic weight", value=0.7) + aesthetic_steps = gr.Slider(minimum=0, maximum=50, step=1, label="Aesthetic steps", value=50) + + aesthetic_imgs = gr.Dropdown(sorted(aesthetic_embeddings.keys()), label="Imgs embedding", value=sorted(aesthetic_embeddings.keys())[0] if len(aesthetic_embeddings) > 0 else None) + aesthetic_slerp = gr.Checkbox(label="Slerp interpolation", value=False) + with gr.Row(): restore_faces = gr.Checkbox(label='Restore faces', value=False, visible=len(shared.face_restorers) > 1) tiling = gr.Checkbox(label='Tiling', value=False) @@ -586,25 +595,30 @@ def create_ui(wrap_gradio_gpu_call): fn=wrap_gradio_gpu_call(modules.txt2img.txt2img), _js="submit", inputs=[ - txt2img_prompt, - txt2img_negative_prompt, - txt2img_prompt_style, - txt2img_prompt_style2, - steps, - sampler_index, - restore_faces, - tiling, - batch_count, - batch_size, - cfg_scale, - seed, - subseed, subseed_strength, seed_resize_from_h, seed_resize_from_w, seed_checkbox, - height, - width, - enable_hr, - scale_latent, - denoising_strength, - ] + custom_inputs, + txt2img_prompt, + txt2img_negative_prompt, + txt2img_prompt_style, + txt2img_prompt_style2, + steps, + sampler_index, + restore_faces, + tiling, + batch_count, + batch_size, + cfg_scale, + seed, + subseed, subseed_strength, seed_resize_from_h, seed_resize_from_w, seed_checkbox, + height, + width, + enable_hr, + scale_latent, + denoising_strength, + aesthetic_lr, + aesthetic_weight, + aesthetic_steps, + aesthetic_imgs, + aesthetic_slerp + ] + custom_inputs, outputs=[ txt2img_gallery, generation_info, @@ -1097,6 +1111,9 @@ def create_ui(wrap_gradio_gpu_call): template_file = gr.Textbox(label='Prompt template file', value=os.path.join(script_path, "textual_inversion_templates", "style_filewords.txt")) training_width = gr.Slider(minimum=64, maximum=2048, step=64, label="Width", value=512) training_height = gr.Slider(minimum=64, maximum=2048, step=64, label="Height", value=512) + batch_size = gr.Slider(minimum=1, maximum=64, step=1, label="Batch Size", value=4) + gradient_accumulation = gr.Slider(minimum=1, maximum=256, step=1, label="Gradient accumulation", + value=1) steps = gr.Number(label='Max steps', value=100000, precision=0) create_image_every = gr.Number(label='Save an image to log directory every N steps, 0 to disable', value=500, precision=0) save_embedding_every = gr.Number(label='Save a copy of embedding to log directory every N steps, 0 to disable', value=500, precision=0) @@ -1180,6 +1197,8 @@ def create_ui(wrap_gradio_gpu_call): template_file, save_image_with_stored_embedding, preview_image_prompt, + batch_size, + gradient_accumulation ], outputs=[ ti_output, From 37d7ffb415cd8c69b3c0bb5f61844dde0b169f78 Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Sat, 15 Oct 2022 15:59:37 +0200 Subject: [PATCH 02/82] fix to tokens lenght, addend embs generator, add new features to edit the embedding before the generation using text --- modules/aesthetic_clip.py | 78 ++++++++++++++++++++ modules/processing.py | 148 +++++++++++++++++++++++++------------- modules/sd_hijack.py | 111 ++++++++++++++++++---------- modules/shared.py | 4 ++ modules/txt2img.py | 10 ++- modules/ui.py | 47 ++++++++++-- 6 files changed, 302 insertions(+), 96 deletions(-) create mode 100644 modules/aesthetic_clip.py diff --git a/modules/aesthetic_clip.py b/modules/aesthetic_clip.py new file mode 100644 index 000000000..f15cfd475 --- /dev/null +++ b/modules/aesthetic_clip.py @@ -0,0 +1,78 @@ +import itertools +import os +from pathlib import Path +import html +import gc + +import gradio as gr +import torch +from PIL import Image +from modules import shared +from modules.shared import device, aesthetic_embeddings +from transformers import CLIPModel, CLIPProcessor + +from tqdm.auto import tqdm + + +def get_all_images_in_folder(folder): + return [os.path.join(folder, f) for f in os.listdir(folder) if + os.path.isfile(os.path.join(folder, f)) and check_is_valid_image_file(f)] + + +def check_is_valid_image_file(filename): + return filename.lower().endswith(('.png', '.jpg', '.jpeg')) + + +def batched(dataset, total, n=1): + for ndx in range(0, total, n): + yield [dataset.__getitem__(i) for i in range(ndx, min(ndx + n, total))] + + +def iter_to_batched(iterable, n=1): + it = iter(iterable) + while True: + chunk = tuple(itertools.islice(it, n)) + if not chunk: + return + yield chunk + + +def generate_imgs_embd(name, folder, batch_size): + # clipModel = CLIPModel.from_pretrained( + # shared.sd_model.cond_stage_model.clipModel.name_or_path + # ) + model = CLIPModel.from_pretrained(shared.sd_model.cond_stage_model.clipModel.name_or_path).to(device) + processor = CLIPProcessor.from_pretrained(shared.sd_model.cond_stage_model.clipModel.name_or_path) + + with torch.no_grad(): + embs = [] + for paths in tqdm(iter_to_batched(get_all_images_in_folder(folder), batch_size), + desc=f"Generating embeddings for {name}"): + if shared.state.interrupted: + break + inputs = processor(images=[Image.open(path) for path in paths], return_tensors="pt").to(device) + outputs = model.get_image_features(**inputs).cpu() + embs.append(torch.clone(outputs)) + inputs.to("cpu") + del inputs, outputs + + embs = torch.cat(embs, dim=0).mean(dim=0, keepdim=True) + + # The generated embedding will be located here + path = str(Path(shared.cmd_opts.aesthetic_embeddings_dir) / f"{name}.pt") + torch.save(embs, path) + + model = model.cpu() + del model + del processor + del embs + gc.collect() + torch.cuda.empty_cache() + res = f""" + Done generating embedding for {name}! + Hypernetwork saved to {html.escape(path)} + """ + shared.update_aesthetic_embeddings() + return gr.Dropdown(sorted(aesthetic_embeddings.keys()), label="Imgs embedding", + value=sorted(aesthetic_embeddings.keys())[0] if len( + aesthetic_embeddings) > 0 else None), res, "" diff --git a/modules/processing.py b/modules/processing.py index 9a0337591..ab68d63a4 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -20,7 +20,6 @@ import modules.images as images import modules.styles import logging - # some of those options should not be changed at all because they would break the model, so I removed them from options. opt_C = 4 opt_f = 8 @@ -52,8 +51,13 @@ def get_correct_sampler(p): elif isinstance(p, modules.processing.StableDiffusionProcessingImg2Img): return sd_samplers.samplers_for_img2img + class StableDiffusionProcessing: - def __init__(self, sd_model=None, outpath_samples=None, outpath_grids=None, prompt="", styles=None, seed=-1, subseed=-1, subseed_strength=0, seed_resize_from_h=-1, seed_resize_from_w=-1, seed_enable_extras=True, sampler_index=0, batch_size=1, n_iter=1, steps=50, cfg_scale=7.0, width=512, height=512, restore_faces=False, tiling=False, do_not_save_samples=False, do_not_save_grid=False, extra_generation_params=None, overlay_images=None, negative_prompt=None, eta=None): + def __init__(self, sd_model=None, outpath_samples=None, outpath_grids=None, prompt="", styles=None, seed=-1, + subseed=-1, subseed_strength=0, seed_resize_from_h=-1, seed_resize_from_w=-1, seed_enable_extras=True, + sampler_index=0, batch_size=1, n_iter=1, steps=50, cfg_scale=7.0, width=512, height=512, + restore_faces=False, tiling=False, do_not_save_samples=False, do_not_save_grid=False, + extra_generation_params=None, overlay_images=None, negative_prompt=None, eta=None): self.sd_model = sd_model self.outpath_samples: str = outpath_samples self.outpath_grids: str = outpath_grids @@ -104,7 +108,8 @@ class StableDiffusionProcessing: class Processed: - def __init__(self, p: StableDiffusionProcessing, images_list, seed=-1, info="", subseed=None, all_prompts=None, all_seeds=None, all_subseeds=None, index_of_first_image=0, infotexts=None): + def __init__(self, p: StableDiffusionProcessing, images_list, seed=-1, info="", subseed=None, all_prompts=None, + all_seeds=None, all_subseeds=None, index_of_first_image=0, infotexts=None): self.images = images_list self.prompt = p.prompt self.negative_prompt = p.negative_prompt @@ -141,7 +146,8 @@ class Processed: self.prompt = self.prompt if type(self.prompt) != list else self.prompt[0] self.negative_prompt = self.negative_prompt if type(self.negative_prompt) != list else self.negative_prompt[0] self.seed = int(self.seed if type(self.seed) != list else self.seed[0]) - self.subseed = int(self.subseed if type(self.subseed) != list else self.subseed[0]) if self.subseed is not None else -1 + self.subseed = int( + self.subseed if type(self.subseed) != list else self.subseed[0]) if self.subseed is not None else -1 self.all_prompts = all_prompts or [self.prompt] self.all_seeds = all_seeds or [self.seed] @@ -181,39 +187,43 @@ class Processed: return json.dumps(obj) - def infotext(self, p: StableDiffusionProcessing, index): - return create_infotext(p, self.all_prompts, self.all_seeds, self.all_subseeds, comments=[], position_in_batch=index % self.batch_size, iteration=index // self.batch_size) + def infotext(self, p: StableDiffusionProcessing, index): + return create_infotext(p, self.all_prompts, self.all_seeds, self.all_subseeds, comments=[], + position_in_batch=index % self.batch_size, iteration=index // self.batch_size) # from https://discuss.pytorch.org/t/help-regarding-slerp-function-for-generative-model-sampling/32475/3 def slerp(val, low, high): - low_norm = low/torch.norm(low, dim=1, keepdim=True) - high_norm = high/torch.norm(high, dim=1, keepdim=True) - dot = (low_norm*high_norm).sum(1) + low_norm = low / torch.norm(low, dim=1, keepdim=True) + high_norm = high / torch.norm(high, dim=1, keepdim=True) + dot = (low_norm * high_norm).sum(1) if dot.mean() > 0.9995: return low * val + high * (1 - val) omega = torch.acos(dot) so = torch.sin(omega) - res = (torch.sin((1.0-val)*omega)/so).unsqueeze(1)*low + (torch.sin(val*omega)/so).unsqueeze(1) * high + res = (torch.sin((1.0 - val) * omega) / so).unsqueeze(1) * low + (torch.sin(val * omega) / so).unsqueeze(1) * high return res -def create_random_tensors(shape, seeds, subseeds=None, subseed_strength=0.0, seed_resize_from_h=0, seed_resize_from_w=0, p=None): +def create_random_tensors(shape, seeds, subseeds=None, subseed_strength=0.0, seed_resize_from_h=0, seed_resize_from_w=0, + p=None): xs = [] # if we have multiple seeds, this means we are working with batch size>1; this then # enables the generation of additional tensors with noise that the sampler will use during its processing. # Using those pre-generated tensors instead of simple torch.randn allows a batch with seeds [100, 101] to # produce the same images as with two batches [100], [101]. - if p is not None and p.sampler is not None and (len(seeds) > 1 and opts.enable_batch_seeds or opts.eta_noise_seed_delta > 0): + if p is not None and p.sampler is not None and ( + len(seeds) > 1 and opts.enable_batch_seeds or opts.eta_noise_seed_delta > 0): sampler_noises = [[] for _ in range(p.sampler.number_of_needed_noises(p))] else: sampler_noises = None for i, seed in enumerate(seeds): - noise_shape = shape if seed_resize_from_h <= 0 or seed_resize_from_w <= 0 else (shape[0], seed_resize_from_h//8, seed_resize_from_w//8) + noise_shape = shape if seed_resize_from_h <= 0 or seed_resize_from_w <= 0 else ( + shape[0], seed_resize_from_h // 8, seed_resize_from_w // 8) subnoise = None if subseeds is not None: @@ -241,7 +251,7 @@ def create_random_tensors(shape, seeds, subseeds=None, subseed_strength=0.0, see dx = max(-dx, 0) dy = max(-dy, 0) - x[:, ty:ty+h, tx:tx+w] = noise[:, dy:dy+h, dx:dx+w] + x[:, ty:ty + h, tx:tx + w] = noise[:, dy:dy + h, dx:dx + w] noise = x if sampler_noises is not None: @@ -293,14 +303,20 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration "Seed": all_seeds[index], "Face restoration": (opts.face_restoration_model if p.restore_faces else None), "Size": f"{p.width}x{p.height}", - "Model hash": getattr(p, 'sd_model_hash', None if not opts.add_model_hash_to_info or not shared.sd_model.sd_model_hash else shared.sd_model.sd_model_hash), - "Model": (None if not opts.add_model_name_to_info or not shared.sd_model.sd_checkpoint_info.model_name else shared.sd_model.sd_checkpoint_info.model_name.replace(',', '').replace(':', '')), - "Hypernet": (None if shared.loaded_hypernetwork is None else shared.loaded_hypernetwork.name.replace(',', '').replace(':', '')), + "Model hash": getattr(p, 'sd_model_hash', + None if not opts.add_model_hash_to_info or not shared.sd_model.sd_model_hash else shared.sd_model.sd_model_hash), + "Model": ( + None if not opts.add_model_name_to_info or not shared.sd_model.sd_checkpoint_info.model_name else shared.sd_model.sd_checkpoint_info.model_name.replace( + ',', '').replace(':', '')), + "Hypernet": ( + None if shared.loaded_hypernetwork is None else shared.loaded_hypernetwork.name.replace(',', '').replace( + ':', '')), "Batch size": (None if p.batch_size < 2 else p.batch_size), "Batch pos": (None if p.batch_size < 2 else position_in_batch), "Variation seed": (None if p.subseed_strength == 0 else all_subseeds[index]), "Variation seed strength": (None if p.subseed_strength == 0 else p.subseed_strength), - "Seed resize from": (None if p.seed_resize_from_w == 0 or p.seed_resize_from_h == 0 else f"{p.seed_resize_from_w}x{p.seed_resize_from_h}"), + "Seed resize from": ( + None if p.seed_resize_from_w == 0 or p.seed_resize_from_h == 0 else f"{p.seed_resize_from_w}x{p.seed_resize_from_h}"), "Denoising strength": getattr(p, 'denoising_strength', None), "Eta": (None if p.sampler is None or p.sampler.eta == p.sampler.default_eta else p.sampler.eta), "Clip skip": None if clip_skip <= 1 else clip_skip, @@ -309,7 +325,8 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration generation_params.update(p.extra_generation_params) - generation_params_text = ", ".join([k if k == v else f'{k}: {v}' for k, v in generation_params.items() if v is not None]) + generation_params_text = ", ".join( + [k if k == v else f'{k}: {v}' for k, v in generation_params.items() if v is not None]) negative_prompt_text = "\nNegative prompt: " + p.negative_prompt if p.negative_prompt else "" @@ -317,7 +334,9 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration def process_images(p: StableDiffusionProcessing, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, - aesthetic_imgs=None,aesthetic_slerp=False) -> Processed: + aesthetic_imgs=None, aesthetic_slerp=False, aesthetic_imgs_text="", + aesthetic_slerp_angle=0.15, + aesthetic_text_negative=False) -> Processed: """this is the main loop that both txt2img and img2img use; it calls func_init once inside all the scopes and func_sample once per batch""" aesthetic_lr = float(aesthetic_lr) @@ -385,7 +404,7 @@ def process_images(p: StableDiffusionProcessing, aesthetic_lr=0, aesthetic_weigh for n in range(p.n_iter): if state.skipped: state.skipped = False - + if state.interrupted: break @@ -396,16 +415,19 @@ def process_images(p: StableDiffusionProcessing, aesthetic_lr=0, aesthetic_weigh if (len(prompts) == 0): break - #uc = p.sd_model.get_learned_conditioning(len(prompts) * [p.negative_prompt]) - #c = p.sd_model.get_learned_conditioning(prompts) + # uc = p.sd_model.get_learned_conditioning(len(prompts) * [p.negative_prompt]) + # c = p.sd_model.get_learned_conditioning(prompts) with devices.autocast(): if hasattr(shared.sd_model.cond_stage_model, "set_aesthetic_params"): - shared.sd_model.cond_stage_model.set_aesthetic_params(0, 0, 0) + shared.sd_model.cond_stage_model.set_aesthetic_params() uc = prompt_parser.get_learned_conditioning(shared.sd_model, len(prompts) * [p.negative_prompt], p.steps) if hasattr(shared.sd_model.cond_stage_model, "set_aesthetic_params"): shared.sd_model.cond_stage_model.set_aesthetic_params(aesthetic_lr, aesthetic_weight, - aesthetic_steps, aesthetic_imgs,aesthetic_slerp) + aesthetic_steps, aesthetic_imgs, + aesthetic_slerp, aesthetic_imgs_text, + aesthetic_slerp_angle, + aesthetic_text_negative) c = prompt_parser.get_multicond_learned_conditioning(shared.sd_model, prompts, p.steps) if len(model_hijack.comments) > 0: @@ -413,13 +435,13 @@ def process_images(p: StableDiffusionProcessing, aesthetic_lr=0, aesthetic_weigh comments[comment] = 1 if p.n_iter > 1: - shared.state.job = f"Batch {n+1} out of {p.n_iter}" + shared.state.job = f"Batch {n + 1} out of {p.n_iter}" with devices.autocast(): - samples_ddim = p.sample(conditioning=c, unconditional_conditioning=uc, seeds=seeds, subseeds=subseeds, subseed_strength=p.subseed_strength) + samples_ddim = p.sample(conditioning=c, unconditional_conditioning=uc, seeds=seeds, subseeds=subseeds, + subseed_strength=p.subseed_strength) if state.interrupted or state.skipped: - # if we are interrupted, sample returns just noise # use the image collected previously in sampler loop samples_ddim = shared.state.current_latent @@ -445,7 +467,9 @@ def process_images(p: StableDiffusionProcessing, aesthetic_lr=0, aesthetic_weigh if p.restore_faces: if opts.save and not p.do_not_save_samples and opts.save_images_before_face_restoration: - images.save_image(Image.fromarray(x_sample), p.outpath_samples, "", seeds[i], prompts[i], opts.samples_format, info=infotext(n, i), p=p, suffix="-before-face-restoration") + images.save_image(Image.fromarray(x_sample), p.outpath_samples, "", seeds[i], prompts[i], + opts.samples_format, info=infotext(n, i), p=p, + suffix="-before-face-restoration") devices.torch_gc() @@ -456,7 +480,8 @@ def process_images(p: StableDiffusionProcessing, aesthetic_lr=0, aesthetic_weigh if p.color_corrections is not None and i < len(p.color_corrections): if opts.save and not p.do_not_save_samples and opts.save_images_before_color_correction: - images.save_image(image, p.outpath_samples, "", seeds[i], prompts[i], opts.samples_format, info=infotext(n, i), p=p, suffix="-before-color-correction") + images.save_image(image, p.outpath_samples, "", seeds[i], prompts[i], opts.samples_format, + info=infotext(n, i), p=p, suffix="-before-color-correction") image = apply_color_correction(p.color_corrections[i], image) if p.overlay_images is not None and i < len(p.overlay_images): @@ -474,7 +499,8 @@ def process_images(p: StableDiffusionProcessing, aesthetic_lr=0, aesthetic_weigh image = image.convert('RGB') if opts.samples_save and not p.do_not_save_samples: - images.save_image(image, p.outpath_samples, "", seeds[i], prompts[i], opts.samples_format, info=infotext(n, i), p=p) + images.save_image(image, p.outpath_samples, "", seeds[i], prompts[i], opts.samples_format, + info=infotext(n, i), p=p) text = infotext(n, i) infotexts.append(text) @@ -482,7 +508,7 @@ def process_images(p: StableDiffusionProcessing, aesthetic_lr=0, aesthetic_weigh image.info["parameters"] = text output_images.append(image) - del x_samples_ddim + del x_samples_ddim devices.torch_gc() @@ -504,10 +530,13 @@ def process_images(p: StableDiffusionProcessing, aesthetic_lr=0, aesthetic_weigh index_of_first_image = 1 if opts.grid_save: - images.save_image(grid, p.outpath_grids, "grid", all_seeds[0], all_prompts[0], opts.grid_format, info=infotext(), short_filename=not opts.grid_extended_filename, p=p, grid=True) + images.save_image(grid, p.outpath_grids, "grid", all_seeds[0], all_prompts[0], opts.grid_format, + info=infotext(), short_filename=not opts.grid_extended_filename, p=p, grid=True) devices.torch_gc() - return Processed(p, output_images, all_seeds[0], infotext() + "".join(["\n\n" + x for x in comments]), subseed=all_subseeds[0], all_prompts=all_prompts, all_seeds=all_seeds, all_subseeds=all_subseeds, index_of_first_image=index_of_first_image, infotexts=infotexts) + return Processed(p, output_images, all_seeds[0], infotext() + "".join(["\n\n" + x for x in comments]), + subseed=all_subseeds[0], all_prompts=all_prompts, all_seeds=all_seeds, all_subseeds=all_subseeds, + index_of_first_image=index_of_first_image, infotexts=infotexts) class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): @@ -543,25 +572,34 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): self.sampler = sd_samplers.create_sampler_with_index(sd_samplers.samplers, self.sampler_index, self.sd_model) if not self.enable_hr: - x = create_random_tensors([opt_C, self.height // opt_f, self.width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) + x = create_random_tensors([opt_C, self.height // opt_f, self.width // opt_f], seeds=seeds, + subseeds=subseeds, subseed_strength=self.subseed_strength, + seed_resize_from_h=self.seed_resize_from_h, + seed_resize_from_w=self.seed_resize_from_w, p=self) samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning) return samples - x = create_random_tensors([opt_C, self.firstphase_height // opt_f, self.firstphase_width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) + x = create_random_tensors([opt_C, self.firstphase_height // opt_f, self.firstphase_width // opt_f], seeds=seeds, + subseeds=subseeds, subseed_strength=self.subseed_strength, + seed_resize_from_h=self.seed_resize_from_h, + seed_resize_from_w=self.seed_resize_from_w, p=self) samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning) truncate_x = (self.firstphase_width - self.firstphase_width_truncated) // opt_f truncate_y = (self.firstphase_height - self.firstphase_height_truncated) // opt_f - samples = samples[:, :, truncate_y//2:samples.shape[2]-truncate_y//2, truncate_x//2:samples.shape[3]-truncate_x//2] + samples = samples[:, :, truncate_y // 2:samples.shape[2] - truncate_y // 2, + truncate_x // 2:samples.shape[3] - truncate_x // 2] if self.scale_latent: - samples = torch.nn.functional.interpolate(samples, size=(self.height // opt_f, self.width // opt_f), mode="bilinear") + samples = torch.nn.functional.interpolate(samples, size=(self.height // opt_f, self.width // opt_f), + mode="bilinear") else: decoded_samples = decode_first_stage(self.sd_model, samples) if opts.upscaler_for_img2img is None or opts.upscaler_for_img2img == "None": - decoded_samples = torch.nn.functional.interpolate(decoded_samples, size=(self.height, self.width), mode="bilinear") + decoded_samples = torch.nn.functional.interpolate(decoded_samples, size=(self.height, self.width), + mode="bilinear") else: lowres_samples = torch.clamp((decoded_samples + 1.0) / 2.0, min=0.0, max=1.0) @@ -585,13 +623,16 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): self.sampler = sd_samplers.create_sampler_with_index(sd_samplers.samplers, self.sampler_index, self.sd_model) - noise = create_random_tensors(samples.shape[1:], seeds=seeds, subseeds=subseeds, subseed_strength=subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) + noise = create_random_tensors(samples.shape[1:], seeds=seeds, subseeds=subseeds, + subseed_strength=subseed_strength, seed_resize_from_h=self.seed_resize_from_h, + seed_resize_from_w=self.seed_resize_from_w, p=self) # GC now before running the next img2img to prevent running out of memory x = None devices.torch_gc() - samples = self.sampler.sample_img2img(self, samples, noise, conditioning, unconditional_conditioning, steps=self.steps) + samples = self.sampler.sample_img2img(self, samples, noise, conditioning, unconditional_conditioning, + steps=self.steps) return samples @@ -599,7 +640,9 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): sampler = None - def __init__(self, init_images=None, resize_mode=0, denoising_strength=0.75, mask=None, mask_blur=4, inpainting_fill=0, inpaint_full_res=True, inpaint_full_res_padding=0, inpainting_mask_invert=0, **kwargs): + def __init__(self, init_images=None, resize_mode=0, denoising_strength=0.75, mask=None, mask_blur=4, + inpainting_fill=0, inpaint_full_res=True, inpaint_full_res_padding=0, inpainting_mask_invert=0, + **kwargs): super().__init__(**kwargs) self.init_images = init_images @@ -607,7 +650,7 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): self.denoising_strength: float = denoising_strength self.init_latent = None self.image_mask = mask - #self.image_unblurred_mask = None + # self.image_unblurred_mask = None self.latent_mask = None self.mask_for_overlay = None self.mask_blur = mask_blur @@ -619,7 +662,8 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): self.nmask = None def init(self, all_prompts, all_seeds, all_subseeds): - self.sampler = sd_samplers.create_sampler_with_index(sd_samplers.samplers_for_img2img, self.sampler_index, self.sd_model) + self.sampler = sd_samplers.create_sampler_with_index(sd_samplers.samplers_for_img2img, self.sampler_index, + self.sd_model) crop_region = None if self.image_mask is not None: @@ -628,7 +672,7 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): if self.inpainting_mask_invert: self.image_mask = ImageOps.invert(self.image_mask) - #self.image_unblurred_mask = self.image_mask + # self.image_unblurred_mask = self.image_mask if self.mask_blur > 0: self.image_mask = self.image_mask.filter(ImageFilter.GaussianBlur(self.mask_blur)) @@ -642,7 +686,7 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): mask = mask.crop(crop_region) self.image_mask = images.resize_image(2, mask, self.width, self.height) - self.paste_to = (x1, y1, x2-x1, y2-y1) + self.paste_to = (x1, y1, x2 - x1, y2 - y1) else: self.image_mask = images.resize_image(self.resize_mode, self.image_mask, self.width, self.height) np_mask = np.array(self.image_mask) @@ -665,7 +709,8 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): if self.image_mask is not None: image_masked = Image.new('RGBa', (image.width, image.height)) - image_masked.paste(image.convert("RGBA").convert("RGBa"), mask=ImageOps.invert(self.mask_for_overlay.convert('L'))) + image_masked.paste(image.convert("RGBA").convert("RGBa"), + mask=ImageOps.invert(self.mask_for_overlay.convert('L'))) self.overlay_images.append(image_masked.convert('RGBA')) @@ -714,12 +759,17 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): # this needs to be fixed to be done in sample() using actual seeds for batches if self.inpainting_fill == 2: - self.init_latent = self.init_latent * self.mask + create_random_tensors(self.init_latent.shape[1:], all_seeds[0:self.init_latent.shape[0]]) * self.nmask + self.init_latent = self.init_latent * self.mask + create_random_tensors(self.init_latent.shape[1:], + all_seeds[ + 0:self.init_latent.shape[ + 0]]) * self.nmask elif self.inpainting_fill == 3: self.init_latent = self.init_latent * self.mask def sample(self, conditioning, unconditional_conditioning, seeds, subseeds, subseed_strength): - x = create_random_tensors([opt_C, self.height // opt_f, self.width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) + x = create_random_tensors([opt_C, self.height // opt_f, self.width // opt_f], seeds=seeds, subseeds=subseeds, + subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, + seed_resize_from_w=self.seed_resize_from_w, p=self) samples = self.sampler.sample_img2img(self, self.init_latent, x, conditioning, unconditional_conditioning) diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py index 6d5196fea..192883b2d 100644 --- a/modules/sd_hijack.py +++ b/modules/sd_hijack.py @@ -14,7 +14,8 @@ from modules.sd_hijack_optimizations import invokeAI_mps_available import ldm.modules.attention import ldm.modules.diffusionmodules.model -from transformers import CLIPVisionModel, CLIPModel +from tqdm import trange +from transformers import CLIPVisionModel, CLIPModel, CLIPTokenizer import torch.optim as optim import copy @@ -22,21 +23,25 @@ attention_CrossAttention_forward = ldm.modules.attention.CrossAttention.forward diffusionmodules_model_nonlinearity = ldm.modules.diffusionmodules.model.nonlinearity diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.AttnBlock.forward + def apply_optimizations(): undo_optimizations() ldm.modules.diffusionmodules.model.nonlinearity = silu - if cmd_opts.force_enable_xformers or (cmd_opts.xformers and shared.xformers_available and torch.version.cuda and (6, 0) <= torch.cuda.get_device_capability(shared.device) <= (8, 6)): + if cmd_opts.force_enable_xformers or (cmd_opts.xformers and shared.xformers_available and torch.version.cuda and ( + 6, 0) <= torch.cuda.get_device_capability(shared.device) <= (8, 6)): print("Applying xformers cross attention optimization.") ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.xformers_attnblock_forward elif cmd_opts.opt_split_attention_v1: print("Applying v1 cross attention optimization.") ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1 - elif not cmd_opts.disable_opt_split_attention and (cmd_opts.opt_split_attention_invokeai or not torch.cuda.is_available()): + elif not cmd_opts.disable_opt_split_attention and ( + cmd_opts.opt_split_attention_invokeai or not torch.cuda.is_available()): if not invokeAI_mps_available and shared.device.type == 'mps': - print("The InvokeAI cross attention optimization for MPS requires the psutil package which is not installed.") + print( + "The InvokeAI cross attention optimization for MPS requires the psutil package which is not installed.") print("Applying v1 cross attention optimization.") ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1 else: @@ -112,14 +117,16 @@ class StableDiffusionModelHijack: _, remade_batch_tokens, _, _, _, token_count = self.clip.process_text([text]) return remade_batch_tokens[0], token_count, get_target_prompt_token_count(token_count) + def slerp(low, high, val): - low_norm = low/torch.norm(low, dim=1, keepdim=True) - high_norm = high/torch.norm(high, dim=1, keepdim=True) - omega = torch.acos((low_norm*high_norm).sum(1)) + low_norm = low / torch.norm(low, dim=1, keepdim=True) + high_norm = high / torch.norm(high, dim=1, keepdim=True) + omega = torch.acos((low_norm * high_norm).sum(1)) so = torch.sin(omega) - res = (torch.sin((1.0-val)*omega)/so).unsqueeze(1)*low + (torch.sin(val*omega)/so).unsqueeze(1) * high + res = (torch.sin((1.0 - val) * omega) / so).unsqueeze(1) * low + (torch.sin(val * omega) / so).unsqueeze(1) * high return res + class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): def __init__(self, wrapped, hijack): super().__init__() @@ -128,6 +135,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): self.wrapped.transformer.name_or_path ) del self.clipModel.vision_model + self.tokenizer = CLIPTokenizer.from_pretrained(self.wrapped.transformer.name_or_path) self.hijack: StableDiffusionModelHijack = hijack self.tokenizer = wrapped.tokenizer # self.vision = CLIPVisionModel.from_pretrained(self.wrapped.transformer.name_or_path).eval() @@ -139,7 +147,8 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): self.comma_token = [v for k, v in self.tokenizer.get_vocab().items() if k == ','][0] - tokens_with_parens = [(k, v) for k, v in self.tokenizer.get_vocab().items() if '(' in k or ')' in k or '[' in k or ']' in k] + tokens_with_parens = [(k, v) for k, v in self.tokenizer.get_vocab().items() if + '(' in k or ')' in k or '[' in k or ']' in k] for text, ident in tokens_with_parens: mult = 1.0 for c in text: @@ -155,8 +164,13 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): if mult != 1.0: self.token_mults[ident] = mult - def set_aesthetic_params(self, aesthetic_lr, aesthetic_weight, aesthetic_steps, image_embs_name=None, - aesthetic_slerp=True): + def set_aesthetic_params(self, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, image_embs_name=None, + aesthetic_slerp=True, aesthetic_imgs_text="", + aesthetic_slerp_angle=0.15, + aesthetic_text_negative=False): + self.aesthetic_imgs_text = aesthetic_imgs_text + self.aesthetic_slerp_angle = aesthetic_slerp_angle + self.aesthetic_text_negative = aesthetic_text_negative self.slerp = aesthetic_slerp self.aesthetic_lr = aesthetic_lr self.aesthetic_weight = aesthetic_weight @@ -180,7 +194,8 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): else: parsed = [[line, 1.0]] - tokenized = self.wrapped.tokenizer([text for text, _ in parsed], truncation=False, add_special_tokens=False)["input_ids"] + tokenized = self.wrapped.tokenizer([text for text, _ in parsed], truncation=False, add_special_tokens=False)[ + "input_ids"] fixes = [] remade_tokens = [] @@ -196,18 +211,20 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): if token == self.comma_token: last_comma = len(remade_tokens) - elif opts.comma_padding_backtrack != 0 and max(len(remade_tokens), 1) % 75 == 0 and last_comma != -1 and len(remade_tokens) - last_comma <= opts.comma_padding_backtrack: + elif opts.comma_padding_backtrack != 0 and max(len(remade_tokens), + 1) % 75 == 0 and last_comma != -1 and len( + remade_tokens) - last_comma <= opts.comma_padding_backtrack: last_comma += 1 reloc_tokens = remade_tokens[last_comma:] reloc_mults = multipliers[last_comma:] remade_tokens = remade_tokens[:last_comma] length = len(remade_tokens) - + rem = int(math.ceil(length / 75)) * 75 - length remade_tokens += [id_end] * rem + reloc_tokens multipliers = multipliers[:last_comma] + [1.0] * rem + reloc_mults - + if embedding is None: remade_tokens.append(token) multipliers.append(weight) @@ -248,7 +265,8 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): if line in cache: remade_tokens, fixes, multipliers = cache[line] else: - remade_tokens, fixes, multipliers, current_token_count = self.tokenize_line(line, used_custom_terms, hijack_comments) + remade_tokens, fixes, multipliers, current_token_count = self.tokenize_line(line, used_custom_terms, + hijack_comments) token_count = max(current_token_count, token_count) cache[line] = (remade_tokens, fixes, multipliers) @@ -259,7 +277,6 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count - def process_text_old(self, text): id_start = self.wrapped.tokenizer.bos_token_id id_end = self.wrapped.tokenizer.eos_token_id @@ -289,7 +306,8 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): while i < len(tokens): token = tokens[i] - embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i) + embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, + i) mult_change = self.token_mults.get(token) if opts.enable_emphasis else None if mult_change is not None: @@ -312,11 +330,12 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): ovf = remade_tokens[maxlen - 2:] overflowing_words = [vocab.get(int(x), "") for x in ovf] overflowing_text = self.wrapped.tokenizer.convert_tokens_to_string(''.join(overflowing_words)) - hijack_comments.append(f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n") + hijack_comments.append( + f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n") token_count = len(remade_tokens) remade_tokens = remade_tokens + [id_end] * (maxlen - 2 - len(remade_tokens)) - remade_tokens = [id_start] + remade_tokens[0:maxlen-2] + [id_end] + remade_tokens = [id_start] + remade_tokens[0:maxlen - 2] + [id_end] cache[tuple_tokens] = (remade_tokens, fixes, multipliers) multipliers = multipliers + [1.0] * (maxlen - 2 - len(multipliers)) @@ -326,23 +345,26 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): hijack_fixes.append(fixes) batch_multipliers.append(multipliers) return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count - + def forward(self, text): use_old = opts.use_old_emphasis_implementation if use_old: - batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text_old(text) + batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text_old( + text) else: - batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text(text) + batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text( + text) self.hijack.comments += hijack_comments if len(used_custom_terms) > 0: - self.hijack.comments.append("Used embeddings: " + ", ".join([f'{word} [{checksum}]' for word, checksum in used_custom_terms])) - + self.hijack.comments.append( + "Used embeddings: " + ", ".join([f'{word} [{checksum}]' for word, checksum in used_custom_terms])) + if use_old: self.hijack.fixes = hijack_fixes return self.process_tokens(remade_batch_tokens, batch_multipliers) - + z = None i = 0 while max(map(len, remade_batch_tokens)) != 0: @@ -356,7 +378,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): if fix[0] == i: fixes.append(fix[1]) self.hijack.fixes.append(fixes) - + tokens = [] multipliers = [] for j in range(len(remade_batch_tokens)): @@ -378,19 +400,30 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): remade_batch_tokens] tokens = torch.asarray(remade_batch_tokens).to(device) + + model = copy.deepcopy(self.clipModel).to(device) + model.requires_grad_(True) + if self.aesthetic_imgs_text is not None and len(self.aesthetic_imgs_text) > 0: + text_embs_2 = model.get_text_features( + **self.tokenizer([self.aesthetic_imgs_text], padding=True, return_tensors="pt").to(device)) + if self.aesthetic_text_negative: + text_embs_2 = self.image_embs - text_embs_2 + text_embs_2 /= text_embs_2.norm(dim=-1, keepdim=True) + img_embs = slerp(self.image_embs, text_embs_2, self.aesthetic_slerp_angle) + else: + img_embs = self.image_embs + with torch.enable_grad(): - model = copy.deepcopy(self.clipModel).to(device) - model.requires_grad_(True) # We optimize the model to maximize the similarity optimizer = optim.Adam( model.text_model.parameters(), lr=self.aesthetic_lr ) - for i in range(self.aesthetic_steps): + for i in trange(self.aesthetic_steps, desc="Aesthetic optimization"): text_embs = model.get_text_features(input_ids=tokens) text_embs = text_embs / text_embs.norm(dim=-1, keepdim=True) - sim = text_embs @ self.image_embs.T + sim = text_embs @ img_embs.T loss = -sim optimizer.zero_grad() loss.mean().backward() @@ -405,6 +438,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): model.cpu() del model + zn = torch.concat([zn for i in range(z.shape[1] // 77)], 1) if self.slerp: z = slerp(z, zn, self.aesthetic_weight) else: @@ -413,15 +447,16 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): remade_batch_tokens = rem_tokens batch_multipliers = rem_multipliers i += 1 - + return z - - + def process_tokens(self, remade_batch_tokens, batch_multipliers): if not opts.use_old_emphasis_implementation: - remade_batch_tokens = [[self.wrapped.tokenizer.bos_token_id] + x[:75] + [self.wrapped.tokenizer.eos_token_id] for x in remade_batch_tokens] + remade_batch_tokens = [ + [self.wrapped.tokenizer.bos_token_id] + x[:75] + [self.wrapped.tokenizer.eos_token_id] for x in + remade_batch_tokens] batch_multipliers = [[1.0] + x[:75] + [1.0] for x in batch_multipliers] - + tokens = torch.asarray(remade_batch_tokens).to(device) outputs = self.wrapped.transformer(input_ids=tokens, output_hidden_states=-opts.CLIP_stop_at_last_layers) @@ -461,8 +496,8 @@ class EmbeddingsWithFixes(torch.nn.Module): for fixes, tensor in zip(batch_fixes, inputs_embeds): for offset, embedding in fixes: emb = embedding.vec - emb_len = min(tensor.shape[0]-offset-1, emb.shape[0]) - tensor = torch.cat([tensor[0:offset+1], emb[0:emb_len], tensor[offset+1+emb_len:]]) + emb_len = min(tensor.shape[0] - offset - 1, emb.shape[0]) + tensor = torch.cat([tensor[0:offset + 1], emb[0:emb_len], tensor[offset + 1 + emb_len:]]) vecs.append(tensor) diff --git a/modules/shared.py b/modules/shared.py index cf13a10d9..7cd608cae 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -95,6 +95,10 @@ loaded_hypernetwork = None aesthetic_embeddings = {f.replace(".pt",""): os.path.join(cmd_opts.aesthetic_embeddings_dir, f) for f in os.listdir(cmd_opts.aesthetic_embeddings_dir) if f.endswith(".pt")} +def update_aesthetic_embeddings(): + global aesthetic_embeddings + aesthetic_embeddings = {f.replace(".pt",""): os.path.join(cmd_opts.aesthetic_embeddings_dir, f) for f in + os.listdir(cmd_opts.aesthetic_embeddings_dir) if f.endswith(".pt")} def reload_hypernetworks(): global hypernetworks diff --git a/modules/txt2img.py b/modules/txt2img.py index 783420243..eedcdfe02 100644 --- a/modules/txt2img.py +++ b/modules/txt2img.py @@ -13,7 +13,11 @@ def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, aesthetic_imgs=None, - aesthetic_slerp=False, *args): + aesthetic_slerp=False, + aesthetic_imgs_text="", + aesthetic_slerp_angle=0.15, + aesthetic_text_negative=False, + *args): p = StableDiffusionProcessingTxt2Img( sd_model=shared.sd_model, outpath_samples=opts.outdir_samples or opts.outdir_txt2img_samples, @@ -47,7 +51,9 @@ def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: processed = modules.scripts.scripts_txt2img.run(p, *args) if processed is None: - processed = process_images(p, aesthetic_lr, aesthetic_weight, aesthetic_steps, aesthetic_imgs, aesthetic_slerp) + processed = process_images(p, aesthetic_lr, aesthetic_weight, aesthetic_steps, aesthetic_imgs, aesthetic_slerp,aesthetic_imgs_text, + aesthetic_slerp_angle, + aesthetic_text_negative) shared.total_tqdm.clear() diff --git a/modules/ui.py b/modules/ui.py index d961d126e..e98e21135 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -41,6 +41,7 @@ from modules import prompt_parser from modules.images import save_image import modules.textual_inversion.ui import modules.hypernetworks.ui +import modules.aesthetic_clip # this is a fix for Windows users. Without it, javascript files will be served with text/html content-type and the browser will not show any UI mimetypes.init() @@ -449,7 +450,7 @@ def create_toprow(is_img2img): with gr.Row(): negative_prompt = gr.Textbox(label="Negative prompt", elem_id="negative_prompt", show_label=False, placeholder="Negative prompt", lines=2) with gr.Column(scale=1, elem_id="roll_col"): - sh = gr.Button(elem_id="sh", visible=True) + sh = gr.Button(elem_id="sh", visible=True) with gr.Column(scale=1, elem_id="style_neg_col"): prompt_style2 = gr.Dropdown(label="Style 2", elem_id=f"{id_part}_style2_index", choices=[k for k, v in shared.prompt_styles.styles.items()], value=next(iter(shared.prompt_styles.styles.keys())), visible=len(shared.prompt_styles.styles) > 1) @@ -536,9 +537,13 @@ def create_ui(wrap_gradio_gpu_call): height = gr.Slider(minimum=64, maximum=2048, step=64, label="Height", value=512) with gr.Group(): - aesthetic_lr = gr.Textbox(label='Learning rate', placeholder="Learning rate", value="0.005") - aesthetic_weight = gr.Slider(minimum=0, maximum=1, step=0.01, label="Aesthetic weight", value=0.7) - aesthetic_steps = gr.Slider(minimum=0, maximum=50, step=1, label="Aesthetic steps", value=50) + aesthetic_lr = gr.Textbox(label='Learning rate', placeholder="Learning rate", value="0.0001") + aesthetic_weight = gr.Slider(minimum=0, maximum=1, step=0.01, label="Aesthetic weight", value=0.9) + aesthetic_steps = gr.Slider(minimum=0, maximum=256, step=1, label="Aesthetic steps", value=5) + with gr.Row(): + aesthetic_imgs_text = gr.Textbox(label='Aesthetic text for imgs', placeholder="This text is used to rotate the feature space of the imgs embs", value="") + aesthetic_slerp_angle = gr.Slider(label='Slerp angle',minimum=0, maximum=1, step=0.01, value=0.1) + aesthetic_text_negative = gr.Checkbox(label="Is negative text", value=False) aesthetic_imgs = gr.Dropdown(sorted(aesthetic_embeddings.keys()), label="Imgs embedding", value=sorted(aesthetic_embeddings.keys())[0] if len(aesthetic_embeddings) > 0 else None) aesthetic_slerp = gr.Checkbox(label="Slerp interpolation", value=False) @@ -617,7 +622,10 @@ def create_ui(wrap_gradio_gpu_call): aesthetic_weight, aesthetic_steps, aesthetic_imgs, - aesthetic_slerp + aesthetic_slerp, + aesthetic_imgs_text, + aesthetic_slerp_angle, + aesthetic_text_negative ] + custom_inputs, outputs=[ txt2img_gallery, @@ -721,7 +729,7 @@ def create_ui(wrap_gradio_gpu_call): with gr.Row(): inpaint_full_res = gr.Checkbox(label='Inpaint at full resolution', value=False) - inpaint_full_res_padding = gr.Slider(label='Inpaint at full resolution padding, pixels', minimum=0, maximum=256, step=4, value=32) + inpaint_full_res_padding = gr.Slider(label='Inpaint at full resolution padding, pixels', minimum=0, maximum=1024, step=4, value=32) with gr.TabItem('Batch img2img', id='batch'): hidden = '
Disabled when launched with --hide-ui-dir-config.' if shared.cmd_opts.hide_ui_dir_config else '' @@ -1071,6 +1079,17 @@ def create_ui(wrap_gradio_gpu_call): with gr.Column(): create_embedding = gr.Button(value="Create embedding", variant='primary') + with gr.Tab(label="Create images embedding"): + new_embedding_name_ae = gr.Textbox(label="Name") + process_src_ae = gr.Textbox(label='Source directory') + batch_ae = gr.Slider(minimum=1, maximum=1024, step=1, label="Batch size", value=256) + with gr.Row(): + with gr.Column(scale=3): + gr.HTML(value="") + + with gr.Column(): + create_embedding_ae = gr.Button(value="Create images embedding", variant='primary') + with gr.Tab(label="Create hypernetwork"): new_hypernetwork_name = gr.Textbox(label="Name") new_hypernetwork_sizes = gr.CheckboxGroup(label="Modules", value=["768", "320", "640", "1280"], choices=["768", "320", "640", "1280"]) @@ -1139,7 +1158,7 @@ def create_ui(wrap_gradio_gpu_call): fn=modules.textual_inversion.ui.create_embedding, inputs=[ new_embedding_name, - initialization_text, + process_src, nvpt, ], outputs=[ @@ -1149,6 +1168,20 @@ def create_ui(wrap_gradio_gpu_call): ] ) + create_embedding_ae.click( + fn=modules.aesthetic_clip.generate_imgs_embd, + inputs=[ + new_embedding_name_ae, + process_src_ae, + batch_ae + ], + outputs=[ + aesthetic_imgs, + ti_output, + ti_outcome, + ] + ) + create_hypernetwork.click( fn=modules.hypernetworks.ui.create_hypernetwork, inputs=[ From 4387e4fe6479c08f7bc7e42924c3a1093e3a1872 Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Sat, 15 Oct 2022 18:39:29 +0200 Subject: [PATCH 03/82] Update modules/ui.py MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-authored-by: Víctor Gallego --- modules/ui.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/modules/ui.py b/modules/ui.py index d0696101e..5bb961b2d 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -599,7 +599,8 @@ def create_ui(wrap_gradio_gpu_call): with gr.Group(): aesthetic_lr = gr.Textbox(label='Learning rate', placeholder="Learning rate", value="0.0001") aesthetic_weight = gr.Slider(minimum=0, maximum=1, step=0.01, label="Aesthetic weight", value=0.9) - aesthetic_steps = gr.Slider(minimum=0, maximum=256, step=1, label="Aesthetic steps", value=5) + aesthetic_steps = gr.Slider(minimum=0, maximum=50, step=1, label="Aesthetic steps", value=5) + with gr.Row(): aesthetic_imgs_text = gr.Textbox(label='Aesthetic text for imgs', placeholder="This text is used to rotate the feature space of the imgs embs", value="") aesthetic_slerp_angle = gr.Slider(label='Slerp angle',minimum=0, maximum=1, step=0.01, value=0.1) From f7df06a98180a2a8769b3ceebf7b6a35eca8ffc5 Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Sat, 15 Oct 2022 18:40:06 +0200 Subject: [PATCH 04/82] Update README.md MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-authored-by: Víctor Gallego --- README.md | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index 7b8d018b5..40104833d 100644 --- a/README.md +++ b/README.md @@ -70,7 +70,8 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web - No token limit for prompts (original stable diffusion lets you use up to 75 tokens) - DeepDanbooru integration, creates danbooru style tags for anime prompts (add --deepdanbooru to commandline args) - [xformers](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Xformers), major speed increase for select cards: (add --xformers to commandline args) -- Aesthetic, a way to generate images with a specific aesthetic by using clip images embds (implementation of https://github.com/vicgalle/stable-diffusion-aesthetic-gradients) +- Aesthetic Gradients, a way to generate images with a specific aesthetic by using clip images embds (implementation of [https://github.com/vicgalle/stable-diffusion-aesthetic-gradients](https://github.com/vicgalle/stable-diffusion-aesthetic-gradients)) + ## Installation and Running Make sure the required [dependencies](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Dependencies) are met and follow the instructions available for both [NVidia](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs) (recommended) and [AMD](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs) GPUs. From 9b7705e0573bddde26df4575c71f994d73a4d519 Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Sat, 15 Oct 2022 18:40:34 +0200 Subject: [PATCH 05/82] Update modules/aesthetic_clip.py MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-authored-by: Víctor Gallego --- modules/aesthetic_clip.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/aesthetic_clip.py b/modules/aesthetic_clip.py index f15cfd475..bcf2b0734 100644 --- a/modules/aesthetic_clip.py +++ b/modules/aesthetic_clip.py @@ -70,7 +70,7 @@ def generate_imgs_embd(name, folder, batch_size): torch.cuda.empty_cache() res = f""" Done generating embedding for {name}! - Hypernetwork saved to {html.escape(path)} + Aesthetic embedding saved to {html.escape(path)} """ shared.update_aesthetic_embeddings() return gr.Dropdown(sorted(aesthetic_embeddings.keys()), label="Imgs embedding", From 0d4f5db235357aeb4c7a8738179ba33aaf5a6b75 Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Sat, 15 Oct 2022 18:40:58 +0200 Subject: [PATCH 06/82] Update modules/ui.py MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-authored-by: Víctor Gallego --- modules/ui.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/modules/ui.py b/modules/ui.py index 5bb961b2d..25eba5480 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -597,7 +597,8 @@ def create_ui(wrap_gradio_gpu_call): height = gr.Slider(minimum=64, maximum=2048, step=64, label="Height", value=512) with gr.Group(): - aesthetic_lr = gr.Textbox(label='Learning rate', placeholder="Learning rate", value="0.0001") + aesthetic_lr = gr.Textbox(label='Aesthetic learning rate', placeholder="Aesthetic learning rate", value="0.0001") + aesthetic_weight = gr.Slider(minimum=0, maximum=1, step=0.01, label="Aesthetic weight", value=0.9) aesthetic_steps = gr.Slider(minimum=0, maximum=50, step=1, label="Aesthetic steps", value=5) From ad9bc604a8fadcfebe72be37f66cec51e7e87fb5 Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Sat, 15 Oct 2022 18:41:18 +0200 Subject: [PATCH 07/82] Update modules/ui.py MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-authored-by: Víctor Gallego --- modules/ui.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/modules/ui.py b/modules/ui.py index 25eba5480..3b28b69c0 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -607,7 +607,8 @@ def create_ui(wrap_gradio_gpu_call): aesthetic_slerp_angle = gr.Slider(label='Slerp angle',minimum=0, maximum=1, step=0.01, value=0.1) aesthetic_text_negative = gr.Checkbox(label="Is negative text", value=False) - aesthetic_imgs = gr.Dropdown(sorted(aesthetic_embeddings.keys()), label="Imgs embedding", value=sorted(aesthetic_embeddings.keys())[0] if len(aesthetic_embeddings) > 0 else None) + aesthetic_imgs = gr.Dropdown(sorted(aesthetic_embeddings.keys()), label="Aesthetic imgs embedding", value=sorted(aesthetic_embeddings.keys())[0] if len(aesthetic_embeddings) > 0 else None) + aesthetic_slerp = gr.Checkbox(label="Slerp interpolation", value=False) with gr.Row(): From 3f5c3b981e46c16bb10948d012575b25170efb3b Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Sat, 15 Oct 2022 18:41:46 +0200 Subject: [PATCH 08/82] Update modules/ui.py MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-authored-by: Víctor Gallego --- modules/ui.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/modules/ui.py b/modules/ui.py index 3b28b69c0..1f6fcdc90 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1190,7 +1190,8 @@ def create_ui(wrap_gradio_gpu_call): with gr.Column(): create_embedding = gr.Button(value="Create embedding", variant='primary') - with gr.Tab(label="Create images embedding"): + with gr.Tab(label="Create aesthetic images embedding"): + new_embedding_name_ae = gr.Textbox(label="Name") process_src_ae = gr.Textbox(label='Source directory') batch_ae = gr.Slider(minimum=1, maximum=1024, step=1, label="Batch size", value=256) From 3d21684ee30ca5734126b8d08c05b3a0f513fe75 Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Sun, 16 Oct 2022 00:01:00 +0200 Subject: [PATCH 09/82] Add support to other img format, fixed dropbox update --- modules/aesthetic_clip.py | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/modules/aesthetic_clip.py b/modules/aesthetic_clip.py index bcf2b0734..68264284a 100644 --- a/modules/aesthetic_clip.py +++ b/modules/aesthetic_clip.py @@ -8,7 +8,7 @@ import gradio as gr import torch from PIL import Image from modules import shared -from modules.shared import device, aesthetic_embeddings +from modules.shared import device from transformers import CLIPModel, CLIPProcessor from tqdm.auto import tqdm @@ -20,7 +20,7 @@ def get_all_images_in_folder(folder): def check_is_valid_image_file(filename): - return filename.lower().endswith(('.png', '.jpg', '.jpeg')) + return filename.lower().endswith(('.png', '.jpg', '.jpeg', ".gif", ".tiff", ".webp")) def batched(dataset, total, n=1): @@ -73,6 +73,6 @@ def generate_imgs_embd(name, folder, batch_size): Aesthetic embedding saved to {html.escape(path)} """ shared.update_aesthetic_embeddings() - return gr.Dropdown(sorted(aesthetic_embeddings.keys()), label="Imgs embedding", - value=sorted(aesthetic_embeddings.keys())[0] if len( - aesthetic_embeddings) > 0 else None), res, "" + return gr.Dropdown.update(choices=sorted(shared.aesthetic_embeddings.keys()), label="Imgs embedding", + value=sorted(shared.aesthetic_embeddings.keys())[0] if len( + shared.aesthetic_embeddings) > 0 else None), res, "" From 9325c85f780c569d1823e422eaf51b2e497e0d3e Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Sun, 16 Oct 2022 00:23:47 +0200 Subject: [PATCH 10/82] fixed dropbox update --- modules/sd_hijack.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py index 192883b2d..491312b42 100644 --- a/modules/sd_hijack.py +++ b/modules/sd_hijack.py @@ -9,7 +9,7 @@ from torch.nn.functional import silu import modules.textual_inversion.textual_inversion from modules import prompt_parser, devices, sd_hijack_optimizations, shared -from modules.shared import opts, device, cmd_opts, aesthetic_embeddings +from modules.shared import opts, device, cmd_opts from modules.sd_hijack_optimizations import invokeAI_mps_available import ldm.modules.attention @@ -182,7 +182,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): image_embs_name = None if image_embs_name is not None and self.image_embs_name != image_embs_name: self.image_embs_name = image_embs_name - self.image_embs = torch.load(aesthetic_embeddings[self.image_embs_name], map_location=device) + self.image_embs = torch.load(shared.aesthetic_embeddings[self.image_embs_name], map_location=device) self.image_embs /= self.image_embs.norm(dim=-1, keepdim=True) self.image_embs.requires_grad_(False) From 523140d7805c644700009b8a2483ff4eb4a22304 Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Sun, 16 Oct 2022 10:23:30 +0200 Subject: [PATCH 11/82] ui fix --- modules/aesthetic_clip.py | 3 +-- modules/sd_hijack.py | 3 +-- modules/shared.py | 2 ++ modules/ui.py | 24 ++++++++++++++---------- 4 files changed, 18 insertions(+), 14 deletions(-) diff --git a/modules/aesthetic_clip.py b/modules/aesthetic_clip.py index 68264284a..ccb35c73f 100644 --- a/modules/aesthetic_clip.py +++ b/modules/aesthetic_clip.py @@ -74,5 +74,4 @@ def generate_imgs_embd(name, folder, batch_size): """ shared.update_aesthetic_embeddings() return gr.Dropdown.update(choices=sorted(shared.aesthetic_embeddings.keys()), label="Imgs embedding", - value=sorted(shared.aesthetic_embeddings.keys())[0] if len( - shared.aesthetic_embeddings) > 0 else None), res, "" + value="None"), res, "" diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py index 01fcb78fb..2de2eed5e 100644 --- a/modules/sd_hijack.py +++ b/modules/sd_hijack.py @@ -392,8 +392,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): z1 = self.process_tokens(tokens, multipliers) z = z1 if z is None else torch.cat((z, z1), axis=-2) - if len(text[ - 0]) != 0 and self.aesthetic_steps != 0 and self.aesthetic_lr != 0 and self.aesthetic_weight != 0 and self.image_embs_name != None: + if self.aesthetic_steps != 0 and self.aesthetic_lr != 0 and self.aesthetic_weight != 0 and self.image_embs_name != None: if not opts.use_old_emphasis_implementation: remade_batch_tokens = [ [self.wrapped.tokenizer.bos_token_id] + x[:75] + [self.wrapped.tokenizer.eos_token_id] for x in diff --git a/modules/shared.py b/modules/shared.py index 3c5ffef1e..e2c98b2dd 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -96,11 +96,13 @@ loaded_hypernetwork = None aesthetic_embeddings = {f.replace(".pt",""): os.path.join(cmd_opts.aesthetic_embeddings_dir, f) for f in os.listdir(cmd_opts.aesthetic_embeddings_dir) if f.endswith(".pt")} +aesthetic_embeddings = aesthetic_embeddings | {"None": None} def update_aesthetic_embeddings(): global aesthetic_embeddings aesthetic_embeddings = {f.replace(".pt",""): os.path.join(cmd_opts.aesthetic_embeddings_dir, f) for f in os.listdir(cmd_opts.aesthetic_embeddings_dir) if f.endswith(".pt")} + aesthetic_embeddings = aesthetic_embeddings | {"None": None} def reload_hypernetworks(): global hypernetworks diff --git a/modules/ui.py b/modules/ui.py index 13ba31420..4069f0d20 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -594,19 +594,23 @@ def create_ui(wrap_gradio_gpu_call): height = gr.Slider(minimum=64, maximum=2048, step=64, label="Height", value=512) with gr.Group(): - aesthetic_lr = gr.Textbox(label='Aesthetic learning rate', placeholder="Aesthetic learning rate", value="0.0001") + with gr.Accordion("Open for Clip Aesthetic!",open=False): + with gr.Row(): + aesthetic_weight = gr.Slider(minimum=0, maximum=1, step=0.01, label="Aesthetic weight", value=0.9) + aesthetic_steps = gr.Slider(minimum=0, maximum=50, step=1, label="Aesthetic steps", value=5) - aesthetic_weight = gr.Slider(minimum=0, maximum=1, step=0.01, label="Aesthetic weight", value=0.9) - aesthetic_steps = gr.Slider(minimum=0, maximum=50, step=1, label="Aesthetic steps", value=5) + with gr.Row(): + aesthetic_lr = gr.Textbox(label='Aesthetic learning rate', placeholder="Aesthetic learning rate", value="0.0001") + aesthetic_slerp = gr.Checkbox(label="Slerp interpolation", value=False) + aesthetic_imgs = gr.Dropdown(sorted(aesthetic_embeddings.keys()), + label="Aesthetic imgs embedding", + value="None") - with gr.Row(): - aesthetic_imgs_text = gr.Textbox(label='Aesthetic text for imgs', placeholder="This text is used to rotate the feature space of the imgs embs", value="") - aesthetic_slerp_angle = gr.Slider(label='Slerp angle',minimum=0, maximum=1, step=0.01, value=0.1) - aesthetic_text_negative = gr.Checkbox(label="Is negative text", value=False) + with gr.Row(): + aesthetic_imgs_text = gr.Textbox(label='Aesthetic text for imgs', placeholder="This text is used to rotate the feature space of the imgs embs", value="") + aesthetic_slerp_angle = gr.Slider(label='Slerp angle',minimum=0, maximum=1, step=0.01, value=0.1) + aesthetic_text_negative = gr.Checkbox(label="Is negative text", value=False) - aesthetic_imgs = gr.Dropdown(sorted(aesthetic_embeddings.keys()), label="Aesthetic imgs embedding", value=sorted(aesthetic_embeddings.keys())[0] if len(aesthetic_embeddings) > 0 else None) - - aesthetic_slerp = gr.Checkbox(label="Slerp interpolation", value=False) with gr.Row(): restore_faces = gr.Checkbox(label='Restore faces', value=False, visible=len(shared.face_restorers) > 1) From e4f8b5f00dd33b7547cc6b76fbed26bb83b37a64 Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Sun, 16 Oct 2022 10:28:21 +0200 Subject: [PATCH 12/82] ui fix --- modules/sd_hijack.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py index 2de2eed5e..5d0590afc 100644 --- a/modules/sd_hijack.py +++ b/modules/sd_hijack.py @@ -178,7 +178,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): self.load_image_embs(image_embs_name) def load_image_embs(self, image_embs_name): - if image_embs_name is None or len(image_embs_name) == 0: + if image_embs_name is None or len(image_embs_name) == 0 or image_embs_name == "None": image_embs_name = None if image_embs_name is not None and self.image_embs_name != image_embs_name: self.image_embs_name = image_embs_name From 9324cdaa3199d65c182858785dd1eca42b192b8e Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Sun, 16 Oct 2022 17:53:56 +0200 Subject: [PATCH 13/82] ui fix, re organization of the code --- modules/aesthetic_clip.py | 154 +++++++++++++++++++++++++-- modules/img2img.py | 14 ++- modules/processing.py | 29 ++--- modules/sd_hijack.py | 102 +----------------- modules/sd_models.py | 5 +- modules/shared.py | 14 ++- modules/textual_inversion/dataset.py | 2 +- modules/txt2img.py | 18 ++-- modules/ui.py | 50 +++++---- 9 files changed, 232 insertions(+), 156 deletions(-) diff --git a/modules/aesthetic_clip.py b/modules/aesthetic_clip.py index ccb35c73f..34efa931b 100644 --- a/modules/aesthetic_clip.py +++ b/modules/aesthetic_clip.py @@ -1,3 +1,4 @@ +import copy import itertools import os from pathlib import Path @@ -7,11 +8,12 @@ import gc import gradio as gr import torch from PIL import Image -from modules import shared -from modules.shared import device -from transformers import CLIPModel, CLIPProcessor +from torch import optim -from tqdm.auto import tqdm +from modules import shared +from transformers import CLIPModel, CLIPProcessor, CLIPTokenizer +from tqdm.auto import tqdm, trange +from modules.shared import opts, device def get_all_images_in_folder(folder): @@ -37,12 +39,39 @@ def iter_to_batched(iterable, n=1): yield chunk +def create_ui(): + with gr.Group(): + with gr.Accordion("Open for Clip Aesthetic!", open=False): + with gr.Row(): + aesthetic_weight = gr.Slider(minimum=0, maximum=1, step=0.01, label="Aesthetic weight", + value=0.9) + aesthetic_steps = gr.Slider(minimum=0, maximum=50, step=1, label="Aesthetic steps", value=5) + + with gr.Row(): + aesthetic_lr = gr.Textbox(label='Aesthetic learning rate', + placeholder="Aesthetic learning rate", value="0.0001") + aesthetic_slerp = gr.Checkbox(label="Slerp interpolation", value=False) + aesthetic_imgs = gr.Dropdown(sorted(shared.aesthetic_embeddings.keys()), + label="Aesthetic imgs embedding", + value="None") + + with gr.Row(): + aesthetic_imgs_text = gr.Textbox(label='Aesthetic text for imgs', + placeholder="This text is used to rotate the feature space of the imgs embs", + value="") + aesthetic_slerp_angle = gr.Slider(label='Slerp angle', minimum=0, maximum=1, step=0.01, + value=0.1) + aesthetic_text_negative = gr.Checkbox(label="Is negative text", value=False) + + return aesthetic_weight, aesthetic_steps, aesthetic_lr, aesthetic_slerp, aesthetic_imgs, aesthetic_imgs_text, aesthetic_slerp_angle, aesthetic_text_negative + + def generate_imgs_embd(name, folder, batch_size): # clipModel = CLIPModel.from_pretrained( # shared.sd_model.cond_stage_model.clipModel.name_or_path # ) - model = CLIPModel.from_pretrained(shared.sd_model.cond_stage_model.clipModel.name_or_path).to(device) - processor = CLIPProcessor.from_pretrained(shared.sd_model.cond_stage_model.clipModel.name_or_path) + model = shared.clip_model.to(device) + processor = CLIPProcessor.from_pretrained(model.name_or_path) with torch.no_grad(): embs = [] @@ -63,7 +92,6 @@ def generate_imgs_embd(name, folder, batch_size): torch.save(embs, path) model = model.cpu() - del model del processor del embs gc.collect() @@ -74,4 +102,114 @@ def generate_imgs_embd(name, folder, batch_size): """ shared.update_aesthetic_embeddings() return gr.Dropdown.update(choices=sorted(shared.aesthetic_embeddings.keys()), label="Imgs embedding", - value="None"), res, "" + value="None"), \ + gr.Dropdown.update(choices=sorted(shared.aesthetic_embeddings.keys()), + label="Imgs embedding", + value="None"), res, "" + + +def slerp(low, high, val): + low_norm = low / torch.norm(low, dim=1, keepdim=True) + high_norm = high / torch.norm(high, dim=1, keepdim=True) + omega = torch.acos((low_norm * high_norm).sum(1)) + so = torch.sin(omega) + res = (torch.sin((1.0 - val) * omega) / so).unsqueeze(1) * low + (torch.sin(val * omega) / so).unsqueeze(1) * high + return res + + +class AestheticCLIP: + def __init__(self): + self.skip = False + self.aesthetic_steps = 0 + self.aesthetic_weight = 0 + self.aesthetic_lr = 0 + self.slerp = False + self.aesthetic_text_negative = "" + self.aesthetic_slerp_angle = 0 + self.aesthetic_imgs_text = "" + + self.image_embs_name = None + self.image_embs = None + self.load_image_embs(None) + + def set_aesthetic_params(self, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, image_embs_name=None, + aesthetic_slerp=True, aesthetic_imgs_text="", + aesthetic_slerp_angle=0.15, + aesthetic_text_negative=False): + self.aesthetic_imgs_text = aesthetic_imgs_text + self.aesthetic_slerp_angle = aesthetic_slerp_angle + self.aesthetic_text_negative = aesthetic_text_negative + self.slerp = aesthetic_slerp + self.aesthetic_lr = aesthetic_lr + self.aesthetic_weight = aesthetic_weight + self.aesthetic_steps = aesthetic_steps + self.load_image_embs(image_embs_name) + + def set_skip(self, skip): + self.skip = skip + + def load_image_embs(self, image_embs_name): + if image_embs_name is None or len(image_embs_name) == 0 or image_embs_name == "None": + image_embs_name = None + self.image_embs_name = None + if image_embs_name is not None and self.image_embs_name != image_embs_name: + self.image_embs_name = image_embs_name + self.image_embs = torch.load(shared.aesthetic_embeddings[self.image_embs_name], map_location=device) + self.image_embs /= self.image_embs.norm(dim=-1, keepdim=True) + self.image_embs.requires_grad_(False) + + def __call__(self, z, remade_batch_tokens): + if not self.skip and self.aesthetic_steps != 0 and self.aesthetic_lr != 0 and self.aesthetic_weight != 0 and self.image_embs_name is not None: + tokenizer = shared.sd_model.cond_stage_model.tokenizer + if not opts.use_old_emphasis_implementation: + remade_batch_tokens = [ + [tokenizer.bos_token_id] + x[:75] + [tokenizer.eos_token_id] for x in + remade_batch_tokens] + + tokens = torch.asarray(remade_batch_tokens).to(device) + + model = copy.deepcopy(shared.clip_model).to(device) + model.requires_grad_(True) + if self.aesthetic_imgs_text is not None and len(self.aesthetic_imgs_text) > 0: + text_embs_2 = model.get_text_features( + **tokenizer([self.aesthetic_imgs_text], padding=True, return_tensors="pt").to(device)) + if self.aesthetic_text_negative: + text_embs_2 = self.image_embs - text_embs_2 + text_embs_2 /= text_embs_2.norm(dim=-1, keepdim=True) + img_embs = slerp(self.image_embs, text_embs_2, self.aesthetic_slerp_angle) + else: + img_embs = self.image_embs + + with torch.enable_grad(): + + # We optimize the model to maximize the similarity + optimizer = optim.Adam( + model.text_model.parameters(), lr=self.aesthetic_lr + ) + + for _ in trange(self.aesthetic_steps, desc="Aesthetic optimization"): + text_embs = model.get_text_features(input_ids=tokens) + text_embs = text_embs / text_embs.norm(dim=-1, keepdim=True) + sim = text_embs @ img_embs.T + loss = -sim + optimizer.zero_grad() + loss.mean().backward() + optimizer.step() + + zn = model.text_model(input_ids=tokens, output_hidden_states=-opts.CLIP_stop_at_last_layers) + if opts.CLIP_stop_at_last_layers > 1: + zn = zn.hidden_states[-opts.CLIP_stop_at_last_layers] + zn = model.text_model.final_layer_norm(zn) + else: + zn = zn.last_hidden_state + model.cpu() + del model + gc.collect() + torch.cuda.empty_cache() + zn = torch.concat([zn[77 * i:77 * (i + 1)] for i in range(max(z.shape[1] // 77, 1))], 1) + if self.slerp: + z = slerp(z, zn, self.aesthetic_weight) + else: + z = z * (1 - self.aesthetic_weight) + zn * self.aesthetic_weight + + return z diff --git a/modules/img2img.py b/modules/img2img.py index 241267745..4ed80c4bf 100644 --- a/modules/img2img.py +++ b/modules/img2img.py @@ -56,7 +56,14 @@ def process_batch(p, input_dir, output_dir, args): processed_image.save(os.path.join(output_dir, filename)) -def img2img(mode: int, prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: str, init_img, init_img_with_mask, init_img_inpaint, init_mask_inpaint, mask_mode, steps: int, sampler_index: int, mask_blur: int, inpainting_fill: int, restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, denoising_strength: float, seed: int, subseed: int, subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool, height: int, width: int, resize_mode: int, inpaint_full_res: bool, inpaint_full_res_padding: int, inpainting_mask_invert: int, img2img_batch_input_dir: str, img2img_batch_output_dir: str, *args): +def img2img(mode: int, prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: str, init_img, init_img_with_mask, init_img_inpaint, init_mask_inpaint, mask_mode, steps: int, sampler_index: int, mask_blur: int, inpainting_fill: int, restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, denoising_strength: float, seed: int, subseed: int, subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool, height: int, width: int, resize_mode: int, inpaint_full_res: bool, inpaint_full_res_padding: int, inpainting_mask_invert: int, img2img_batch_input_dir: str, img2img_batch_output_dir: str, + aesthetic_lr=0, + aesthetic_weight=0, aesthetic_steps=0, + aesthetic_imgs=None, + aesthetic_slerp=False, + aesthetic_imgs_text="", + aesthetic_slerp_angle=0.15, + aesthetic_text_negative=False, *args): is_inpaint = mode == 1 is_batch = mode == 2 @@ -109,6 +116,11 @@ def img2img(mode: int, prompt: str, negative_prompt: str, prompt_style: str, pro inpainting_mask_invert=inpainting_mask_invert, ) + shared.aesthetic_clip.set_aesthetic_params(float(aesthetic_lr), float(aesthetic_weight), int(aesthetic_steps), + aesthetic_imgs, aesthetic_slerp, aesthetic_imgs_text, + aesthetic_slerp_angle, + aesthetic_text_negative) + if shared.cmd_opts.enable_console_prompts: print(f"\nimg2img: {prompt}", file=shared.progress_print_out) diff --git a/modules/processing.py b/modules/processing.py index 1db26c3ea..685f9fcd8 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -146,7 +146,8 @@ class Processed: self.prompt = self.prompt if type(self.prompt) != list else self.prompt[0] self.negative_prompt = self.negative_prompt if type(self.negative_prompt) != list else self.negative_prompt[0] self.seed = int(self.seed if type(self.seed) != list else self.seed[0]) if self.seed is not None else -1 - self.subseed = int(self.subseed if type(self.subseed) != list else self.subseed[0]) if self.subseed is not None else -1 + self.subseed = int( + self.subseed if type(self.subseed) != list else self.subseed[0]) if self.subseed is not None else -1 self.all_prompts = all_prompts or [self.prompt] self.all_seeds = all_seeds or [self.seed] @@ -332,16 +333,9 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration return f"{all_prompts[index]}{negative_prompt_text}\n{generation_params_text}".strip() -def process_images(p: StableDiffusionProcessing, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, - aesthetic_imgs=None, aesthetic_slerp=False, aesthetic_imgs_text="", - aesthetic_slerp_angle=0.15, - aesthetic_text_negative=False) -> Processed: +def process_images(p: StableDiffusionProcessing) -> Processed: """this is the main loop that both txt2img and img2img use; it calls func_init once inside all the scopes and func_sample once per batch""" - aesthetic_lr = float(aesthetic_lr) - aesthetic_weight = float(aesthetic_weight) - aesthetic_steps = int(aesthetic_steps) - if type(p.prompt) == list: assert (len(p.prompt) > 0) else: @@ -417,16 +411,10 @@ def process_images(p: StableDiffusionProcessing, aesthetic_lr=0, aesthetic_weigh # uc = p.sd_model.get_learned_conditioning(len(prompts) * [p.negative_prompt]) # c = p.sd_model.get_learned_conditioning(prompts) with devices.autocast(): - if hasattr(shared.sd_model.cond_stage_model, "set_aesthetic_params"): - shared.sd_model.cond_stage_model.set_aesthetic_params() + shared.aesthetic_clip.set_skip(True) uc = prompt_parser.get_learned_conditioning(shared.sd_model, len(prompts) * [p.negative_prompt], p.steps) - if hasattr(shared.sd_model.cond_stage_model, "set_aesthetic_params"): - shared.sd_model.cond_stage_model.set_aesthetic_params(aesthetic_lr, aesthetic_weight, - aesthetic_steps, aesthetic_imgs, - aesthetic_slerp, aesthetic_imgs_text, - aesthetic_slerp_angle, - aesthetic_text_negative) + shared.aesthetic_clip.set_skip(False) c = prompt_parser.get_multicond_learned_conditioning(shared.sd_model, prompts, p.steps) if len(model_hijack.comments) > 0: @@ -582,7 +570,6 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): self.truncate_x = int(self.firstphase_width - firstphase_width_truncated) // opt_f self.truncate_y = int(self.firstphase_height - firstphase_height_truncated) // opt_f - def sample(self, conditioning, unconditional_conditioning, seeds, subseeds, subseed_strength): self.sampler = sd_samplers.create_sampler_with_index(sd_samplers.samplers, self.sampler_index, self.sd_model) @@ -600,10 +587,12 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): seed_resize_from_w=self.seed_resize_from_w, p=self) samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning) - samples = samples[:, :, self.truncate_y//2:samples.shape[2]-self.truncate_y//2, self.truncate_x//2:samples.shape[3]-self.truncate_x//2] + samples = samples[:, :, self.truncate_y // 2:samples.shape[2] - self.truncate_y // 2, + self.truncate_x // 2:samples.shape[3] - self.truncate_x // 2] if opts.use_scale_latent_for_hires_fix: - samples = torch.nn.functional.interpolate(samples, size=(self.height // opt_f, self.width // opt_f), mode="bilinear") + samples = torch.nn.functional.interpolate(samples, size=(self.height // opt_f, self.width // opt_f), + mode="bilinear") else: decoded_samples = decode_first_stage(self.sd_model, samples) lowres_samples = torch.clamp((decoded_samples + 1.0) / 2.0, min=0.0, max=1.0) diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py index 5d0590afc..227e76705 100644 --- a/modules/sd_hijack.py +++ b/modules/sd_hijack.py @@ -29,8 +29,8 @@ def apply_optimizations(): ldm.modules.diffusionmodules.model.nonlinearity = silu - - if cmd_opts.force_enable_xformers or (cmd_opts.xformers and shared.xformers_available and torch.version.cuda and (6, 0) <= torch.cuda.get_device_capability(shared.device) <= (9, 0)): + if cmd_opts.force_enable_xformers or (cmd_opts.xformers and shared.xformers_available and torch.version.cuda and ( + 6, 0) <= torch.cuda.get_device_capability(shared.device) <= (9, 0)): print("Applying xformers cross attention optimization.") ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.xformers_attnblock_forward @@ -118,33 +118,14 @@ class StableDiffusionModelHijack: return remade_batch_tokens[0], token_count, get_target_prompt_token_count(token_count) -def slerp(low, high, val): - low_norm = low / torch.norm(low, dim=1, keepdim=True) - high_norm = high / torch.norm(high, dim=1, keepdim=True) - omega = torch.acos((low_norm * high_norm).sum(1)) - so = torch.sin(omega) - res = (torch.sin((1.0 - val) * omega) / so).unsqueeze(1) * low + (torch.sin(val * omega) / so).unsqueeze(1) * high - return res - - class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): def __init__(self, wrapped, hijack): super().__init__() self.wrapped = wrapped - self.clipModel = CLIPModel.from_pretrained( - self.wrapped.transformer.name_or_path - ) - del self.clipModel.vision_model - self.tokenizer = CLIPTokenizer.from_pretrained(self.wrapped.transformer.name_or_path) - self.hijack: StableDiffusionModelHijack = hijack - self.tokenizer = wrapped.tokenizer - # self.vision = CLIPVisionModel.from_pretrained(self.wrapped.transformer.name_or_path).eval() - self.image_embs_name = None - self.image_embs = None - self.load_image_embs(None) self.token_mults = {} - + self.hijack: StableDiffusionModelHijack = hijack + self.tokenizer = wrapped.tokenizer self.comma_token = [v for k, v in self.tokenizer.get_vocab().items() if k == ','][0] tokens_with_parens = [(k, v) for k, v in self.tokenizer.get_vocab().items() if @@ -164,28 +145,6 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): if mult != 1.0: self.token_mults[ident] = mult - def set_aesthetic_params(self, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, image_embs_name=None, - aesthetic_slerp=True, aesthetic_imgs_text="", - aesthetic_slerp_angle=0.15, - aesthetic_text_negative=False): - self.aesthetic_imgs_text = aesthetic_imgs_text - self.aesthetic_slerp_angle = aesthetic_slerp_angle - self.aesthetic_text_negative = aesthetic_text_negative - self.slerp = aesthetic_slerp - self.aesthetic_lr = aesthetic_lr - self.aesthetic_weight = aesthetic_weight - self.aesthetic_steps = aesthetic_steps - self.load_image_embs(image_embs_name) - - def load_image_embs(self, image_embs_name): - if image_embs_name is None or len(image_embs_name) == 0 or image_embs_name == "None": - image_embs_name = None - if image_embs_name is not None and self.image_embs_name != image_embs_name: - self.image_embs_name = image_embs_name - self.image_embs = torch.load(shared.aesthetic_embeddings[self.image_embs_name], map_location=device) - self.image_embs /= self.image_embs.norm(dim=-1, keepdim=True) - self.image_embs.requires_grad_(False) - def tokenize_line(self, line, used_custom_terms, hijack_comments): id_end = self.wrapped.tokenizer.eos_token_id @@ -391,58 +350,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): z1 = self.process_tokens(tokens, multipliers) z = z1 if z is None else torch.cat((z, z1), axis=-2) - - if self.aesthetic_steps != 0 and self.aesthetic_lr != 0 and self.aesthetic_weight != 0 and self.image_embs_name != None: - if not opts.use_old_emphasis_implementation: - remade_batch_tokens = [ - [self.wrapped.tokenizer.bos_token_id] + x[:75] + [self.wrapped.tokenizer.eos_token_id] for x in - remade_batch_tokens] - - tokens = torch.asarray(remade_batch_tokens).to(device) - - model = copy.deepcopy(self.clipModel).to(device) - model.requires_grad_(True) - if self.aesthetic_imgs_text is not None and len(self.aesthetic_imgs_text) > 0: - text_embs_2 = model.get_text_features( - **self.tokenizer([self.aesthetic_imgs_text], padding=True, return_tensors="pt").to(device)) - if self.aesthetic_text_negative: - text_embs_2 = self.image_embs - text_embs_2 - text_embs_2 /= text_embs_2.norm(dim=-1, keepdim=True) - img_embs = slerp(self.image_embs, text_embs_2, self.aesthetic_slerp_angle) - else: - img_embs = self.image_embs - - with torch.enable_grad(): - - # We optimize the model to maximize the similarity - optimizer = optim.Adam( - model.text_model.parameters(), lr=self.aesthetic_lr - ) - - for i in trange(self.aesthetic_steps, desc="Aesthetic optimization"): - text_embs = model.get_text_features(input_ids=tokens) - text_embs = text_embs / text_embs.norm(dim=-1, keepdim=True) - sim = text_embs @ img_embs.T - loss = -sim - optimizer.zero_grad() - loss.mean().backward() - optimizer.step() - - zn = model.text_model(input_ids=tokens, output_hidden_states=-opts.CLIP_stop_at_last_layers) - if opts.CLIP_stop_at_last_layers > 1: - zn = zn.hidden_states[-opts.CLIP_stop_at_last_layers] - zn = model.text_model.final_layer_norm(zn) - else: - zn = zn.last_hidden_state - model.cpu() - del model - - zn = torch.concat([zn for i in range(z.shape[1] // 77)], 1) - if self.slerp: - z = slerp(z, zn, self.aesthetic_weight) - else: - z = z * (1 - self.aesthetic_weight) + zn * self.aesthetic_weight - + z = shared.aesthetic_clip(z, remade_batch_tokens) remade_batch_tokens = rem_tokens batch_multipliers = rem_multipliers i += 1 diff --git a/modules/sd_models.py b/modules/sd_models.py index 3aa21ec14..8e4ee435a 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -20,7 +20,7 @@ checkpoints_loaded = collections.OrderedDict() try: # this silences the annoying "Some weights of the model checkpoint were not used when initializing..." message at start. - from transformers import logging + from transformers import logging, CLIPModel logging.set_verbosity_error() except Exception: @@ -196,6 +196,9 @@ def load_model(): sd_hijack.model_hijack.hijack(sd_model) + if shared.clip_model is None or shared.clip_model.transformer.name_or_path != sd_model.cond_stage_model.wrapped.transformer.name_or_path: + shared.clip_model = CLIPModel.from_pretrained(sd_model.cond_stage_model.wrapped.transformer.name_or_path) + sd_model.eval() print(f"Model loaded.") diff --git a/modules/shared.py b/modules/shared.py index e2c98b2dd..e19ca7797 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -3,6 +3,7 @@ import datetime import json import os import sys +from collections import OrderedDict import gradio as gr import tqdm @@ -94,15 +95,15 @@ os.makedirs(cmd_opts.hypernetwork_dir, exist_ok=True) hypernetworks = hypernetwork.list_hypernetworks(cmd_opts.hypernetwork_dir) loaded_hypernetwork = None -aesthetic_embeddings = {f.replace(".pt",""): os.path.join(cmd_opts.aesthetic_embeddings_dir, f) for f in - os.listdir(cmd_opts.aesthetic_embeddings_dir) if f.endswith(".pt")} -aesthetic_embeddings = aesthetic_embeddings | {"None": None} +aesthetic_embeddings = {} def update_aesthetic_embeddings(): global aesthetic_embeddings aesthetic_embeddings = {f.replace(".pt",""): os.path.join(cmd_opts.aesthetic_embeddings_dir, f) for f in os.listdir(cmd_opts.aesthetic_embeddings_dir) if f.endswith(".pt")} - aesthetic_embeddings = aesthetic_embeddings | {"None": None} + aesthetic_embeddings = OrderedDict(**{"None": None}, **aesthetic_embeddings) + +update_aesthetic_embeddings() def reload_hypernetworks(): global hypernetworks @@ -381,6 +382,11 @@ sd_upscalers = [] sd_model = None +clip_model = None + +from modules.aesthetic_clip import AestheticCLIP +aesthetic_clip = AestheticCLIP() + progress_print_out = sys.stdout diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 68ceffe3f..23bb4b6a6 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -49,7 +49,7 @@ class PersonalizedBase(Dataset): print("Preparing dataset...") for path in tqdm.tqdm(self.image_paths): try: - image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.Resampling.BICUBIC) + image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.BICUBIC) except Exception: continue diff --git a/modules/txt2img.py b/modules/txt2img.py index 8f394d05b..6cbc50fca 100644 --- a/modules/txt2img.py +++ b/modules/txt2img.py @@ -1,12 +1,17 @@ import modules.scripts -from modules.processing import StableDiffusionProcessing, Processed, StableDiffusionProcessingTxt2Img, StableDiffusionProcessingImg2Img, process_images +from modules.processing import StableDiffusionProcessing, Processed, StableDiffusionProcessingTxt2Img, \ + StableDiffusionProcessingImg2Img, process_images from modules.shared import opts, cmd_opts import modules.shared as shared import modules.processing as processing from modules.ui import plaintext_to_html -def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: str, steps: int, sampler_index: int, restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, seed: int, subseed: int, subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool, height: int, width: int, enable_hr: bool, denoising_strength: float, firstphase_width: int, firstphase_height: int,aesthetic_lr=0, +def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: str, steps: int, sampler_index: int, + restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, seed: int, subseed: int, + subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool, + height: int, width: int, enable_hr: bool, denoising_strength: float, firstphase_width: int, + firstphase_height: int, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, aesthetic_imgs=None, aesthetic_slerp=False, @@ -41,15 +46,17 @@ def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: firstphase_height=firstphase_height if enable_hr else None, ) + shared.aesthetic_clip.set_aesthetic_params(float(aesthetic_lr), float(aesthetic_weight), int(aesthetic_steps), + aesthetic_imgs, aesthetic_slerp, aesthetic_imgs_text, aesthetic_slerp_angle, + aesthetic_text_negative) + if cmd_opts.enable_console_prompts: print(f"\ntxt2img: {prompt}", file=shared.progress_print_out) processed = modules.scripts.scripts_txt2img.run(p, *args) if processed is None: - processed = process_images(p, aesthetic_lr, aesthetic_weight, aesthetic_steps, aesthetic_imgs, aesthetic_slerp,aesthetic_imgs_text, - aesthetic_slerp_angle, - aesthetic_text_negative) + processed = process_images(p) shared.total_tqdm.clear() @@ -61,4 +68,3 @@ def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: processed.images = [] return processed.images, generation_info_js, plaintext_to_html(processed.info) - diff --git a/modules/ui.py b/modules/ui.py index 4069f0d20..0e5d73f00 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -43,7 +43,7 @@ from modules.images import save_image import modules.textual_inversion.ui import modules.hypernetworks.ui -import modules.aesthetic_clip +import modules.aesthetic_clip as aesthetic_clip import modules.images_history as img_his @@ -593,23 +593,25 @@ def create_ui(wrap_gradio_gpu_call): width = gr.Slider(minimum=64, maximum=2048, step=64, label="Width", value=512) height = gr.Slider(minimum=64, maximum=2048, step=64, label="Height", value=512) - with gr.Group(): - with gr.Accordion("Open for Clip Aesthetic!",open=False): - with gr.Row(): - aesthetic_weight = gr.Slider(minimum=0, maximum=1, step=0.01, label="Aesthetic weight", value=0.9) - aesthetic_steps = gr.Slider(minimum=0, maximum=50, step=1, label="Aesthetic steps", value=5) + # with gr.Group(): + # with gr.Accordion("Open for Clip Aesthetic!",open=False): + # with gr.Row(): + # aesthetic_weight = gr.Slider(minimum=0, maximum=1, step=0.01, label="Aesthetic weight", value=0.9) + # aesthetic_steps = gr.Slider(minimum=0, maximum=50, step=1, label="Aesthetic steps", value=5) + # + # with gr.Row(): + # aesthetic_lr = gr.Textbox(label='Aesthetic learning rate', placeholder="Aesthetic learning rate", value="0.0001") + # aesthetic_slerp = gr.Checkbox(label="Slerp interpolation", value=False) + # aesthetic_imgs = gr.Dropdown(sorted(aesthetic_embeddings.keys()), + # label="Aesthetic imgs embedding", + # value="None") + # + # with gr.Row(): + # aesthetic_imgs_text = gr.Textbox(label='Aesthetic text for imgs', placeholder="This text is used to rotate the feature space of the imgs embs", value="") + # aesthetic_slerp_angle = gr.Slider(label='Slerp angle',minimum=0, maximum=1, step=0.01, value=0.1) + # aesthetic_text_negative = gr.Checkbox(label="Is negative text", value=False) - with gr.Row(): - aesthetic_lr = gr.Textbox(label='Aesthetic learning rate', placeholder="Aesthetic learning rate", value="0.0001") - aesthetic_slerp = gr.Checkbox(label="Slerp interpolation", value=False) - aesthetic_imgs = gr.Dropdown(sorted(aesthetic_embeddings.keys()), - label="Aesthetic imgs embedding", - value="None") - - with gr.Row(): - aesthetic_imgs_text = gr.Textbox(label='Aesthetic text for imgs', placeholder="This text is used to rotate the feature space of the imgs embs", value="") - aesthetic_slerp_angle = gr.Slider(label='Slerp angle',minimum=0, maximum=1, step=0.01, value=0.1) - aesthetic_text_negative = gr.Checkbox(label="Is negative text", value=False) + aesthetic_weight, aesthetic_steps, aesthetic_lr, aesthetic_slerp, aesthetic_imgs, aesthetic_imgs_text, aesthetic_slerp_angle, aesthetic_text_negative = aesthetic_clip.create_ui() with gr.Row(): @@ -840,6 +842,9 @@ def create_ui(wrap_gradio_gpu_call): width = gr.Slider(minimum=64, maximum=2048, step=64, label="Width", value=512) height = gr.Slider(minimum=64, maximum=2048, step=64, label="Height", value=512) + aesthetic_weight_im, aesthetic_steps_im, aesthetic_lr_im, aesthetic_slerp_im, aesthetic_imgs_im, aesthetic_imgs_text_im, aesthetic_slerp_angle_im, aesthetic_text_negative_im = aesthetic_clip.create_ui() + + with gr.Row(): restore_faces = gr.Checkbox(label='Restore faces', value=False, visible=len(shared.face_restorers) > 1) tiling = gr.Checkbox(label='Tiling', value=False) @@ -944,6 +949,14 @@ def create_ui(wrap_gradio_gpu_call): inpainting_mask_invert, img2img_batch_input_dir, img2img_batch_output_dir, + aesthetic_lr_im, + aesthetic_weight_im, + aesthetic_steps_im, + aesthetic_imgs_im, + aesthetic_slerp_im, + aesthetic_imgs_text_im, + aesthetic_slerp_angle_im, + aesthetic_text_negative_im, ] + custom_inputs, outputs=[ img2img_gallery, @@ -1283,7 +1296,7 @@ def create_ui(wrap_gradio_gpu_call): ) create_embedding_ae.click( - fn=modules.aesthetic_clip.generate_imgs_embd, + fn=aesthetic_clip.generate_imgs_embd, inputs=[ new_embedding_name_ae, process_src_ae, @@ -1291,6 +1304,7 @@ def create_ui(wrap_gradio_gpu_call): ], outputs=[ aesthetic_imgs, + aesthetic_imgs_im, ti_output, ti_outcome, ] From ca023f8a459717be19d333513a0e388cf3944e74 Mon Sep 17 00:00:00 2001 From: ClashSAN <98228077+ClashSAN@users.noreply.github.com> Date: Tue, 18 Oct 2022 08:57:05 +0000 Subject: [PATCH 14/82] Update README.md --- README.md | 20 +++++++++++++------- 1 file changed, 13 insertions(+), 7 deletions(-) diff --git a/README.md b/README.md index 859a91b6f..a89593bf1 100644 --- a/README.md +++ b/README.md @@ -11,6 +11,7 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web - One click install and run script (but you still must install python and git) - Outpainting - Inpainting +- Color Sketch - Prompt Matrix - Stable Diffusion Upscale - Attention, specify parts of text that the model should pay more attention to @@ -37,14 +38,14 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web - Interrupt processing at any time - 4GB video card support (also reports of 2GB working) - Correct seeds for batches -- Prompt length validation - - get length of prompt in tokens as you type - - get a warning after generation if some text was truncated +- Live prompt token length validation - Generation parameters - parameters you used to generate images are saved with that image - in PNG chunks for PNG, in EXIF for JPEG - can drag the image to PNG info tab to restore generation parameters and automatically copy them into UI - can be disabled in settings + - drag and drop an image/text-parameters to promptbox +- Read Generation Parameters Button, loads parameters in promptbox to UI - Settings page - Running arbitrary python code from UI (must run with --allow-code to enable) - Mouseover hints for most UI elements @@ -62,7 +63,7 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web - Img2img Alternative - Highres Fix, a convenience option to produce high resolution pictures in one click without usual distortions - Reloading checkpoints on the fly -- Checkpoint Merger, a tab that allows you to merge two checkpoints into one +- Checkpoint Merger, a tab that allows you to merge up to 3 checkpoints into one - [Custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts) with many extensions from community - [Composable-Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/), a way to use multiple prompts at once - separate prompts using uppercase `AND` @@ -70,14 +71,19 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web - No token limit for prompts (original stable diffusion lets you use up to 75 tokens) - DeepDanbooru integration, creates danbooru style tags for anime prompts (add --deepdanbooru to commandline args) - [xformers](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Xformers), major speed increase for select cards: (add --xformers to commandline args) +- History tab: view, direct and delete images conveniently within the UI +- Generate forever option +- Training Tab +- Preprocessing Image Datasets: cropping, mirroring, autotagging using BLIP or deepdanbooru (for anime) + + ## Installation and Running Make sure the required [dependencies](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Dependencies) are met and follow the instructions available for both [NVidia](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs) (recommended) and [AMD](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs) GPUs. -Alternatively, use Google Colab: +Alternatively, use online services(like Google Colab): -- [Colab, maintained by Akaibu](https://colab.research.google.com/drive/1kw3egmSn-KgWsikYvOMjJkVDsPLjEMzl) -- [Colab, original by me, outdated](https://colab.research.google.com/drive/1Iy-xW9t1-OQWhb0hNxueGij8phCyluOh). +- [List of Online Services](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Online-Services) ### Automatic Installation on Windows 1. Install [Python 3.10.6](https://www.python.org/downloads/windows/), checking "Add Python to PATH" From 019a3a88f07766f2d32c32fbe8e41625f28ecb5e Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Wed, 19 Oct 2022 17:15:47 +0100 Subject: [PATCH 15/82] Update ui.py --- modules/ui.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/ui.py b/modules/ui.py index d2e248801..1573ef827 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1247,7 +1247,7 @@ def create_ui(wrap_gradio_gpu_call): run_preprocess = gr.Button(value="Preprocess", variant='primary') with gr.Tab(label="Train"): - gr.HTML(value="

Train an embedding; must specify a directory with a set of 1:1 ratio images

") + gr.HTML(value="

Train an embedding or Hypernetwork; you must specify a directory with a set of 1:1 ratio images
Initial learning rates: 0.005 for an Embedding, 0.00001 for Hypernetwork wiki

") with gr.Row(): train_embedding_name = gr.Dropdown(label='Embedding', elem_id="train_embedding", choices=sorted(sd_hijack.model_hijack.embedding_db.word_embeddings.keys())) create_refresh_button(train_embedding_name, sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings, lambda: {"choices": sorted(sd_hijack.model_hijack.embedding_db.word_embeddings.keys())}, "refresh_train_embedding_name") From 2ce52d32e41fb523d1494f45073fd18496e52d35 Mon Sep 17 00:00:00 2001 From: discus0434 Date: Wed, 19 Oct 2022 16:31:12 +0000 Subject: [PATCH 16/82] fix for #3086 failing to load any previous hypernet --- modules/hypernetworks/hypernetwork.py | 60 +++++++++++++-------------- 1 file changed, 28 insertions(+), 32 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 7d519cd9b..74300122b 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -24,11 +24,10 @@ class HypernetworkModule(torch.nn.Module): def __init__(self, dim, state_dict=None, layer_structure=None, add_layer_norm=False): super().__init__() - if layer_structure is not None: - assert layer_structure[0] == 1, "Multiplier Sequence should start with size 1!" - assert layer_structure[-1] == 1, "Multiplier Sequence should end with size 1!" - else: - layer_structure = parse_layer_structure(dim, state_dict) + + assert layer_structure is not None, "layer_structure mut not be None" + assert layer_structure[0] == 1, "Multiplier Sequence should start with size 1!" + assert layer_structure[-1] == 1, "Multiplier Sequence should end with size 1!" linears = [] for i in range(len(layer_structure) - 1): @@ -39,23 +38,30 @@ class HypernetworkModule(torch.nn.Module): self.linear = torch.nn.Sequential(*linears) if state_dict is not None: - try: - self.load_state_dict(state_dict) - except RuntimeError: - self.try_load_previous(state_dict) + self.fix_old_state_dict(state_dict) + self.load_state_dict(state_dict) else: for layer in self.linear: - layer.weight.data.normal_(mean = 0.0, std = 0.01) + layer.weight.data.normal_(mean=0.0, std=0.01) layer.bias.data.zero_() self.to(devices.device) - def try_load_previous(self, state_dict): - states = self.state_dict() - states['linear.0.bias'].copy_(state_dict['linear1.bias']) - states['linear.0.weight'].copy_(state_dict['linear1.weight']) - states['linear.1.bias'].copy_(state_dict['linear2.bias']) - states['linear.1.weight'].copy_(state_dict['linear2.weight']) + def fix_old_state_dict(self, state_dict): + changes = { + 'linear1.bias': 'linear.0.bias', + 'linear1.weight': 'linear.0.weight', + 'linear2.bias': 'linear.1.bias', + 'linear2.weight': 'linear.1.weight', + } + + for fr, to in changes.items(): + x = state_dict.get(fr, None) + if x is None: + continue + + del state_dict[fr] + state_dict[to] = x def forward(self, x): return x + self.linear(x) * self.multiplier @@ -71,18 +77,6 @@ def apply_strength(value=None): HypernetworkModule.multiplier = value if value is not None else shared.opts.sd_hypernetwork_strength -def parse_layer_structure(dim, state_dict): - i = 0 - layer_structure = [1] - - while (key := "linear.{}.weight".format(i)) in state_dict: - weight = state_dict[key] - layer_structure.append(len(weight) // dim) - i += 1 - - return layer_structure - - class Hypernetwork: filename = None name = None @@ -135,17 +129,18 @@ class Hypernetwork: state_dict = torch.load(filename, map_location='cpu') + self.layer_structure = state_dict.get('layer_structure', [1, 2, 1]) + self.add_layer_norm = state_dict.get('is_layer_norm', False) + for size, sd in state_dict.items(): if type(size) == int: self.layers[size] = ( - HypernetworkModule(size, sd[0], state_dict["layer_structure"], state_dict["is_layer_norm"]), - HypernetworkModule(size, sd[1], state_dict["layer_structure"], state_dict["is_layer_norm"]), + HypernetworkModule(size, sd[0], self.layer_structure, self.add_layer_norm), + HypernetworkModule(size, sd[1], self.layer_structure, self.add_layer_norm), ) self.name = state_dict.get('name', self.name) self.step = state_dict.get('step', 0) - self.layer_structure = state_dict.get('layer_structure', None) - self.add_layer_norm = state_dict.get('is_layer_norm', False) self.sd_checkpoint = state_dict.get('sd_checkpoint', None) self.sd_checkpoint_name = state_dict.get('sd_checkpoint_name', None) @@ -244,6 +239,7 @@ def stack_conds(conds): return torch.stack(conds) + def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): assert hypernetwork_name, 'hypernetwork not selected' From eb7ba4b713ac2fb960ecf6365b1de0c89451e583 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Wed, 19 Oct 2022 19:50:46 +0100 Subject: [PATCH 17/82] update training header text --- modules/ui.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/ui.py b/modules/ui.py index 1573ef827..93c0767cf 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1247,7 +1247,7 @@ def create_ui(wrap_gradio_gpu_call): run_preprocess = gr.Button(value="Preprocess", variant='primary') with gr.Tab(label="Train"): - gr.HTML(value="

Train an embedding or Hypernetwork; you must specify a directory with a set of 1:1 ratio images
Initial learning rates: 0.005 for an Embedding, 0.00001 for Hypernetwork wiki

") + gr.HTML(value="

Train an embedding or Hypernetwork; you must specify a directory with a set of 1:1 ratio images
Initial learning rates: 0.005 for an Embedding, 0.00001 for Hypernetwork [wiki]

") with gr.Row(): train_embedding_name = gr.Dropdown(label='Embedding', elem_id="train_embedding", choices=sorted(sd_hijack.model_hijack.embedding_db.word_embeddings.keys())) create_refresh_button(train_embedding_name, sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings, lambda: {"choices": sorted(sd_hijack.model_hijack.embedding_db.word_embeddings.keys())}, "refresh_train_embedding_name") From 4d663055ded968831ec97f047dfa8e94036cf1c1 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Wed, 19 Oct 2022 20:33:18 +0100 Subject: [PATCH 18/82] update ui with extra training options --- modules/ui.py | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/modules/ui.py b/modules/ui.py index 93c0767cf..cdb9d335f 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1206,6 +1206,7 @@ def create_ui(wrap_gradio_gpu_call): new_embedding_name = gr.Textbox(label="Name") initialization_text = gr.Textbox(label="Initialization text", value="*") nvpt = gr.Slider(label="Number of vectors per token", minimum=1, maximum=75, step=1, value=1) + overwrite_old_embedding = gr.Checkbox(value=False, label="Overwrite Old Embedding") with gr.Row(): with gr.Column(scale=3): @@ -1219,6 +1220,7 @@ def create_ui(wrap_gradio_gpu_call): new_hypernetwork_sizes = gr.CheckboxGroup(label="Modules", value=["768", "320", "640", "1280"], choices=["768", "320", "640", "1280"]) new_hypernetwork_layer_structure = gr.Textbox("1, 2, 1", label="Enter hypernetwork layer structure", placeholder="1st and last digit must be 1. ex:'1, 2, 1'") new_hypernetwork_add_layer_norm = gr.Checkbox(label="Add layer normalization") + overwrite_old_hypernetwork = gr.Checkbox(value=False, label="Overwrite Old Hypernetwork") with gr.Row(): with gr.Column(scale=3): @@ -1247,14 +1249,17 @@ def create_ui(wrap_gradio_gpu_call): run_preprocess = gr.Button(value="Preprocess", variant='primary') with gr.Tab(label="Train"): - gr.HTML(value="

Train an embedding or Hypernetwork; you must specify a directory with a set of 1:1 ratio images
Initial learning rates: 0.005 for an Embedding, 0.00001 for Hypernetwork [wiki]

") + gr.HTML(value="

Train an embedding or Hypernetwork; you must specify a directory with a set of 1:1 ratio images [wiki]

") with gr.Row(): train_embedding_name = gr.Dropdown(label='Embedding', elem_id="train_embedding", choices=sorted(sd_hijack.model_hijack.embedding_db.word_embeddings.keys())) create_refresh_button(train_embedding_name, sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings, lambda: {"choices": sorted(sd_hijack.model_hijack.embedding_db.word_embeddings.keys())}, "refresh_train_embedding_name") with gr.Row(): train_hypernetwork_name = gr.Dropdown(label='Hypernetwork', elem_id="train_hypernetwork", choices=[x for x in shared.hypernetworks.keys()]) create_refresh_button(train_hypernetwork_name, shared.reload_hypernetworks, lambda: {"choices": sorted([x for x in shared.hypernetworks.keys()])}, "refresh_train_hypernetwork_name") - learn_rate = gr.Textbox(label='Learning rate', placeholder="Learning rate", value="0.005") + with gr.Row(): + embedding_learn_rate = gr.Textbox(label='Embedding Learning rate', placeholder="Embedding Learning rate", value="0.005") + hypernetwork_learn_rate = gr.Textbox(label='Hypernetwork Learning rate', placeholder="Hypernetwork Learning rate", value="0.00001") + batch_size = gr.Number(label='Batch size', value=1, precision=0) dataset_directory = gr.Textbox(label='Dataset directory', placeholder="Path to directory with input images") log_directory = gr.Textbox(label='Log directory', placeholder="Path to directory where to write outputs", value="textual_inversion") @@ -1288,6 +1293,7 @@ def create_ui(wrap_gradio_gpu_call): new_embedding_name, initialization_text, nvpt, + overwrite_old_embedding, ], outputs=[ train_embedding_name, @@ -1303,6 +1309,7 @@ def create_ui(wrap_gradio_gpu_call): new_hypernetwork_sizes, new_hypernetwork_layer_structure, new_hypernetwork_add_layer_norm, + overwrite_old_hypernetwork, ], outputs=[ train_hypernetwork_name, From 8e7097d06a6a261580d34375c9d2a9e4ffc63ffa Mon Sep 17 00:00:00 2001 From: random_thoughtss Date: Wed, 19 Oct 2022 13:47:45 -0700 Subject: [PATCH 19/82] Added support for RunwayML inpainting model --- modules/processing.py | 34 +++++- modules/sd_hijack_inpainting.py | 208 ++++++++++++++++++++++++++++++++ modules/sd_models.py | 16 ++- modules/sd_samplers.py | 50 ++++++-- 4 files changed, 293 insertions(+), 15 deletions(-) create mode 100644 modules/sd_hijack_inpainting.py diff --git a/modules/processing.py b/modules/processing.py index bcb0c32c3..a6c308f9e 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -546,7 +546,16 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): if not self.enable_hr: x = create_random_tensors([opt_C, self.height // opt_f, self.width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) - samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning) + + # The "masked-image" in this case will just be all zeros since the entire image is masked. + image_conditioning = torch.zeros(x.shape[0], 3, self.height, self.width, device=x.device) + image_conditioning = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(image_conditioning)) + + # Add the fake full 1s mask to the first dimension. + image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0) + image_conditioning = image_conditioning.to(x.dtype) + + samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning, image_conditioning=image_conditioning) return samples x = create_random_tensors([opt_C, self.firstphase_height // opt_f, self.firstphase_width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) @@ -714,10 +723,31 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): elif self.inpainting_fill == 3: self.init_latent = self.init_latent * self.mask + if self.image_mask is not None: + conditioning_mask = np.array(self.image_mask.convert("L")) + conditioning_mask = conditioning_mask.astype(np.float32) / 255.0 + conditioning_mask = torch.from_numpy(conditioning_mask[None, None]) + + # Inpainting model uses a discretized mask as input, so we round to either 1.0 or 0.0 + conditioning_mask = torch.round(conditioning_mask) + else: + conditioning_mask = torch.ones(1, 1, *image.shape[-2:]) + + # Create another latent image, this time with a masked version of the original input. + conditioning_mask = conditioning_mask.to(image.device) + conditioning_image = image * (1.0 - conditioning_mask) + conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(conditioning_image)) + + # Create the concatenated conditioning tensor to be fed to `c_concat` + conditioning_mask = torch.nn.functional.interpolate(conditioning_mask, size=self.init_latent.shape[-2:]) + conditioning_mask = conditioning_mask.expand(conditioning_image.shape[0], -1, -1, -1) + self.image_conditioning = torch.cat([conditioning_mask, conditioning_image], dim=1) + self.image_conditioning = self.image_conditioning.to(shared.device).type(self.sd_model.dtype) + def sample(self, conditioning, unconditional_conditioning, seeds, subseeds, subseed_strength): x = create_random_tensors([opt_C, self.height // opt_f, self.width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) - samples = self.sampler.sample_img2img(self, self.init_latent, x, conditioning, unconditional_conditioning) + samples = self.sampler.sample_img2img(self, self.init_latent, x, conditioning, unconditional_conditioning, image_conditioning=self.image_conditioning) if self.mask is not None: samples = samples * self.nmask + self.init_latent * self.mask diff --git a/modules/sd_hijack_inpainting.py b/modules/sd_hijack_inpainting.py new file mode 100644 index 000000000..7e5670d66 --- /dev/null +++ b/modules/sd_hijack_inpainting.py @@ -0,0 +1,208 @@ +import torch +import numpy as np + +from tqdm import tqdm +from einops import rearrange, repeat +from omegaconf import ListConfig + +from types import MethodType + +import ldm.models.diffusion.ddpm +import ldm.models.diffusion.ddim + +from ldm.models.diffusion.ddpm import LatentDiffusion +from ldm.models.diffusion.ddim import DDIMSampler, noise_like + +# ================================================================================================= +# Monkey patch DDIMSampler methods from RunwayML repo directly. +# Adapted from: +# https://github.com/runwayml/stable-diffusion/blob/main/ldm/models/diffusion/ddim.py +# ================================================================================================= +@torch.no_grad() +def sample( + self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + ctmp = conditioning[list(conditioning.keys())[0]] + while isinstance(ctmp, list): + ctmp = elf.inpainting_fill == 2: + self.init_latent = self.init_latent * self.mask + create_random_tensors(self.init_latent.shape[1:], all_seeds[0:self.init_latent.shape[0]]) * self.nmask + elif self.inpainting_fill == 3: + self.init_latent = self.init_latent * self.mask + + if self.image_mask is not None: + conditioning_mask = np.array(self.image_mask.convert("L")) + conditioning_mask = conditioning_mask.astype(np.float32) / 255.0 + conditioning_mask = torch.from_numpy(conditioning_mask[None, None]) + + # Inpainting model uses a discretized mask as input, so we round to either 1.0 or 0.0 + conditioning_mask = torch.round(conditioning_mask) + else: + conditioning_mask = torch.ones(1, 1, *image.shape[-2:]) + + # Create another latent image, this time with a masked version of the original input. + conditioning_mask = conditioning_mask.to(image.device) + conditioning_image = image * (1.0 - conditioning_mask) + conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(conditioning_image)) + + # Create the concatenated conditioning tensor to be fed to `c_concat` + conditioning_mask = torch.nn.functional.interpolate(conditioning_mask, size=self.init_latent.shape[-2:]) + conditioning_mask = conditioning_mask.expand(conditioning_image.shape[0], -1, -1, -1) + self.image_conditioning = torch.cat([conditioning_mask, conditioning_image], dim=1) + self.image_conditioning = self.image_conditioning.to(shared.device).type(self.sd_model.dtype) + + def sample(self, conditioning, unconditional_conditioning, seeds, subseeds, subseed_strength): + x = create_random_tensors([opctmp[0] + cbs = ctmp.shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for DDIM sampling is {size}, eta {eta}') + + samples, intermediates = self.ddim_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + ) + return samples, intermediates + + +@torch.no_grad() +def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None): + b, *_, device = *x.shape, x.device + + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + e_t = self.model.apply_model(x, t, c) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t] * 2) + if isinstance(c, dict): + assert isinstance(unconditional_conditioning, dict) + c_in = dict() + for k in c: + if isinstance(c[k], list): + c_in[k] = [ + torch.cat([unconditional_conditioning[k][i], c[k][i]]) + for i in range(len(c[k])) + ] + else: + c_in[k] = torch.cat([unconditional_conditioning[k], c[k]]) + else: + c_in = torch.cat([unconditional_conditioning, c]) + e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2) + e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond) + + if score_corrector is not None: + assert self.model.parameterization == "eps" + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + return x_prev, pred_x0 + + +# ================================================================================================= +# Monkey patch LatentInpaintDiffusion to load the checkpoint with a proper config. +# Adapted from: +# https://github.com/runwayml/stable-diffusion/blob/main/ldm/models/diffusion/ddpm.py +# ================================================================================================= + +@torch.no_grad() +def get_unconditional_conditioning(self, batch_size, null_label=None): + if null_label is not None: + xc = null_label + if isinstance(xc, ListConfig): + xc = list(xc) + if isinstance(xc, dict) or isinstance(xc, list): + c = self.get_learned_conditioning(xc) + else: + if hasattr(xc, "to"): + xc = xc.to(self.device) + c = self.get_learned_conditioning(xc) + else: + # todo: get null label from cond_stage_model + raise NotImplementedError() + c = repeat(c, "1 ... -> b ...", b=batch_size).to(self.device) + return c + +class LatentInpaintDiffusion(LatentDiffusion): + def __init__( + self, + concat_keys=("mask", "masked_image"), + masked_image_key="masked_image", + *args, + **kwargs, + ): + super().__init__(*args, **kwargs) + self.masked_image_key = masked_image_key + assert self.masked_image_key in concat_keys + self.concat_keys = concat_keys + +def should_hijack_inpainting(checkpoint_info): + return str(checkpoint_info.filename).endswith("inpainting.ckpt") and not checkpoint_info.config.endswith("inpainting.yaml") + +def do_inpainting_hijack(): + ldm.models.diffusion.ddpm.get_unconditional_conditioning = get_unconditional_conditioning + ldm.models.diffusion.ddpm.LatentInpaintDiffusion = LatentInpaintDiffusion + ldm.models.diffusion.ddim.DDIMSampler.p_sample_ddim = p_sample_ddim + ldm.models.diffusion.ddim.DDIMSampler.sample = sample \ No newline at end of file diff --git a/modules/sd_models.py b/modules/sd_models.py index eae22e872..47836d256 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -9,6 +9,7 @@ from ldm.util import instantiate_from_config from modules import shared, modelloader, devices from modules.paths import models_path +from modules.sd_hijack_inpainting import do_inpainting_hijack, should_hijack_inpainting model_dir = "Stable-diffusion" model_path = os.path.abspath(os.path.join(models_path, model_dir)) @@ -211,6 +212,19 @@ def load_model(): print(f"Loading config from: {checkpoint_info.config}") sd_config = OmegaConf.load(checkpoint_info.config) + + if should_hijack_inpainting(checkpoint_info): + do_inpainting_hijack() + + # Hardcoded config for now... + sd_config.model.target = "ldm.models.diffusion.ddpm.LatentInpaintDiffusion" + sd_config.model.params.use_ema = False + sd_config.model.params.conditioning_key = "hybrid" + sd_config.model.params.unet_config.params.in_channels = 9 + + # Create a "fake" config with a different name so that we know to unload it when switching models. + checkpoint_info = checkpoint_info._replace(config=checkpoint_info.config.replace(".yaml", "-inpainting.yaml")) + sd_model = instantiate_from_config(sd_config.model) load_model_weights(sd_model, checkpoint_info) @@ -234,7 +248,7 @@ def reload_model_weights(sd_model, info=None): if sd_model.sd_model_checkpoint == checkpoint_info.filename: return - if sd_model.sd_checkpoint_info.config != checkpoint_info.config: + if sd_model.sd_checkpoint_info.config != checkpoint_info.config or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info): checkpoints_loaded.clear() shared.sd_model = load_model() return shared.sd_model diff --git a/modules/sd_samplers.py b/modules/sd_samplers.py index b58e810b3..9d3cf289e 100644 --- a/modules/sd_samplers.py +++ b/modules/sd_samplers.py @@ -136,9 +136,15 @@ class VanillaStableDiffusionSampler: if self.stop_at is not None and self.step > self.stop_at: raise InterruptedException + # Have to unwrap the inpainting conditioning here to perform pre-preocessing + image_conditioning = None + if isinstance(cond, dict): + image_conditioning = cond["c_concat"][0] + cond = cond["c_crossattn"][0] + unconditional_conditioning = unconditional_conditioning["c_crossattn"][0] conds_list, tensor = prompt_parser.reconstruct_multicond_batch(cond, self.step) - unconditional_conditioning = prompt_parser.reconstruct_cond_batch(unconditional_conditioning, self.step) + unconditional_conditioning = prompt_parser.reconstruct_cond_batch(unconditional_conditioning, self.step) assert all([len(conds) == 1 for conds in conds_list]), 'composition via AND is not supported for DDIM/PLMS samplers' cond = tensor @@ -157,6 +163,10 @@ class VanillaStableDiffusionSampler: img_orig = self.sampler.model.q_sample(self.init_latent, ts) x_dec = img_orig * self.mask + self.nmask * x_dec + if image_conditioning is not None: + cond = {"c_concat": [image_conditioning], "c_crossattn": [cond]} + unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]} + res = self.orig_p_sample_ddim(x_dec, cond, ts, unconditional_conditioning=unconditional_conditioning, *args, **kwargs) if self.mask is not None: @@ -182,7 +192,7 @@ class VanillaStableDiffusionSampler: self.mask = p.mask if hasattr(p, 'mask') else None self.nmask = p.nmask if hasattr(p, 'nmask') else None - def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None): + def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None, image_conditioning=None): steps, t_enc = setup_img2img_steps(p, steps) self.initialize(p) @@ -202,7 +212,7 @@ class VanillaStableDiffusionSampler: return samples - def sample(self, p, x, conditioning, unconditional_conditioning, steps=None): + def sample(self, p, x, conditioning, unconditional_conditioning, steps=None, image_conditioning=None): self.initialize(p) self.init_latent = None @@ -210,6 +220,11 @@ class VanillaStableDiffusionSampler: steps = steps or p.steps + # Wrap the conditioning models with additional image conditioning for inpainting model + if image_conditioning is not None: + conditioning = {"c_concat": [image_conditioning], "c_crossattn": [conditioning]} + unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]} + # existing code fails with certain step counts, like 9 try: samples_ddim = self.launch_sampling(steps, lambda: self.sampler.sample(S=steps, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning, x_T=x, eta=self.eta)[0]) @@ -228,7 +243,7 @@ class CFGDenoiser(torch.nn.Module): self.init_latent = None self.step = 0 - def forward(self, x, sigma, uncond, cond, cond_scale): + def forward(self, x, sigma, uncond, cond, cond_scale, image_cond): if state.interrupted or state.skipped: raise InterruptedException @@ -239,28 +254,29 @@ class CFGDenoiser(torch.nn.Module): repeats = [len(conds_list[i]) for i in range(batch_size)] x_in = torch.cat([torch.stack([x[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [x]) + image_cond_in = torch.cat([torch.stack([image_cond[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [image_cond]) sigma_in = torch.cat([torch.stack([sigma[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [sigma]) if tensor.shape[1] == uncond.shape[1]: cond_in = torch.cat([tensor, uncond]) if shared.batch_cond_uncond: - x_out = self.inner_model(x_in, sigma_in, cond=cond_in) + x_out = self.inner_model(x_in, sigma_in, cond={"c_crossattn": [cond_in], "c_concat": [image_cond_in]}) else: x_out = torch.zeros_like(x_in) for batch_offset in range(0, x_out.shape[0], batch_size): a = batch_offset b = a + batch_size - x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=cond_in[a:b]) + x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond={"c_crossattn": [cond_in[a:b]], "c_concat": [image_cond_in[a:b]]}) else: x_out = torch.zeros_like(x_in) batch_size = batch_size*2 if shared.batch_cond_uncond else batch_size for batch_offset in range(0, tensor.shape[0], batch_size): a = batch_offset b = min(a + batch_size, tensor.shape[0]) - x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=tensor[a:b]) + x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond={"c_crossattn": [tensor[a:b]], "c_concat": [image_cond_in[a:b]]}) - x_out[-uncond.shape[0]:] = self.inner_model(x_in[-uncond.shape[0]:], sigma_in[-uncond.shape[0]:], cond=uncond) + x_out[-uncond.shape[0]:] = self.inner_model(x_in[-uncond.shape[0]:], sigma_in[-uncond.shape[0]:], cond={"c_crossattn": [uncond], "c_concat": [image_cond_in[-uncond.shape[0]:]]}) denoised_uncond = x_out[-uncond.shape[0]:] denoised = torch.clone(denoised_uncond) @@ -361,7 +377,7 @@ class KDiffusionSampler: return extra_params_kwargs - def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None): + def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None, image_conditioning=None): steps, t_enc = setup_img2img_steps(p, steps) if p.sampler_noise_scheduler_override: @@ -389,11 +405,16 @@ class KDiffusionSampler: self.model_wrap_cfg.init_latent = x - samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, xi, extra_args={'cond': conditioning, 'uncond': unconditional_conditioning, 'cond_scale': p.cfg_scale}, disable=False, callback=self.callback_state, **extra_params_kwargs)) + samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, xi, extra_args={ + 'cond': conditioning, + 'image_cond': image_conditioning, + 'uncond': unconditional_conditioning, + 'cond_scale': p.cfg_scale + }, disable=False, callback=self.callback_state, **extra_params_kwargs)) return samples - def sample(self, p, x, conditioning, unconditional_conditioning, steps=None): + def sample(self, p, x, conditioning, unconditional_conditioning, steps=None, image_conditioning = None): steps = steps or p.steps if p.sampler_noise_scheduler_override: @@ -414,7 +435,12 @@ class KDiffusionSampler: else: extra_params_kwargs['sigmas'] = sigmas - samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, x, extra_args={'cond': conditioning, 'uncond': unconditional_conditioning, 'cond_scale': p.cfg_scale}, disable=False, callback=self.callback_state, **extra_params_kwargs)) + samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, x, extra_args={ + 'cond': conditioning, + 'image_cond': image_conditioning, + 'uncond': unconditional_conditioning, + 'cond_scale': p.cfg_scale + }, disable=False, callback=self.callback_state, **extra_params_kwargs)) return samples From 0719c10bf1b817364a498ee11b90d30d3d527344 Mon Sep 17 00:00:00 2001 From: random_thoughtss Date: Wed, 19 Oct 2022 13:56:26 -0700 Subject: [PATCH 20/82] Fixed copying mistake --- modules/sd_hijack_inpainting.py | 79 +++++++++++---------------------- 1 file changed, 25 insertions(+), 54 deletions(-) diff --git a/modules/sd_hijack_inpainting.py b/modules/sd_hijack_inpainting.py index 7e5670d66..d4d28d2e7 100644 --- a/modules/sd_hijack_inpainting.py +++ b/modules/sd_hijack_inpainting.py @@ -19,63 +19,35 @@ from ldm.models.diffusion.ddim import DDIMSampler, noise_like # https://github.com/runwayml/stable-diffusion/blob/main/ldm/models/diffusion/ddim.py # ================================================================================================= @torch.no_grad() -def sample( - self, - S, - batch_size, - shape, - conditioning=None, - callback=None, - normals_sequence=None, - img_callback=None, - quantize_x0=False, - eta=0., - mask=None, - x0=None, - temperature=1., - noise_dropout=0., - score_corrector=None, - corrector_kwargs=None, - verbose=True, - x_T=None, - log_every_t=100, - unconditional_guidance_scale=1., - unconditional_conditioning=None, - # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... - **kwargs - ): +def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + **kwargs + ): if conditioning is not None: if isinstance(conditioning, dict): ctmp = conditioning[list(conditioning.keys())[0]] while isinstance(ctmp, list): - ctmp = elf.inpainting_fill == 2: - self.init_latent = self.init_latent * self.mask + create_random_tensors(self.init_latent.shape[1:], all_seeds[0:self.init_latent.shape[0]]) * self.nmask - elif self.inpainting_fill == 3: - self.init_latent = self.init_latent * self.mask - - if self.image_mask is not None: - conditioning_mask = np.array(self.image_mask.convert("L")) - conditioning_mask = conditioning_mask.astype(np.float32) / 255.0 - conditioning_mask = torch.from_numpy(conditioning_mask[None, None]) - - # Inpainting model uses a discretized mask as input, so we round to either 1.0 or 0.0 - conditioning_mask = torch.round(conditioning_mask) - else: - conditioning_mask = torch.ones(1, 1, *image.shape[-2:]) - - # Create another latent image, this time with a masked version of the original input. - conditioning_mask = conditioning_mask.to(image.device) - conditioning_image = image * (1.0 - conditioning_mask) - conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(conditioning_image)) - - # Create the concatenated conditioning tensor to be fed to `c_concat` - conditioning_mask = torch.nn.functional.interpolate(conditioning_mask, size=self.init_latent.shape[-2:]) - conditioning_mask = conditioning_mask.expand(conditioning_image.shape[0], -1, -1, -1) - self.image_conditioning = torch.cat([conditioning_mask, conditioning_image], dim=1) - self.image_conditioning = self.image_conditioning.to(shared.device).type(self.sd_model.dtype) - - def sample(self, conditioning, unconditional_conditioning, seeds, subseeds, subseed_strength): - x = create_random_tensors([opctmp[0] + ctmp = ctmp[0] cbs = ctmp.shape[0] if cbs != batch_size: print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") @@ -106,7 +78,6 @@ def sample( ) return samples, intermediates - @torch.no_grad() def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, From dde9f960727bfe151d418e43685a2881cf580a17 Mon Sep 17 00:00:00 2001 From: random_thoughtss Date: Wed, 19 Oct 2022 14:14:24 -0700 Subject: [PATCH 21/82] added support for ddim img2img --- modules/sd_samplers.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/modules/sd_samplers.py b/modules/sd_samplers.py index 9d3cf289e..d270e4df6 100644 --- a/modules/sd_samplers.py +++ b/modules/sd_samplers.py @@ -208,6 +208,12 @@ class VanillaStableDiffusionSampler: self.init_latent = x self.step = 0 + # Wrap the conditioning models with additional image conditioning for inpainting model + if image_conditioning is not None: + conditioning = {"c_concat": [image_conditioning], "c_crossattn": [conditioning]} + unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]} + + samples = self.launch_sampling(steps, lambda: self.sampler.decode(x1, conditioning, t_enc, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning)) return samples From c418467c03db916c3e5312e6ac4a67365e196dbd Mon Sep 17 00:00:00 2001 From: random_thoughtss Date: Wed, 19 Oct 2022 15:09:43 -0700 Subject: [PATCH 22/82] Don't compute latent mask if were not using it. Also added support for fixed highres_fix generation. --- modules/processing.py | 70 ++++++++++++++++++++++++++---------------- modules/sd_samplers.py | 4 +++ 2 files changed, 48 insertions(+), 26 deletions(-) diff --git a/modules/processing.py b/modules/processing.py index a6c308f9e..684e58332 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -541,12 +541,8 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): self.truncate_y = int(self.firstphase_height - firstphase_height_truncated) // opt_f - def sample(self, conditioning, unconditional_conditioning, seeds, subseeds, subseed_strength): - self.sampler = sd_samplers.create_sampler_with_index(sd_samplers.samplers, self.sampler_index, self.sd_model) - - if not self.enable_hr: - x = create_random_tensors([opt_C, self.height // opt_f, self.width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) - + def create_dummy_mask(self, x): + if self.sampler.conditioning_key in {'hybrid', 'concat'}: # The "masked-image" in this case will just be all zeros since the entire image is masked. image_conditioning = torch.zeros(x.shape[0], 3, self.height, self.width, device=x.device) image_conditioning = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(image_conditioning)) @@ -555,11 +551,23 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0) image_conditioning = image_conditioning.to(x.dtype) - samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning, image_conditioning=image_conditioning) + else: + # Dummy zero conditioning if we're not using inpainting model. + # Still takes up a bit of memory, but no encoder call. + image_conditioning = torch.zeros(x.shape[0], 5, x.shape[-2], x.shape[-1], dtype=x.dtype, device=x.device) + + return image_conditioning + + def sample(self, conditioning, unconditional_conditioning, seeds, subseeds, subseed_strength): + self.sampler = sd_samplers.create_sampler_with_index(sd_samplers.samplers, self.sampler_index, self.sd_model) + + if not self.enable_hr: + x = create_random_tensors([opt_C, self.height // opt_f, self.width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) + samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning, image_conditioning=self.create_dummy_mask(x)) return samples x = create_random_tensors([opt_C, self.firstphase_height // opt_f, self.firstphase_width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) - samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning) + samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning, image_conditioning=self.create_dummy_mask(x)) samples = samples[:, :, self.truncate_y//2:samples.shape[2]-self.truncate_y//2, self.truncate_x//2:samples.shape[3]-self.truncate_x//2] @@ -596,7 +604,7 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): x = None devices.torch_gc() - samples = self.sampler.sample_img2img(self, samples, noise, conditioning, unconditional_conditioning, steps=self.steps) + samples = self.sampler.sample_img2img(self, samples, noise, conditioning, unconditional_conditioning, steps=self.steps, image_conditioning=self.create_dummy_mask(samples)) return samples @@ -723,26 +731,36 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): elif self.inpainting_fill == 3: self.init_latent = self.init_latent * self.mask - if self.image_mask is not None: - conditioning_mask = np.array(self.image_mask.convert("L")) - conditioning_mask = conditioning_mask.astype(np.float32) / 255.0 - conditioning_mask = torch.from_numpy(conditioning_mask[None, None]) + conditioning_key = self.sampler.conditioning_key - # Inpainting model uses a discretized mask as input, so we round to either 1.0 or 0.0 - conditioning_mask = torch.round(conditioning_mask) + if conditioning_key in {'hybrid', 'concat'}: + if self.image_mask is not None: + conditioning_mask = np.array(self.image_mask.convert("L")) + conditioning_mask = conditioning_mask.astype(np.float32) / 255.0 + conditioning_mask = torch.from_numpy(conditioning_mask[None, None]) + + # Inpainting model uses a discretized mask as input, so we round to either 1.0 or 0.0 + conditioning_mask = torch.round(conditioning_mask) + else: + conditioning_mask = torch.ones(1, 1, *image.shape[-2:]) + + # Create another latent image, this time with a masked version of the original input. + conditioning_mask = conditioning_mask.to(image.device) + conditioning_image = image * (1.0 - conditioning_mask) + conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(conditioning_image)) + + # Create the concatenated conditioning tensor to be fed to `c_concat` + conditioning_mask = torch.nn.functional.interpolate(conditioning_mask, size=self.init_latent.shape[-2:]) + conditioning_mask = conditioning_mask.expand(conditioning_image.shape[0], -1, -1, -1) + self.image_conditioning = torch.cat([conditioning_mask, conditioning_image], dim=1) + self.image_conditioning = self.image_conditioning.to(shared.device).type(self.sd_model.dtype) else: - conditioning_mask = torch.ones(1, 1, *image.shape[-2:]) + self.image_conditioning = torch.zeros( + self.init_latent.shape[0], 5, self.init_latent.shape[-2], self.init_latent.shape[-1], + dtype=self.init_latent.dtype, + device=self.init_latent.device + ) - # Create another latent image, this time with a masked version of the original input. - conditioning_mask = conditioning_mask.to(image.device) - conditioning_image = image * (1.0 - conditioning_mask) - conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(conditioning_image)) - - # Create the concatenated conditioning tensor to be fed to `c_concat` - conditioning_mask = torch.nn.functional.interpolate(conditioning_mask, size=self.init_latent.shape[-2:]) - conditioning_mask = conditioning_mask.expand(conditioning_image.shape[0], -1, -1, -1) - self.image_conditioning = torch.cat([conditioning_mask, conditioning_image], dim=1) - self.image_conditioning = self.image_conditioning.to(shared.device).type(self.sd_model.dtype) def sample(self, conditioning, unconditional_conditioning, seeds, subseeds, subseed_strength): x = create_random_tensors([opt_C, self.height // opt_f, self.width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) diff --git a/modules/sd_samplers.py b/modules/sd_samplers.py index d270e4df6..c21be26e8 100644 --- a/modules/sd_samplers.py +++ b/modules/sd_samplers.py @@ -117,6 +117,8 @@ class VanillaStableDiffusionSampler: self.config = None self.last_latent = None + self.conditioning_key = sd_model.model.conditioning_key + def number_of_needed_noises(self, p): return 0 @@ -328,6 +330,8 @@ class KDiffusionSampler: self.config = None self.last_latent = None + self.conditioning_key = sd_model.model.conditioning_key + def callback_state(self, d): step = d['i'] latent = d["denoised"] From d6ea5841374a28f3f6deb73abc251c8f0bcb240f Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Thu, 20 Oct 2022 00:07:57 +0100 Subject: [PATCH 23/82] change html output --- modules/hypernetworks/hypernetwork.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 7d519cd9b..73c1cb801 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -380,7 +380,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log Loss: {mean_loss:.7f}
Step: {hypernetwork.step}
Last prompt: {html.escape(entries[0].cond_text)}
-Last saved embedding: {html.escape(last_saved_file)}
+Last saved hypernetwork: {html.escape(last_saved_file)}
Last saved image: {html.escape(last_saved_image)}

""" From 166be3919b817cee5e702fd01c34afe9081b952c Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Thu, 20 Oct 2022 00:09:40 +0100 Subject: [PATCH 24/82] allow overwrite old hn --- modules/hypernetworks/ui.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py index 08f75f15d..f45345ea0 100644 --- a/modules/hypernetworks/ui.py +++ b/modules/hypernetworks/ui.py @@ -10,9 +10,10 @@ from modules import sd_hijack, shared, devices from modules.hypernetworks import hypernetwork -def create_hypernetwork(name, enable_sizes, layer_structure=None, add_layer_norm=False): +def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None, add_layer_norm=False): fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt") - assert not os.path.exists(fn), f"file {fn} already exists" + if not overwrite_old: + assert not os.path.exists(fn), f"file {fn} already exists" if type(layer_structure) == str: layer_structure = tuple(map(int, re.sub(r'\D', '', layer_structure))) From 0087079c2d487b67b06ffc30f36ce486a74e6318 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Thu, 20 Oct 2022 00:10:59 +0100 Subject: [PATCH 25/82] allow overwrite old embedding --- modules/textual_inversion/textual_inversion.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 3be69562b..5776778b5 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -153,7 +153,7 @@ class EmbeddingDatabase: return None, None -def create_embedding(name, num_vectors_per_token, init_text='*'): +def create_embedding(name, num_vectors_per_token, overwrite_old, init_text='*'): cond_model = shared.sd_model.cond_stage_model embedding_layer = cond_model.wrapped.transformer.text_model.embeddings @@ -165,7 +165,8 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): vec[i] = embedded[i * int(embedded.shape[0]) // num_vectors_per_token] fn = os.path.join(shared.cmd_opts.embeddings_dir, f"{name}.pt") - assert not os.path.exists(fn), f"file {fn} already exists" + if not overwrite_old: + assert not os.path.exists(fn), f"file {fn} already exists" embedding = Embedding(vec, name) embedding.step = 0 From 632e8d660293081cadb145d8062e5aff0a4a8f0d Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Thu, 20 Oct 2022 00:19:40 +0100 Subject: [PATCH 26/82] split learn rates --- modules/ui.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/modules/ui.py b/modules/ui.py index cdb9d335f..d07184eef 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1342,7 +1342,7 @@ def create_ui(wrap_gradio_gpu_call): _js="start_training_textual_inversion", inputs=[ train_embedding_name, - learn_rate, + embedding_learn_rate, batch_size, dataset_directory, log_directory, @@ -1367,7 +1367,7 @@ def create_ui(wrap_gradio_gpu_call): _js="start_training_textual_inversion", inputs=[ train_hypernetwork_name, - learn_rate, + hypernetwork_learn_rate, batch_size, dataset_directory, log_directory, From c3835ec85cbb44fa3c46fa871c622b6fee235c89 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Thu, 20 Oct 2022 00:24:24 +0100 Subject: [PATCH 27/82] pass overwrite old flag --- modules/textual_inversion/ui.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/modules/textual_inversion/ui.py b/modules/textual_inversion/ui.py index 36881e7ad..e712284d1 100644 --- a/modules/textual_inversion/ui.py +++ b/modules/textual_inversion/ui.py @@ -7,8 +7,8 @@ import modules.textual_inversion.preprocess from modules import sd_hijack, shared -def create_embedding(name, initialization_text, nvpt): - filename = modules.textual_inversion.textual_inversion.create_embedding(name, nvpt, init_text=initialization_text) +def create_embedding(name, initialization_text, nvpt, overwrite_old): + filename = modules.textual_inversion.textual_inversion.create_embedding(name, nvpt, overwrite_old, init_text=initialization_text) sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings() From 4d6b9f76a55fd0ac0f72634071032dd9c6efb409 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Thu, 20 Oct 2022 00:27:16 +0100 Subject: [PATCH 28/82] reorder create_hypernetwork params --- modules/ui.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/ui.py b/modules/ui.py index d07184eef..322c082be 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1307,9 +1307,9 @@ def create_ui(wrap_gradio_gpu_call): inputs=[ new_hypernetwork_name, new_hypernetwork_sizes, + overwrite_old_hypernetwork, new_hypernetwork_layer_structure, new_hypernetwork_add_layer_norm, - overwrite_old_hypernetwork, ], outputs=[ train_hypernetwork_name, From fbcce66601994f6ed370db36d9c238840fed6bd2 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Thu, 20 Oct 2022 00:46:54 +0100 Subject: [PATCH 29/82] add existing caption file handling --- modules/textual_inversion/preprocess.py | 32 ++++++++++++++++++------- 1 file changed, 24 insertions(+), 8 deletions(-) diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py index 886cf0c3b..5c43fe13e 100644 --- a/modules/textual_inversion/preprocess.py +++ b/modules/textual_inversion/preprocess.py @@ -48,7 +48,7 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pro shared.state.textinfo = "Preprocessing..." shared.state.job_count = len(files) - def save_pic_with_caption(image, index): + def save_pic_with_caption(image, index, existing_caption=None): caption = "" if process_caption: @@ -66,17 +66,26 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pro basename = f"{index:05}-{subindex[0]}-{filename_part}" image.save(os.path.join(dst, f"{basename}.png")) + if preprocess_txt_action == 'prepend' and existing_caption: + caption = existing_caption + ' ' + caption + elif preprocess_txt_action == 'append' and existing_caption: + caption = caption + ' ' + existing_caption + elif preprocess_txt_action == 'copy' and existing_caption: + caption = existing_caption + + caption = caption.strip() + if len(caption) > 0: with open(os.path.join(dst, f"{basename}.txt"), "w", encoding="utf8") as file: file.write(caption) subindex[0] += 1 - def save_pic(image, index): + def save_pic(image, index, existing_caption=None): save_pic_with_caption(image, index) if process_flip: - save_pic_with_caption(ImageOps.mirror(image), index) + save_pic_with_caption(ImageOps.mirror(image), index, existing_caption=existing_caption) for index, imagefile in enumerate(tqdm.tqdm(files)): subindex = [0] @@ -86,6 +95,13 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pro except Exception: continue + existing_caption = None + + try: + existing_caption = open(os.path.splitext(filename)[0] + '.txt', 'r').read() + except Exception as e: + print(e) + if shared.state.interrupted: break @@ -97,20 +113,20 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pro img = img.resize((width, height * img.height // img.width)) top = img.crop((0, 0, width, height)) - save_pic(top, index) + save_pic(top, index, existing_caption=existing_caption) bot = img.crop((0, img.height - height, width, img.height)) - save_pic(bot, index) + save_pic(bot, index, existing_caption=existing_caption) elif process_split and is_wide: img = img.resize((width * img.width // img.height, height)) left = img.crop((0, 0, width, height)) - save_pic(left, index) + save_pic(left, index, existing_caption=existing_caption) right = img.crop((img.width - width, 0, img.width, height)) - save_pic(right, index) + save_pic(right, index, existing_caption=existing_caption) else: img = images.resize_image(1, img, width, height) - save_pic(img, index) + save_pic(img, index, existing_caption=existing_caption) shared.state.nextjob() From ab353b141df8eee042b0964bcb645015dabf3459 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Thu, 20 Oct 2022 00:48:07 +0100 Subject: [PATCH 30/82] link existing txt option --- modules/ui.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/modules/ui.py b/modules/ui.py index 322c082be..7f52ac0cb 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1234,6 +1234,7 @@ def create_ui(wrap_gradio_gpu_call): process_dst = gr.Textbox(label='Destination directory') process_width = gr.Slider(minimum=64, maximum=2048, step=64, label="Width", value=512) process_height = gr.Slider(minimum=64, maximum=2048, step=64, label="Height", value=512) + preprocess_txt_action = gr.Dropdown(label='Existing Caption txt Action', choices=['ignore', 'copy', 'prepend', 'append']) with gr.Row(): process_flip = gr.Checkbox(label='Create flipped copies') @@ -1326,6 +1327,7 @@ def create_ui(wrap_gradio_gpu_call): process_dst, process_width, process_height, + preprocess_txt_action, process_flip, process_split, process_caption, From 9b65c4ecf4f8eb6187ee721918adebe68e9bc631 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Thu, 20 Oct 2022 00:49:23 +0100 Subject: [PATCH 31/82] pass preprocess_txt_action param --- modules/textual_inversion/preprocess.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py index 5c43fe13e..3713bc897 100644 --- a/modules/textual_inversion/preprocess.py +++ b/modules/textual_inversion/preprocess.py @@ -11,7 +11,7 @@ if cmd_opts.deepdanbooru: import modules.deepbooru as deepbooru -def preprocess(process_src, process_dst, process_width, process_height, process_flip, process_split, process_caption, process_caption_deepbooru=False): +def preprocess(process_src, process_dst, process_width, process_height, preprocess_txt_action, process_flip, process_split, process_caption, process_caption_deepbooru=False): try: if process_caption: shared.interrogator.load() @@ -21,7 +21,7 @@ def preprocess(process_src, process_dst, process_width, process_height, process_ db_opts[deepbooru.OPT_INCLUDE_RANKS] = False deepbooru.create_deepbooru_process(opts.interrogate_deepbooru_score_threshold, db_opts) - preprocess_work(process_src, process_dst, process_width, process_height, process_flip, process_split, process_caption, process_caption_deepbooru) + preprocess_work(process_src, process_dst, process_width, process_height, preprocess_txt_action, process_flip, process_split, process_caption, process_caption_deepbooru) finally: @@ -33,7 +33,7 @@ def preprocess(process_src, process_dst, process_width, process_height, process_ -def preprocess_work(process_src, process_dst, process_width, process_height, process_flip, process_split, process_caption, process_caption_deepbooru=False): +def preprocess_work(process_src, process_dst, process_width, process_height, preprocess_txt_action, process_flip, process_split, process_caption, process_caption_deepbooru=False): width = process_width height = process_height src = os.path.abspath(process_src) From 55d8c6cce6d3aef848b9f194adad2ce53064d8b7 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Thu, 20 Oct 2022 00:53:29 +0100 Subject: [PATCH 32/82] default to ignore existing captions --- modules/ui.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/ui.py b/modules/ui.py index 7f52ac0cb..bd5f1b059 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1234,7 +1234,7 @@ def create_ui(wrap_gradio_gpu_call): process_dst = gr.Textbox(label='Destination directory') process_width = gr.Slider(minimum=64, maximum=2048, step=64, label="Width", value=512) process_height = gr.Slider(minimum=64, maximum=2048, step=64, label="Height", value=512) - preprocess_txt_action = gr.Dropdown(label='Existing Caption txt Action', choices=['ignore', 'copy', 'prepend', 'append']) + preprocess_txt_action = gr.Dropdown(label='Existing Caption txt Action', value="ignore", choices=["ignore", "copy", "prepend", "append"]) with gr.Row(): process_flip = gr.Checkbox(label='Create flipped copies') From 6f98e89486f55b0e4657e96ce640cf1c4675d187 Mon Sep 17 00:00:00 2001 From: discus0434 Date: Thu, 20 Oct 2022 00:10:45 +0000 Subject: [PATCH 33/82] update --- modules/hypernetworks/hypernetwork.py | 29 +++++++++++------ modules/hypernetworks/ui.py | 3 +- modules/ui.py | 45 ++++++++++++++------------- 3 files changed, 45 insertions(+), 32 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 74300122b..7d617680f 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -22,16 +22,20 @@ from modules.textual_inversion.learn_schedule import LearnRateScheduler class HypernetworkModule(torch.nn.Module): multiplier = 1.0 - def __init__(self, dim, state_dict=None, layer_structure=None, add_layer_norm=False): + def __init__(self, dim, state_dict=None, layer_structure=None, add_layer_norm=False, activation_func=None): super().__init__() - assert layer_structure is not None, "layer_structure mut not be None" + assert layer_structure is not None, "layer_structure must not be None" assert layer_structure[0] == 1, "Multiplier Sequence should start with size 1!" assert layer_structure[-1] == 1, "Multiplier Sequence should end with size 1!" linears = [] for i in range(len(layer_structure) - 1): linears.append(torch.nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1]))) + if activation_func == "relu": + linears.append(torch.nn.ReLU()) + if activation_func == "leakyrelu": + linears.append(torch.nn.LeakyReLU()) if add_layer_norm: linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1]))) @@ -42,8 +46,9 @@ class HypernetworkModule(torch.nn.Module): self.load_state_dict(state_dict) else: for layer in self.linear: - layer.weight.data.normal_(mean=0.0, std=0.01) - layer.bias.data.zero_() + if not "ReLU" in layer.__str__(): + layer.weight.data.normal_(mean=0.0, std=0.01) + layer.bias.data.zero_() self.to(devices.device) @@ -69,7 +74,8 @@ class HypernetworkModule(torch.nn.Module): def trainables(self): layer_structure = [] for layer in self.linear: - layer_structure += [layer.weight, layer.bias] + if not "ReLU" in layer.__str__(): + layer_structure += [layer.weight, layer.bias] return layer_structure @@ -81,7 +87,7 @@ class Hypernetwork: filename = None name = None - def __init__(self, name=None, enable_sizes=None, layer_structure=None, add_layer_norm=False): + def __init__(self, name=None, enable_sizes=None, layer_structure=None, add_layer_norm=False, activation_func=None): self.filename = None self.name = name self.layers = {} @@ -90,11 +96,12 @@ class Hypernetwork: self.sd_checkpoint_name = None self.layer_structure = layer_structure self.add_layer_norm = add_layer_norm + self.activation_func = activation_func for size in enable_sizes or []: self.layers[size] = ( - HypernetworkModule(size, None, self.layer_structure, self.add_layer_norm), - HypernetworkModule(size, None, self.layer_structure, self.add_layer_norm), + HypernetworkModule(size, None, self.layer_structure, self.add_layer_norm, self.activation_func), + HypernetworkModule(size, None, self.layer_structure, self.add_layer_norm, self.activation_func), ) def weights(self): @@ -117,6 +124,7 @@ class Hypernetwork: state_dict['name'] = self.name state_dict['layer_structure'] = self.layer_structure state_dict['is_layer_norm'] = self.add_layer_norm + state_dict['activation_func'] = self.activation_func state_dict['sd_checkpoint'] = self.sd_checkpoint state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name @@ -131,12 +139,13 @@ class Hypernetwork: self.layer_structure = state_dict.get('layer_structure', [1, 2, 1]) self.add_layer_norm = state_dict.get('is_layer_norm', False) + self.activation_func = state_dict.get('activation_func', None) for size, sd in state_dict.items(): if type(size) == int: self.layers[size] = ( - HypernetworkModule(size, sd[0], self.layer_structure, self.add_layer_norm), - HypernetworkModule(size, sd[1], self.layer_structure, self.add_layer_norm), + HypernetworkModule(size, sd[0], self.layer_structure, self.add_layer_norm, self.activation_func), + HypernetworkModule(size, sd[1], self.layer_structure, self.add_layer_norm, self.activation_func), ) self.name = state_dict.get('name', self.name) diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py index 08f75f15d..83f9547b4 100644 --- a/modules/hypernetworks/ui.py +++ b/modules/hypernetworks/ui.py @@ -10,7 +10,7 @@ from modules import sd_hijack, shared, devices from modules.hypernetworks import hypernetwork -def create_hypernetwork(name, enable_sizes, layer_structure=None, add_layer_norm=False): +def create_hypernetwork(name, enable_sizes, layer_structure=None, add_layer_norm=False, activation_func=None): fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt") assert not os.path.exists(fn), f"file {fn} already exists" @@ -22,6 +22,7 @@ def create_hypernetwork(name, enable_sizes, layer_structure=None, add_layer_norm enable_sizes=[int(x) for x in enable_sizes], layer_structure=layer_structure, add_layer_norm=add_layer_norm, + activation_func=activation_func, ) hypernet.save(fn) diff --git a/modules/ui.py b/modules/ui.py index d2e248801..8751fa9c3 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -5,43 +5,44 @@ import json import math import mimetypes import os +import platform import random +import subprocess as sp import sys import tempfile import time import traceback -import platform -import subprocess as sp from functools import partial, reduce +import gradio as gr +import gradio.routes +import gradio.utils import numpy as np +import piexif import torch from PIL import Image, PngImagePlugin -import piexif -import gradio as gr -import gradio.utils -import gradio.routes - -from modules import sd_hijack, sd_models, localization +from modules import localization, sd_hijack, sd_models from modules.paths import script_path -from modules.shared import opts, cmd_opts, restricted_opts +from modules.shared import cmd_opts, opts, restricted_opts + if cmd_opts.deepdanbooru: from modules.deepbooru import get_deepbooru_tags -import modules.shared as shared -from modules.sd_samplers import samplers, samplers_for_img2img -from modules.sd_hijack import model_hijack -import modules.ldsr_model -import modules.scripts -import modules.gfpgan_model + import modules.codeformer_model -import modules.styles import modules.generation_parameters_copypaste -from modules import prompt_parser -from modules.images import save_image -import modules.textual_inversion.ui +import modules.gfpgan_model import modules.hypernetworks.ui import modules.images_history as img_his +import modules.ldsr_model +import modules.scripts +import modules.shared as shared +import modules.styles +import modules.textual_inversion.ui +from modules import prompt_parser +from modules.images import save_image +from modules.sd_hijack import model_hijack +from modules.sd_samplers import samplers, samplers_for_img2img # this is a fix for Windows users. Without it, javascript files will be served with text/html content-type and the browser will not show any UI mimetypes.init() @@ -268,8 +269,8 @@ def calc_time_left(progress, threshold, label, force_display): time_since_start = time.time() - shared.state.time_start eta = (time_since_start/progress) eta_relative = eta-time_since_start - if (eta_relative > threshold and progress > 0.02) or force_display: - return label + time.strftime('%H:%M:%S', time.gmtime(eta_relative)) + if (eta_relative > threshold and progress > 0.02) or force_display: + return label + time.strftime('%H:%M:%S', time.gmtime(eta_relative)) else: return "" @@ -1219,6 +1220,7 @@ def create_ui(wrap_gradio_gpu_call): new_hypernetwork_sizes = gr.CheckboxGroup(label="Modules", value=["768", "320", "640", "1280"], choices=["768", "320", "640", "1280"]) new_hypernetwork_layer_structure = gr.Textbox("1, 2, 1", label="Enter hypernetwork layer structure", placeholder="1st and last digit must be 1. ex:'1, 2, 1'") new_hypernetwork_add_layer_norm = gr.Checkbox(label="Add layer normalization") + new_hypernetwork_activation_func = gr.Dropdown(value="relu", label="Select activation function of hypernetwork", choices=["relu", "leakyrelu"]) with gr.Row(): with gr.Column(scale=3): @@ -1303,6 +1305,7 @@ def create_ui(wrap_gradio_gpu_call): new_hypernetwork_sizes, new_hypernetwork_layer_structure, new_hypernetwork_add_layer_norm, + new_hypernetwork_activation_func, ], outputs=[ train_hypernetwork_name, From ba469343e6a1c6e23e82acf5feb65c6101dacbb2 Mon Sep 17 00:00:00 2001 From: discus0434 Date: Thu, 20 Oct 2022 00:17:04 +0000 Subject: [PATCH 34/82] align ui.py imports with upstream --- modules/ui.py | 37 ++++++++++++++++++------------------- 1 file changed, 18 insertions(+), 19 deletions(-) diff --git a/modules/ui.py b/modules/ui.py index 987b1d7de..913b23b47 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -5,44 +5,43 @@ import json import math import mimetypes import os -import platform import random -import subprocess as sp import sys import tempfile import time import traceback +import platform +import subprocess as sp from functools import partial, reduce -import gradio as gr -import gradio.routes -import gradio.utils import numpy as np -import piexif import torch from PIL import Image, PngImagePlugin +import piexif -from modules import localization, sd_hijack, sd_models +import gradio as gr +import gradio.utils +import gradio.routes + +from modules import sd_hijack, sd_models, localization from modules.paths import script_path -from modules.shared import cmd_opts, opts, restricted_opts - +from modules.shared import opts, cmd_opts, restricted_opts if cmd_opts.deepdanbooru: from modules.deepbooru import get_deepbooru_tags - -import modules.codeformer_model -import modules.generation_parameters_copypaste -import modules.gfpgan_model -import modules.hypernetworks.ui -import modules.images_history as img_his +import modules.shared as shared +from modules.sd_samplers import samplers, samplers_for_img2img +from modules.sd_hijack import model_hijack import modules.ldsr_model import modules.scripts -import modules.shared as shared +import modules.gfpgan_model +import modules.codeformer_model import modules.styles -import modules.textual_inversion.ui +import modules.generation_parameters_copypaste from modules import prompt_parser from modules.images import save_image -from modules.sd_hijack import model_hijack -from modules.sd_samplers import samplers, samplers_for_img2img +import modules.textual_inversion.ui +import modules.hypernetworks.ui +import modules.images_history as img_his # this is a fix for Windows users. Without it, javascript files will be served with text/html content-type and the browser will not show any UI mimetypes.init() From 858462f719c22ca9f24b94a41699653c34b5f4fb Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Thu, 20 Oct 2022 02:57:18 +0100 Subject: [PATCH 35/82] do caption copy for both flips --- modules/textual_inversion/preprocess.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py index 3713bc897..6bba3852b 100644 --- a/modules/textual_inversion/preprocess.py +++ b/modules/textual_inversion/preprocess.py @@ -82,7 +82,7 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pre subindex[0] += 1 def save_pic(image, index, existing_caption=None): - save_pic_with_caption(image, index) + save_pic_with_caption(image, index, existing_caption=existing_caption) if process_flip: save_pic_with_caption(ImageOps.mirror(image), index, existing_caption=existing_caption) From aa7ff2a1972f3865883e10ba28c5414cdebe8e3b Mon Sep 17 00:00:00 2001 From: random_thoughtss Date: Wed, 19 Oct 2022 21:46:13 -0700 Subject: [PATCH 36/82] Fixed non-square highres fix generation --- modules/processing.py | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/modules/processing.py b/modules/processing.py index 684e58332..3caac25e5 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -541,10 +541,13 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): self.truncate_y = int(self.firstphase_height - firstphase_height_truncated) // opt_f - def create_dummy_mask(self, x): + def create_dummy_mask(self, x, first_phase: bool = False): if self.sampler.conditioning_key in {'hybrid', 'concat'}: + height = self.firstphase_height if first_phase else self.height + width = self.firstphase_width if first_phase else self.width + # The "masked-image" in this case will just be all zeros since the entire image is masked. - image_conditioning = torch.zeros(x.shape[0], 3, self.height, self.width, device=x.device) + image_conditioning = torch.zeros(x.shape[0], 3, height, width, device=x.device) image_conditioning = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(image_conditioning)) # Add the fake full 1s mask to the first dimension. @@ -567,7 +570,7 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): return samples x = create_random_tensors([opt_C, self.firstphase_height // opt_f, self.firstphase_width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) - samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning, image_conditioning=self.create_dummy_mask(x)) + samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning, image_conditioning=self.create_dummy_mask(x, first_phase=True)) samples = samples[:, :, self.truncate_y//2:samples.shape[2]-self.truncate_y//2, self.truncate_x//2:samples.shape[3]-self.truncate_x//2] From f8733ad08be08bafb40f4299785590e11f049e96 Mon Sep 17 00:00:00 2001 From: discus0434 Date: Thu, 20 Oct 2022 11:07:37 +0000 Subject: [PATCH 37/82] add linear as a act func (option for doin nothing) --- modules/ui.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/ui.py b/modules/ui.py index 913b23b47..716f14b83 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1224,7 +1224,7 @@ def create_ui(wrap_gradio_gpu_call): new_hypernetwork_sizes = gr.CheckboxGroup(label="Modules", value=["768", "320", "640", "1280"], choices=["768", "320", "640", "1280"]) new_hypernetwork_layer_structure = gr.Textbox("1, 2, 1", label="Enter hypernetwork layer structure", placeholder="1st and last digit must be 1. ex:'1, 2, 1'") new_hypernetwork_add_layer_norm = gr.Checkbox(label="Add layer normalization") - new_hypernetwork_activation_func = gr.Dropdown(value="relu", label="Select activation function of hypernetwork", choices=["relu", "leakyrelu"]) + new_hypernetwork_activation_func = gr.Dropdown(value="relu", label="Select activation function of hypernetwork", choices=["linear", "relu", "leakyrelu"]) with gr.Row(): with gr.Column(scale=3): From 4281f255d5e7c67515d619f53654be59a6fc1e13 Mon Sep 17 00:00:00 2001 From: wywywywy Date: Thu, 20 Oct 2022 15:31:09 +0100 Subject: [PATCH 38/82] Implemented batch count logic to Outpainting mk2 --- scripts/outpainting_mk_2.py | 36 ++++++++++++++++++++++++++---------- 1 file changed, 26 insertions(+), 10 deletions(-) diff --git a/scripts/outpainting_mk_2.py b/scripts/outpainting_mk_2.py index a6468e09a..02e655e91 100644 --- a/scripts/outpainting_mk_2.py +++ b/scripts/outpainting_mk_2.py @@ -242,21 +242,37 @@ class Script(scripts.Script): out = out.crop((0, 0, res_w, res_h)) return out - img = init_image + batch_count = p.n_iter + p.n_iter = 1 + state.job_count = batch_count + all_images = [] - if left > 0: - img = expand(img, left, is_left=True) - if right > 0: - img = expand(img, right, is_right=True) - if up > 0: - img = expand(img, up, is_top=True) - if down > 0: - img = expand(img, down, is_bottom=True) + for i in range(batch_count): + img = init_image + state.job = f"Batch {i + 1} out of {state.job_count}" - res = Processed(p, [img], initial_seed_and_info[0], initial_seed_and_info[1]) + if left > 0: + img = expand(img, left, is_left=True) + if right > 0: + img = expand(img, right, is_right=True) + if up > 0: + img = expand(img, up, is_top=True) + if down > 0: + img = expand(img, down, is_bottom=True) + + all_images.append(img) + + combined_grid_image = images.image_grid(all_images) + if opts.return_grid: + all_images = [combined_grid_image] + all_images + + res = Processed(p, all_images, initial_seed_and_info[0], initial_seed_and_info[1]) if opts.samples_save: images.save_image(img, p.outpath_samples, "", res.seed, p.prompt, opts.grid_format, info=res.info, p=p) + if opts.grid_save: + images.save_image(combined_grid_image, p.outpath_grids, "grid", res.seed, p.prompt, opts.grid_format, info=res.info, short_filename=not opts.grid_extended_filename, grid=True, p=p) + return res From 91efe138b35dda65e83070c14e9eb94f481fe476 Mon Sep 17 00:00:00 2001 From: wywywywy Date: Thu, 20 Oct 2022 16:02:32 +0100 Subject: [PATCH 39/82] Implemented batch_size logic in outpainting_mk2 --- scripts/outpainting_mk_2.py | 106 +++++++++++++++++++----------------- 1 file changed, 57 insertions(+), 49 deletions(-) diff --git a/scripts/outpainting_mk_2.py b/scripts/outpainting_mk_2.py index 02e655e91..0377ab326 100644 --- a/scripts/outpainting_mk_2.py +++ b/scripts/outpainting_mk_2.py @@ -176,50 +176,53 @@ class Script(scripts.Script): state.job_count = (1 if left > 0 else 0) + (1 if right > 0 else 0) + (1 if up > 0 else 0) + (1 if down > 0 else 0) - def expand(init, expand_pixels, is_left=False, is_right=False, is_top=False, is_bottom=False): + def expand(init, count, expand_pixels, is_left=False, is_right=False, is_top=False, is_bottom=False): is_horiz = is_left or is_right is_vert = is_top or is_bottom pixels_horiz = expand_pixels if is_horiz else 0 pixels_vert = expand_pixels if is_vert else 0 - res_w = init.width + pixels_horiz - res_h = init.height + pixels_vert - process_res_w = math.ceil(res_w / 64) * 64 - process_res_h = math.ceil(res_h / 64) * 64 + images_to_process = [] + for n in range(count): + res_w = init[n].width + pixels_horiz + res_h = init[n].height + pixels_vert + process_res_w = math.ceil(res_w / 64) * 64 + process_res_h = math.ceil(res_h / 64) * 64 - img = Image.new("RGB", (process_res_w, process_res_h)) - img.paste(init, (pixels_horiz if is_left else 0, pixels_vert if is_top else 0)) - mask = Image.new("RGB", (process_res_w, process_res_h), "white") - draw = ImageDraw.Draw(mask) - draw.rectangle(( - expand_pixels + mask_blur if is_left else 0, - expand_pixels + mask_blur if is_top else 0, - mask.width - expand_pixels - mask_blur if is_right else res_w, - mask.height - expand_pixels - mask_blur if is_bottom else res_h, - ), fill="black") + img = Image.new("RGB", (process_res_w, process_res_h)) + img.paste(init[n], (pixels_horiz if is_left else 0, pixels_vert if is_top else 0)) + mask = Image.new("RGB", (process_res_w, process_res_h), "white") + draw = ImageDraw.Draw(mask) + draw.rectangle(( + expand_pixels + mask_blur if is_left else 0, + expand_pixels + mask_blur if is_top else 0, + mask.width - expand_pixels - mask_blur if is_right else res_w, + mask.height - expand_pixels - mask_blur if is_bottom else res_h, + ), fill="black") - np_image = (np.asarray(img) / 255.0).astype(np.float64) - np_mask = (np.asarray(mask) / 255.0).astype(np.float64) - noised = get_matched_noise(np_image, np_mask, noise_q, color_variation) - out = Image.fromarray(np.clip(noised * 255., 0., 255.).astype(np.uint8), mode="RGB") + np_image = (np.asarray(img) / 255.0).astype(np.float64) + np_mask = (np.asarray(mask) / 255.0).astype(np.float64) + noised = get_matched_noise(np_image, np_mask, noise_q, color_variation) + out = Image.fromarray(np.clip(noised * 255., 0., 255.).astype(np.uint8), mode="RGB") - target_width = min(process_width, init.width + pixels_horiz) if is_horiz else img.width - target_height = min(process_height, init.height + pixels_vert) if is_vert else img.height + target_width = min(process_width, init[n].width + pixels_horiz) if is_horiz else img.width + target_height = min(process_height, init[n].height + pixels_vert) if is_vert else img.height + p.width = target_width if is_horiz else img.width + p.height = target_height if is_vert else img.height - crop_region = ( - 0 if is_left else out.width - target_width, - 0 if is_top else out.height - target_height, - target_width if is_left else out.width, - target_height if is_top else out.height, - ) + crop_region = ( + 0 if is_left else out.width - target_width, + 0 if is_top else out.height - target_height, + target_width if is_left else out.width, + target_height if is_top else out.height, + ) + mask = mask.crop(crop_region) + p.image_mask = mask - image_to_process = out.crop(crop_region) - mask = mask.crop(crop_region) + image_to_process = out.crop(crop_region) + images_to_process.append(image_to_process) - p.width = target_width if is_horiz else img.width - p.height = target_height if is_vert else img.height - p.init_images = [image_to_process] - p.image_mask = mask + p.init_images = images_to_process latent_mask = Image.new("RGB", (p.width, p.height), "white") draw = ImageDraw.Draw(latent_mask) @@ -232,44 +235,49 @@ class Script(scripts.Script): p.latent_mask = latent_mask proc = process_images(p) - proc_img = proc.images[0] if initial_seed_and_info[0] is None: initial_seed_and_info[0] = proc.seed initial_seed_and_info[1] = proc.info - out.paste(proc_img, (0 if is_left else out.width - proc_img.width, 0 if is_top else out.height - proc_img.height)) - out = out.crop((0, 0, res_w, res_h)) - return out + for proc_img in proc.images: + out.paste(proc_img, (0 if is_left else out.width - proc_img.width, 0 if is_top else out.height - proc_img.height)) + out = out.crop((0, 0, res_w, res_h)) + + return proc.images batch_count = p.n_iter + batch_size = p.batch_size p.n_iter = 1 state.job_count = batch_count - all_images = [] + all_processed_images = [] for i in range(batch_count): - img = init_image - state.job = f"Batch {i + 1} out of {state.job_count}" + imgs = [init_img] * batch_size + state.job = f"Batch {i + 1} out of {batch_count}" if left > 0: - img = expand(img, left, is_left=True) + imgs = expand(imgs, batch_size, left, is_left=True) if right > 0: - img = expand(img, right, is_right=True) + imgs = expand(imgs, batch_size, right, is_right=True) if up > 0: - img = expand(img, up, is_top=True) + imgs = expand(imgs, batch_size, up, is_top=True) if down > 0: - img = expand(img, down, is_bottom=True) + imgs = expand(imgs, batch_size, down, is_bottom=True) - all_images.append(img) + all_processed_images += imgs + + combined_grid_image = images.image_grid(all_processed_images) + all_images = all_processed_images - combined_grid_image = images.image_grid(all_images) if opts.return_grid: - all_images = [combined_grid_image] + all_images - + all_images = [combined_grid_image] + all_processed_images + res = Processed(p, all_images, initial_seed_and_info[0], initial_seed_and_info[1]) if opts.samples_save: - images.save_image(img, p.outpath_samples, "", res.seed, p.prompt, opts.grid_format, info=res.info, p=p) + for img in all_processed_images: + images.save_image(img, p.outpath_samples, "", res.seed, p.prompt, opts.grid_format, info=res.info, p=p) if opts.grid_save: images.save_image(combined_grid_image, p.outpath_grids, "grid", res.seed, p.prompt, opts.grid_format, info=res.info, short_filename=not opts.grid_extended_filename, grid=True, p=p) From 18df060c3e9252f1cf79b494e7173aff4181049a Mon Sep 17 00:00:00 2001 From: wywywywy Date: Thu, 20 Oct 2022 16:16:09 +0100 Subject: [PATCH 40/82] Fixed outpainting_mk2 output cropping --- scripts/outpainting_mk_2.py | 21 +++++++++++---------- 1 file changed, 11 insertions(+), 10 deletions(-) diff --git a/scripts/outpainting_mk_2.py b/scripts/outpainting_mk_2.py index 0377ab326..726417e7e 100644 --- a/scripts/outpainting_mk_2.py +++ b/scripts/outpainting_mk_2.py @@ -183,6 +183,7 @@ class Script(scripts.Script): pixels_vert = expand_pixels if is_vert else 0 images_to_process = [] + output_images = [] for n in range(count): res_w = init[n].width + pixels_horiz res_h = init[n].height + pixels_vert @@ -203,7 +204,7 @@ class Script(scripts.Script): np_image = (np.asarray(img) / 255.0).astype(np.float64) np_mask = (np.asarray(mask) / 255.0).astype(np.float64) noised = get_matched_noise(np_image, np_mask, noise_q, color_variation) - out = Image.fromarray(np.clip(noised * 255., 0., 255.).astype(np.uint8), mode="RGB") + output_images.append(Image.fromarray(np.clip(noised * 255., 0., 255.).astype(np.uint8), mode="RGB")) target_width = min(process_width, init[n].width + pixels_horiz) if is_horiz else img.width target_height = min(process_height, init[n].height + pixels_vert) if is_vert else img.height @@ -211,15 +212,15 @@ class Script(scripts.Script): p.height = target_height if is_vert else img.height crop_region = ( - 0 if is_left else out.width - target_width, - 0 if is_top else out.height - target_height, - target_width if is_left else out.width, - target_height if is_top else out.height, + 0 if is_left else output_images[n].width - target_width, + 0 if is_top else output_images[n].height - target_height, + target_width if is_left else output_images[n].width, + target_height if is_top else output_images[n].height, ) mask = mask.crop(crop_region) p.image_mask = mask - image_to_process = out.crop(crop_region) + image_to_process = output_images[n].crop(crop_region) images_to_process.append(image_to_process) p.init_images = images_to_process @@ -240,11 +241,11 @@ class Script(scripts.Script): initial_seed_and_info[0] = proc.seed initial_seed_and_info[1] = proc.info - for proc_img in proc.images: - out.paste(proc_img, (0 if is_left else out.width - proc_img.width, 0 if is_top else out.height - proc_img.height)) - out = out.crop((0, 0, res_w, res_h)) + for n in range(count): + output_images[n].paste(proc.images[n], (0 if is_left else output_images[n].width - proc.images[n].width, 0 if is_top else output_images[n].height - proc.images[n].height)) + output_images[n] = output_images[n].crop((0, 0, res_w, res_h)) - return proc.images + return output_images batch_count = p.n_iter batch_size = p.batch_size From 92a17a7a4a13fceb3c3e25a2e854b2a7dd6eb5df Mon Sep 17 00:00:00 2001 From: random_thoughtss Date: Thu, 20 Oct 2022 09:45:03 -0700 Subject: [PATCH 41/82] Made dummy latents smaller. Minor code cleanups --- modules/processing.py | 7 ++++--- modules/sd_samplers.py | 6 ++++-- 2 files changed, 8 insertions(+), 5 deletions(-) diff --git a/modules/processing.py b/modules/processing.py index 3caac25e5..539cde38d 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -557,7 +557,8 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): else: # Dummy zero conditioning if we're not using inpainting model. # Still takes up a bit of memory, but no encoder call. - image_conditioning = torch.zeros(x.shape[0], 5, x.shape[-2], x.shape[-1], dtype=x.dtype, device=x.device) + # Pretty sure we can just make this a 1x1 image since its not going to be used besides its batch size. + image_conditioning = torch.zeros(x.shape[0], 5, 1, 1, dtype=x.dtype, device=x.device) return image_conditioning @@ -759,8 +760,8 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): self.image_conditioning = self.image_conditioning.to(shared.device).type(self.sd_model.dtype) else: self.image_conditioning = torch.zeros( - self.init_latent.shape[0], 5, self.init_latent.shape[-2], self.init_latent.shape[-1], - dtype=self.init_latent.dtype, + self.init_latent.shape[0], 5, 1, 1, + dtype=self.init_latent.dtype, device=self.init_latent.device ) diff --git a/modules/sd_samplers.py b/modules/sd_samplers.py index c21be26e8..cc682593c 100644 --- a/modules/sd_samplers.py +++ b/modules/sd_samplers.py @@ -138,7 +138,7 @@ class VanillaStableDiffusionSampler: if self.stop_at is not None and self.step > self.stop_at: raise InterruptedException - # Have to unwrap the inpainting conditioning here to perform pre-preocessing + # Have to unwrap the inpainting conditioning here to perform pre-processing image_conditioning = None if isinstance(cond, dict): image_conditioning = cond["c_concat"][0] @@ -146,7 +146,7 @@ class VanillaStableDiffusionSampler: unconditional_conditioning = unconditional_conditioning["c_crossattn"][0] conds_list, tensor = prompt_parser.reconstruct_multicond_batch(cond, self.step) - unconditional_conditioning = prompt_parser.reconstruct_cond_batch(unconditional_conditioning, self.step) + unconditional_conditioning = prompt_parser.reconstruct_cond_batch(unconditional_conditioning, self.step) assert all([len(conds) == 1 for conds in conds_list]), 'composition via AND is not supported for DDIM/PLMS samplers' cond = tensor @@ -165,6 +165,8 @@ class VanillaStableDiffusionSampler: img_orig = self.sampler.model.q_sample(self.init_latent, ts) x_dec = img_orig * self.mask + self.nmask * x_dec + # Wrap the image conditioning back up since the DDIM code can accept the dict directly. + # Note that they need to be lists because it just concatenates them later. if image_conditioning is not None: cond = {"c_concat": [image_conditioning], "c_crossattn": [cond]} unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]} From d1cb08bfb221cd1b0cfc6078162b4e206ea80a5c Mon Sep 17 00:00:00 2001 From: Vladimir Repin <32306715+mezotaken@users.noreply.github.com> Date: Thu, 20 Oct 2022 22:49:06 +0300 Subject: [PATCH 42/82] fix skip and interrupt for highres. fix option --- modules/processing.py | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/modules/processing.py b/modules/processing.py index bcb0c32c3..6324ca913 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -587,9 +587,7 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): x = None devices.torch_gc() - samples = self.sampler.sample_img2img(self, samples, noise, conditioning, unconditional_conditioning, steps=self.steps) - - return samples + return self.sampler.sample_img2img(self, samples, noise, conditioning, unconditional_conditioning, steps=self.steps) or samples class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): From 708c3a7bd8ce68cbe1aa7c268e5a4b1980affc9f Mon Sep 17 00:00:00 2001 From: random_thoughtss Date: Thu, 20 Oct 2022 13:28:43 -0700 Subject: [PATCH 43/82] Added PLMS hijack and made sure to always replace methods --- modules/sd_hijack_inpainting.py | 163 ++++++++++++++++++++++++++++++-- modules/sd_models.py | 3 +- 2 files changed, 157 insertions(+), 9 deletions(-) diff --git a/modules/sd_hijack_inpainting.py b/modules/sd_hijack_inpainting.py index d4d28d2e7..439380713 100644 --- a/modules/sd_hijack_inpainting.py +++ b/modules/sd_hijack_inpainting.py @@ -1,16 +1,14 @@ import torch -import numpy as np -from tqdm import tqdm -from einops import rearrange, repeat +from einops import repeat from omegaconf import ListConfig -from types import MethodType - import ldm.models.diffusion.ddpm import ldm.models.diffusion.ddim +import ldm.models.diffusion.plms from ldm.models.diffusion.ddpm import LatentDiffusion +from ldm.models.diffusion.plms import PLMSSampler from ldm.models.diffusion.ddim import DDIMSampler, noise_like # ================================================================================================= @@ -19,7 +17,7 @@ from ldm.models.diffusion.ddim import DDIMSampler, noise_like # https://github.com/runwayml/stable-diffusion/blob/main/ldm/models/diffusion/ddim.py # ================================================================================================= @torch.no_grad() -def sample(self, +def sample_ddim(self, S, batch_size, shape, @@ -132,6 +130,153 @@ def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=F return x_prev, pred_x0 +# ================================================================================================= +# Monkey patch PLMSSampler methods. +# This one was not actually patched correctly in the RunwayML repo, but we can replicate the changes. +# Adapted from: +# https://github.com/CompVis/stable-diffusion/blob/main/ldm/models/diffusion/plms.py +# ================================================================================================= +@torch.no_grad() +def sample_plms(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + ctmp = conditioning[list(conditioning.keys())[0]] + while isinstance(ctmp, list): + ctmp = ctmp[0] + cbs = ctmp.shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for PLMS sampling is {size}') + + samples, intermediates = self.plms_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + ) + return samples, intermediates + + +@torch.no_grad() +def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None): + b, *_, device = *x.shape, x.device + + def get_model_output(x, t): + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + e_t = self.model.apply_model(x, t, c) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t] * 2) + + if isinstance(c, dict): + assert isinstance(unconditional_conditioning, dict) + c_in = dict() + for k in c: + if isinstance(c[k], list): + c_in[k] = [ + torch.cat([unconditional_conditioning[k][i], c[k][i]]) + for i in range(len(c[k])) + ] + else: + c_in[k] = torch.cat([unconditional_conditioning[k], c[k]]) + else: + c_in = torch.cat([unconditional_conditioning, c]) + + e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2) + e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond) + + if score_corrector is not None: + assert self.model.parameterization == "eps" + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + return e_t + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + + def get_x_prev_and_pred_x0(e_t, index): + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + return x_prev, pred_x0 + + e_t = get_model_output(x, t) + if len(old_eps) == 0: + # Pseudo Improved Euler (2nd order) + x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index) + e_t_next = get_model_output(x_prev, t_next) + e_t_prime = (e_t + e_t_next) / 2 + elif len(old_eps) == 1: + # 2nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (3 * e_t - old_eps[-1]) / 2 + elif len(old_eps) == 2: + # 3nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12 + elif len(old_eps) >= 3: + # 4nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24 + + x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index) + + return x_prev, pred_x0, e_t + # ================================================================================================= # Monkey patch LatentInpaintDiffusion to load the checkpoint with a proper config. # Adapted from: @@ -175,5 +320,9 @@ def should_hijack_inpainting(checkpoint_info): def do_inpainting_hijack(): ldm.models.diffusion.ddpm.get_unconditional_conditioning = get_unconditional_conditioning ldm.models.diffusion.ddpm.LatentInpaintDiffusion = LatentInpaintDiffusion + ldm.models.diffusion.ddim.DDIMSampler.p_sample_ddim = p_sample_ddim - ldm.models.diffusion.ddim.DDIMSampler.sample = sample \ No newline at end of file + ldm.models.diffusion.ddim.DDIMSampler.sample = sample_ddim + + ldm.models.diffusion.plms.PLMSSampler.p_sample_plms = p_sample_plms + ldm.models.diffusion.plms.PLMSSampler.sample = sample_plms \ No newline at end of file diff --git a/modules/sd_models.py b/modules/sd_models.py index 47836d256..7072db08e 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -214,8 +214,6 @@ def load_model(): sd_config = OmegaConf.load(checkpoint_info.config) if should_hijack_inpainting(checkpoint_info): - do_inpainting_hijack() - # Hardcoded config for now... sd_config.model.target = "ldm.models.diffusion.ddpm.LatentInpaintDiffusion" sd_config.model.params.use_ema = False @@ -225,6 +223,7 @@ def load_model(): # Create a "fake" config with a different name so that we know to unload it when switching models. checkpoint_info = checkpoint_info._replace(config=checkpoint_info.config.replace(".yaml", "-inpainting.yaml")) + do_inpainting_hijack() sd_model = instantiate_from_config(sd_config.model) load_model_weights(sd_model, checkpoint_info) From d23a46ceaa76af2847f11172f32c92665c268b1b Mon Sep 17 00:00:00 2001 From: Vladimir Repin <32306715+mezotaken@users.noreply.github.com> Date: Thu, 20 Oct 2022 23:49:14 +0300 Subject: [PATCH 44/82] Different approach to skip/interrupt with highres fix --- modules/processing.py | 4 +++- modules/sd_samplers.py | 4 ++++ 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/modules/processing.py b/modules/processing.py index 6324ca913..bcb0c32c3 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -587,7 +587,9 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): x = None devices.torch_gc() - return self.sampler.sample_img2img(self, samples, noise, conditioning, unconditional_conditioning, steps=self.steps) or samples + samples = self.sampler.sample_img2img(self, samples, noise, conditioning, unconditional_conditioning, steps=self.steps) + + return samples class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): diff --git a/modules/sd_samplers.py b/modules/sd_samplers.py index b58e810b3..7ff77c010 100644 --- a/modules/sd_samplers.py +++ b/modules/sd_samplers.py @@ -196,6 +196,7 @@ class VanillaStableDiffusionSampler: x1 = self.sampler.stochastic_encode(x, torch.tensor([t_enc] * int(x.shape[0])).to(shared.device), noise=noise) self.init_latent = x + self.last_latent = x self.step = 0 samples = self.launch_sampling(steps, lambda: self.sampler.decode(x1, conditioning, t_enc, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning)) @@ -206,6 +207,7 @@ class VanillaStableDiffusionSampler: self.initialize(p) self.init_latent = None + self.last_latent = x self.step = 0 steps = steps or p.steps @@ -388,6 +390,7 @@ class KDiffusionSampler: extra_params_kwargs['sigmas'] = sigma_sched self.model_wrap_cfg.init_latent = x + self.last_latent = x samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, xi, extra_args={'cond': conditioning, 'uncond': unconditional_conditioning, 'cond_scale': p.cfg_scale}, disable=False, callback=self.callback_state, **extra_params_kwargs)) @@ -414,6 +417,7 @@ class KDiffusionSampler: else: extra_params_kwargs['sigmas'] = sigmas + self.last_latent = x samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, x, extra_args={'cond': conditioning, 'uncond': unconditional_conditioning, 'cond_scale': p.cfg_scale}, disable=False, callback=self.callback_state, **extra_params_kwargs)) return samples From 49533eed9e3aad19e9868ee140708baec4fd44be Mon Sep 17 00:00:00 2001 From: random_thoughtss Date: Thu, 20 Oct 2022 16:01:27 -0700 Subject: [PATCH 45/82] XY grid correctly re-assignes model when config changes --- modules/sd_models.py | 6 +++--- scripts/xy_grid.py | 1 + 2 files changed, 4 insertions(+), 3 deletions(-) diff --git a/modules/sd_models.py b/modules/sd_models.py index 7072db08e..fea84630c 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -204,9 +204,9 @@ def load_model_weights(model, checkpoint_info): model.sd_checkpoint_info = checkpoint_info -def load_model(): +def load_model(checkpoint_info=None): from modules import lowvram, sd_hijack - checkpoint_info = select_checkpoint() + checkpoint_info = checkpoint_info or select_checkpoint() if checkpoint_info.config != shared.cmd_opts.config: print(f"Loading config from: {checkpoint_info.config}") @@ -249,7 +249,7 @@ def reload_model_weights(sd_model, info=None): if sd_model.sd_checkpoint_info.config != checkpoint_info.config or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info): checkpoints_loaded.clear() - shared.sd_model = load_model() + shared.sd_model = load_model(checkpoint_info) return shared.sd_model if shared.cmd_opts.lowvram or shared.cmd_opts.medvram: diff --git a/scripts/xy_grid.py b/scripts/xy_grid.py index 5cca168a1..eff0c942a 100644 --- a/scripts/xy_grid.py +++ b/scripts/xy_grid.py @@ -89,6 +89,7 @@ def apply_checkpoint(p, x, xs): if info is None: raise RuntimeError(f"Unknown checkpoint: {x}") modules.sd_models.reload_model_weights(shared.sd_model, info) + p.sd_model = shared.sd_model def confirm_checkpoints(p, xs): From 1fc278bcc642f720484a77eb169271054d3153b1 Mon Sep 17 00:00:00 2001 From: wywywywy Date: Fri, 21 Oct 2022 02:38:24 +0100 Subject: [PATCH 46/82] Fixed job count & single-output grid --- scripts/outpainting_mk_2.py | 14 +++++--------- 1 file changed, 5 insertions(+), 9 deletions(-) diff --git a/scripts/outpainting_mk_2.py b/scripts/outpainting_mk_2.py index 726417e7e..633dc1198 100644 --- a/scripts/outpainting_mk_2.py +++ b/scripts/outpainting_mk_2.py @@ -172,10 +172,6 @@ class Script(scripts.Script): if down > 0: down = target_h - init_img.height - up - init_image = p.init_images[0] - - state.job_count = (1 if left > 0 else 0) + (1 if right > 0 else 0) + (1 if up > 0 else 0) + (1 if down > 0 else 0) - def expand(init, count, expand_pixels, is_left=False, is_right=False, is_top=False, is_bottom=False): is_horiz = is_left or is_right is_vert = is_top or is_bottom @@ -250,7 +246,7 @@ class Script(scripts.Script): batch_count = p.n_iter batch_size = p.batch_size p.n_iter = 1 - state.job_count = batch_count + state.job_count = batch_count * batch_size * ((1 if left > 0 else 0) + (1 if right > 0 else 0) + (1 if up > 0 else 0) + (1 if down > 0 else 0)) all_processed_images = [] for i in range(batch_count): @@ -268,10 +264,11 @@ class Script(scripts.Script): all_processed_images += imgs - combined_grid_image = images.image_grid(all_processed_images) all_images = all_processed_images - if opts.return_grid: + combined_grid_image = images.image_grid(all_processed_images) + unwanted_grid_because_of_img_count = len(all_processed_images) < 2 and opts.grid_only_if_multiple + if opts.return_grid and not unwanted_grid_because_of_img_count: all_images = [combined_grid_image] + all_processed_images res = Processed(p, all_images, initial_seed_and_info[0], initial_seed_and_info[1]) @@ -280,8 +277,7 @@ class Script(scripts.Script): for img in all_processed_images: images.save_image(img, p.outpath_samples, "", res.seed, p.prompt, opts.grid_format, info=res.info, p=p) - if opts.grid_save: + if opts.grid_save and not unwanted_grid_because_of_img_count: images.save_image(combined_grid_image, p.outpath_grids, "grid", res.seed, p.prompt, opts.grid_format, info=res.info, short_filename=not opts.grid_extended_filename, grid=True, p=p) return res - From 0110429dc4bc004ac56573fe1a6b05cb0123678e Mon Sep 17 00:00:00 2001 From: winterspringsummer Date: Thu, 20 Oct 2022 19:10:32 +0900 Subject: [PATCH 47/82] Fixed path issue while extras batch processing --- modules/extras.py | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-) diff --git a/modules/extras.py b/modules/extras.py index b853fa5b2..f97966249 100644 --- a/modules/extras.py +++ b/modules/extras.py @@ -118,10 +118,14 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_ while len(cached_images) > 2: del cached_images[next(iter(cached_images.keys()))] + + if opts.use_original_name_batch and image_name != None: + basename = os.path.splitext(os.path.basename(image_name))[0] + else: + basename = '' - images.save_image(image, path=outpath, basename="", seed=None, prompt=None, extension=opts.samples_format, info=info, short_filename=True, - no_prompt=True, grid=False, pnginfo_section_name="extras", existing_info=existing_pnginfo, - forced_filename=image_name if opts.use_original_name_batch else None) + images.save_image(image, path=outpath, basename=basename, seed=None, prompt=None, extension=opts.samples_format, info=info, short_filename=True, + no_prompt=True, grid=False, pnginfo_section_name="extras", existing_info=existing_pnginfo, forced_filename=None) if opts.enable_pnginfo: image.info = existing_pnginfo From aacc4c1ecbcba3cef421d8776dc5b4b239df9b42 Mon Sep 17 00:00:00 2001 From: winterspringsummer Date: Thu, 20 Oct 2022 21:31:29 +0900 Subject: [PATCH 48/82] Added try except to extras batch from directory --- modules/extras.py | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-) diff --git a/modules/extras.py b/modules/extras.py index f97966249..0d817cf92 100644 --- a/modules/extras.py +++ b/modules/extras.py @@ -41,7 +41,10 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_ return outputs, "Please select an input directory.", '' image_list = [file for file in [os.path.join(input_dir, x) for x in os.listdir(input_dir)] if os.path.isfile(file)] for img in image_list: - image = Image.open(img) + try: + image = Image.open(img) + except Exception: + continue imageArr.append(image) imageNameArr.append(img) else: @@ -122,10 +125,10 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_ if opts.use_original_name_batch and image_name != None: basename = os.path.splitext(os.path.basename(image_name))[0] else: - basename = '' + basename = None - images.save_image(image, path=outpath, basename=basename, seed=None, prompt=None, extension=opts.samples_format, info=info, short_filename=True, - no_prompt=True, grid=False, pnginfo_section_name="extras", existing_info=existing_pnginfo, forced_filename=None) + images.save_image(image, path=outpath, basename='', seed=None, prompt=None, extension=opts.samples_format, info=info, short_filename=True, + no_prompt=True, grid=False, pnginfo_section_name="extras", existing_info=existing_pnginfo, forced_filename=basename) if opts.enable_pnginfo: image.info = existing_pnginfo From bc16b135b527224545dca555a9d51edb0adcee2d Mon Sep 17 00:00:00 2001 From: winterspringsummer Date: Thu, 20 Oct 2022 21:43:27 +0900 Subject: [PATCH 49/82] Fixed path issue while extras batch processing --- modules/extras.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/modules/extras.py b/modules/extras.py index 0d817cf92..ac85142ca 100644 --- a/modules/extras.py +++ b/modules/extras.py @@ -125,10 +125,10 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_ if opts.use_original_name_batch and image_name != None: basename = os.path.splitext(os.path.basename(image_name))[0] else: - basename = None + basename = '' - images.save_image(image, path=outpath, basename='', seed=None, prompt=None, extension=opts.samples_format, info=info, short_filename=True, - no_prompt=True, grid=False, pnginfo_section_name="extras", existing_info=existing_pnginfo, forced_filename=basename) + images.save_image(image, path=outpath, basename=basename, seed=None, prompt=None, extension=opts.samples_format, info=info, short_filename=True, + no_prompt=True, grid=False, pnginfo_section_name="extras", existing_info=existing_pnginfo, forced_filename=None) if opts.enable_pnginfo: image.info = existing_pnginfo From 991a595686b8d105025d68d0e833d1cbf44cb143 Mon Sep 17 00:00:00 2001 From: winterspringsummer Date: Fri, 21 Oct 2022 09:23:13 +0900 Subject: [PATCH 50/82] sort file list in alphabetical ordering in extras --- modules/extras.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/extras.py b/modules/extras.py index ac85142ca..22c5a1c12 100644 --- a/modules/extras.py +++ b/modules/extras.py @@ -39,7 +39,7 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_ if input_dir == '': return outputs, "Please select an input directory.", '' - image_list = [file for file in [os.path.join(input_dir, x) for x in os.listdir(input_dir)] if os.path.isfile(file)] + image_list = [file for file in [os.path.join(input_dir, x) for x in sorted(os.listdir(input_dir))] if os.path.isfile(file)] for img in image_list: try: image = Image.open(img) From 45872181902ada06267e2de601586d512cf5df1a Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 21 Oct 2022 09:00:39 +0300 Subject: [PATCH 51/82] updated readme and some small stylistic changes to code --- README.md | 1 + modules/processing.py | 14 ++++++-------- modules/sd_hijack_inpainting.py | 3 +++ 3 files changed, 10 insertions(+), 8 deletions(-) diff --git a/README.md b/README.md index 859a91b6f..a98bb00b7 100644 --- a/README.md +++ b/README.md @@ -70,6 +70,7 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web - No token limit for prompts (original stable diffusion lets you use up to 75 tokens) - DeepDanbooru integration, creates danbooru style tags for anime prompts (add --deepdanbooru to commandline args) - [xformers](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Xformers), major speed increase for select cards: (add --xformers to commandline args) +- Support for dedicated [inpainting model](https://github.com/runwayml/stable-diffusion#inpainting-with-stable-diffusion) by RunwayML. ## Installation and Running Make sure the required [dependencies](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Dependencies) are met and follow the instructions available for both [NVidia](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs) (recommended) and [AMD](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs) GPUs. diff --git a/modules/processing.py b/modules/processing.py index 539cde38d..217869685 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -540,11 +540,10 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): self.truncate_x = int(self.firstphase_width - firstphase_width_truncated) // opt_f self.truncate_y = int(self.firstphase_height - firstphase_height_truncated) // opt_f - - def create_dummy_mask(self, x, first_phase: bool = False): + def create_dummy_mask(self, x, width=None, height=None): if self.sampler.conditioning_key in {'hybrid', 'concat'}: - height = self.firstphase_height if first_phase else self.height - width = self.firstphase_width if first_phase else self.width + height = height or self.height + width = width or self.width # The "masked-image" in this case will just be all zeros since the entire image is masked. image_conditioning = torch.zeros(x.shape[0], 3, height, width, device=x.device) @@ -571,7 +570,7 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): return samples x = create_random_tensors([opt_C, self.firstphase_height // opt_f, self.firstphase_width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) - samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning, image_conditioning=self.create_dummy_mask(x, first_phase=True)) + samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning, image_conditioning=self.create_dummy_mask(x, self.firstphase_width, self.firstphase_height)) samples = samples[:, :, self.truncate_y//2:samples.shape[2]-self.truncate_y//2, self.truncate_x//2:samples.shape[3]-self.truncate_x//2] @@ -634,6 +633,7 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): self.inpainting_mask_invert = inpainting_mask_invert self.mask = None self.nmask = None + self.image_conditioning = None def init(self, all_prompts, all_seeds, all_subseeds): self.sampler = sd_samplers.create_sampler_with_index(sd_samplers.samplers_for_img2img, self.sampler_index, self.sd_model) @@ -735,9 +735,7 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): elif self.inpainting_fill == 3: self.init_latent = self.init_latent * self.mask - conditioning_key = self.sampler.conditioning_key - - if conditioning_key in {'hybrid', 'concat'}: + if self.sampler.conditioning_key in {'hybrid', 'concat'}: if self.image_mask is not None: conditioning_mask = np.array(self.image_mask.convert("L")) conditioning_mask = conditioning_mask.astype(np.float32) / 255.0 diff --git a/modules/sd_hijack_inpainting.py b/modules/sd_hijack_inpainting.py index 439380713..fd92a3354 100644 --- a/modules/sd_hijack_inpainting.py +++ b/modules/sd_hijack_inpainting.py @@ -301,6 +301,7 @@ def get_unconditional_conditioning(self, batch_size, null_label=None): c = repeat(c, "1 ... -> b ...", b=batch_size).to(self.device) return c + class LatentInpaintDiffusion(LatentDiffusion): def __init__( self, @@ -314,9 +315,11 @@ class LatentInpaintDiffusion(LatentDiffusion): assert self.masked_image_key in concat_keys self.concat_keys = concat_keys + def should_hijack_inpainting(checkpoint_info): return str(checkpoint_info.filename).endswith("inpainting.ckpt") and not checkpoint_info.config.endswith("inpainting.yaml") + def do_inpainting_hijack(): ldm.models.diffusion.ddpm.get_unconditional_conditioning = get_unconditional_conditioning ldm.models.diffusion.ddpm.LatentInpaintDiffusion = LatentInpaintDiffusion From 74088c2a06a975092806362aede22f82716cb011 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Thu, 20 Oct 2022 08:18:02 +0300 Subject: [PATCH 52/82] allow float sizes for hypernet's layer_structure --- modules/hypernetworks/ui.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py index 08f75f15d..e0741d08d 100644 --- a/modules/hypernetworks/ui.py +++ b/modules/hypernetworks/ui.py @@ -15,7 +15,7 @@ def create_hypernetwork(name, enable_sizes, layer_structure=None, add_layer_norm assert not os.path.exists(fn), f"file {fn} already exists" if type(layer_structure) == str: - layer_structure = tuple(map(int, re.sub(r'\D', '', layer_structure))) + layer_structure = [float(x.strip()) for x in layer_structure.split(",")] hypernet = modules.hypernetworks.hypernetwork.Hypernetwork( name=name, From 60872c5b404114336f9ca0c671ba88fa4a8201c9 Mon Sep 17 00:00:00 2001 From: winterspringsummer Date: Thu, 20 Oct 2022 19:10:32 +0900 Subject: [PATCH 53/82] Fixed path issue while extras batch processing --- modules/extras.py | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-) diff --git a/modules/extras.py b/modules/extras.py index b853fa5b2..f97966249 100644 --- a/modules/extras.py +++ b/modules/extras.py @@ -118,10 +118,14 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_ while len(cached_images) > 2: del cached_images[next(iter(cached_images.keys()))] + + if opts.use_original_name_batch and image_name != None: + basename = os.path.splitext(os.path.basename(image_name))[0] + else: + basename = '' - images.save_image(image, path=outpath, basename="", seed=None, prompt=None, extension=opts.samples_format, info=info, short_filename=True, - no_prompt=True, grid=False, pnginfo_section_name="extras", existing_info=existing_pnginfo, - forced_filename=image_name if opts.use_original_name_batch else None) + images.save_image(image, path=outpath, basename=basename, seed=None, prompt=None, extension=opts.samples_format, info=info, short_filename=True, + no_prompt=True, grid=False, pnginfo_section_name="extras", existing_info=existing_pnginfo, forced_filename=None) if opts.enable_pnginfo: image.info = existing_pnginfo From fb5a8cf0d9ed027ea3aa2e5422c946d8e6e72efe Mon Sep 17 00:00:00 2001 From: winterspringsummer Date: Thu, 20 Oct 2022 21:31:29 +0900 Subject: [PATCH 54/82] Added try except to extras batch from directory --- modules/extras.py | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-) diff --git a/modules/extras.py b/modules/extras.py index f97966249..0d817cf92 100644 --- a/modules/extras.py +++ b/modules/extras.py @@ -41,7 +41,10 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_ return outputs, "Please select an input directory.", '' image_list = [file for file in [os.path.join(input_dir, x) for x in os.listdir(input_dir)] if os.path.isfile(file)] for img in image_list: - image = Image.open(img) + try: + image = Image.open(img) + except Exception: + continue imageArr.append(image) imageNameArr.append(img) else: @@ -122,10 +125,10 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_ if opts.use_original_name_batch and image_name != None: basename = os.path.splitext(os.path.basename(image_name))[0] else: - basename = '' + basename = None - images.save_image(image, path=outpath, basename=basename, seed=None, prompt=None, extension=opts.samples_format, info=info, short_filename=True, - no_prompt=True, grid=False, pnginfo_section_name="extras", existing_info=existing_pnginfo, forced_filename=None) + images.save_image(image, path=outpath, basename='', seed=None, prompt=None, extension=opts.samples_format, info=info, short_filename=True, + no_prompt=True, grid=False, pnginfo_section_name="extras", existing_info=existing_pnginfo, forced_filename=basename) if opts.enable_pnginfo: image.info = existing_pnginfo From a13c3bed3cec27afe3c015d3d62db36e25b10d1f Mon Sep 17 00:00:00 2001 From: winterspringsummer Date: Thu, 20 Oct 2022 21:43:27 +0900 Subject: [PATCH 55/82] Fixed path issue while extras batch processing --- modules/extras.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/modules/extras.py b/modules/extras.py index 0d817cf92..ac85142ca 100644 --- a/modules/extras.py +++ b/modules/extras.py @@ -125,10 +125,10 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_ if opts.use_original_name_batch and image_name != None: basename = os.path.splitext(os.path.basename(image_name))[0] else: - basename = None + basename = '' - images.save_image(image, path=outpath, basename='', seed=None, prompt=None, extension=opts.samples_format, info=info, short_filename=True, - no_prompt=True, grid=False, pnginfo_section_name="extras", existing_info=existing_pnginfo, forced_filename=basename) + images.save_image(image, path=outpath, basename=basename, seed=None, prompt=None, extension=opts.samples_format, info=info, short_filename=True, + no_prompt=True, grid=False, pnginfo_section_name="extras", existing_info=existing_pnginfo, forced_filename=None) if opts.enable_pnginfo: image.info = existing_pnginfo From 9d71eef02e7395e179b8d5e61e6d91ddd8928d2e Mon Sep 17 00:00:00 2001 From: winterspringsummer Date: Fri, 21 Oct 2022 09:23:13 +0900 Subject: [PATCH 56/82] sort file list in alphabetical ordering in extras --- modules/extras.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/extras.py b/modules/extras.py index ac85142ca..22c5a1c12 100644 --- a/modules/extras.py +++ b/modules/extras.py @@ -39,7 +39,7 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_ if input_dir == '': return outputs, "Please select an input directory.", '' - image_list = [file for file in [os.path.join(input_dir, x) for x in os.listdir(input_dir)] if os.path.isfile(file)] + image_list = [file for file in [os.path.join(input_dir, x) for x in sorted(os.listdir(input_dir))] if os.path.isfile(file)] for img in image_list: try: image = Image.open(img) From c23f666dba2b484d521d2dc4be91cf9e09312647 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 21 Oct 2022 09:47:43 +0300 Subject: [PATCH 57/82] a more strict check for activation type and a more reasonable check for type of layer in hypernets --- modules/hypernetworks/hypernetwork.py | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 7d617680f..84e7e350d 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -32,10 +32,16 @@ class HypernetworkModule(torch.nn.Module): linears = [] for i in range(len(layer_structure) - 1): linears.append(torch.nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1]))) + if activation_func == "relu": linears.append(torch.nn.ReLU()) - if activation_func == "leakyrelu": + elif activation_func == "leakyrelu": linears.append(torch.nn.LeakyReLU()) + elif activation_func == 'linear' or activation_func is None: + pass + else: + raise RuntimeError(f'hypernetwork uses an unsupported activation function: {activation_func}') + if add_layer_norm: linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1]))) @@ -46,7 +52,7 @@ class HypernetworkModule(torch.nn.Module): self.load_state_dict(state_dict) else: for layer in self.linear: - if not "ReLU" in layer.__str__(): + if type(layer) == torch.nn.Linear: layer.weight.data.normal_(mean=0.0, std=0.01) layer.bias.data.zero_() @@ -74,7 +80,7 @@ class HypernetworkModule(torch.nn.Module): def trainables(self): layer_structure = [] for layer in self.linear: - if not "ReLU" in layer.__str__(): + if type(layer) == torch.nn.Linear: layer_structure += [layer.weight, layer.bias] return layer_structure From 5f4fec307c14dd7f817244ffa92e8a4a64abed0b Mon Sep 17 00:00:00 2001 From: Stephen Date: Thu, 20 Oct 2022 11:32:17 -0400 Subject: [PATCH 58/82] [Bugfix][API] - Fix API arg in launch script --- webui.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/webui.py b/webui.py index 177bef744..875890641 100644 --- a/webui.py +++ b/webui.py @@ -118,7 +118,8 @@ def api_only(): api.launch(server_name="0.0.0.0" if cmd_opts.listen else "127.0.0.1", port=cmd_opts.port if cmd_opts.port else 7861) -def webui(launch_api=False): +def webui(): + launch_api = cmd_opts.api initialize() while 1: @@ -158,4 +159,4 @@ if __name__ == "__main__": if cmd_opts.nowebui: api_only() else: - webui(cmd_opts.api) + webui() From 7157e5d064741fa57ca81a2c6432a651f21ee82f Mon Sep 17 00:00:00 2001 From: Patryk Wychowaniec Date: Thu, 20 Oct 2022 19:22:59 +0200 Subject: [PATCH 59/82] interrogate: Fix CLIP-interrogation on CPU Currently, trying to perform CLIP interrogation on a CPU fails, saying: ``` RuntimeError: "slow_conv2d_cpu" not implemented for 'Half' ``` This merge request fixes this issue by detecting whether the target device is CPU and, if so, force-enabling `--no-half` and passing `device="cpu"` to `clip.load()` (which then does some extra tricks to ensure it works correctly on CPU). --- modules/interrogate.py | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-) diff --git a/modules/interrogate.py b/modules/interrogate.py index 64b91eb46..65b05d34c 100644 --- a/modules/interrogate.py +++ b/modules/interrogate.py @@ -28,9 +28,11 @@ class InterrogateModels: clip_preprocess = None categories = None dtype = None + running_on_cpu = None def __init__(self, content_dir): self.categories = [] + self.running_on_cpu = devices.device_interrogate == torch.device("cpu") if os.path.exists(content_dir): for filename in os.listdir(content_dir): @@ -53,7 +55,11 @@ class InterrogateModels: def load_clip_model(self): import clip - model, preprocess = clip.load(clip_model_name) + if self.running_on_cpu: + model, preprocess = clip.load(clip_model_name, device="cpu") + else: + model, preprocess = clip.load(clip_model_name) + model.eval() model = model.to(devices.device_interrogate) @@ -62,14 +68,14 @@ class InterrogateModels: def load(self): if self.blip_model is None: self.blip_model = self.load_blip_model() - if not shared.cmd_opts.no_half: + if not shared.cmd_opts.no_half and not self.running_on_cpu: self.blip_model = self.blip_model.half() self.blip_model = self.blip_model.to(devices.device_interrogate) if self.clip_model is None: self.clip_model, self.clip_preprocess = self.load_clip_model() - if not shared.cmd_opts.no_half: + if not shared.cmd_opts.no_half and not self.running_on_cpu: self.clip_model = self.clip_model.half() self.clip_model = self.clip_model.to(devices.device_interrogate) From b69c37d25e4ffc56e8f8c247fa2c38b4648cefb7 Mon Sep 17 00:00:00 2001 From: guaneec Date: Thu, 20 Oct 2022 22:21:12 +0800 Subject: [PATCH 60/82] Allow datasets with only 1 image in TI --- modules/textual_inversion/dataset.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 23bb4b6a6..5b1c5002f 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -83,7 +83,7 @@ class PersonalizedBase(Dataset): self.dataset.append(entry) - assert len(self.dataset) > 1, "No images have been found in the dataset." + assert len(self.dataset) > 0, "No images have been found in the dataset." self.length = len(self.dataset) * repeats // batch_size self.initial_indexes = np.arange(len(self.dataset)) @@ -91,7 +91,7 @@ class PersonalizedBase(Dataset): self.shuffle() def shuffle(self): - self.indexes = self.initial_indexes[torch.randperm(self.initial_indexes.shape[0])] + self.indexes = self.initial_indexes[torch.randperm(self.initial_indexes.shape[0]).numpy()] def create_text(self, filename_text): text = random.choice(self.lines) From 5245c7a4935f67b677da0f5a1fc2b74c074aa0e2 Mon Sep 17 00:00:00 2001 From: timntorres Date: Wed, 19 Oct 2022 12:21:32 -0700 Subject: [PATCH 61/82] Issue #2921-Give PNG info to Hypernet previews. --- modules/hypernetworks/hypernetwork.py | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 84e7e350d..68c8f26da 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -256,6 +256,9 @@ def stack_conds(conds): def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): + # images is required here to give training previews their infotext. Importing this at the very top causes a circular dependency. + from modules import images + assert hypernetwork_name, 'hypernetwork not selected' path = shared.hypernetworks.get(hypernetwork_name, None) @@ -298,6 +301,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log last_saved_file = "" last_saved_image = "" + forced_filename = "" ititial_step = hypernetwork.step or 0 if ititial_step > steps: @@ -345,7 +349,8 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log }) if hypernetwork.step > 0 and images_dir is not None and hypernetwork.step % create_image_every == 0: - last_saved_image = os.path.join(images_dir, f'{hypernetwork_name}-{hypernetwork.step}.png') + forced_filename = f'{hypernetwork_name}-{hypernetwork.step}' + last_saved_image = os.path.join(images_dir, forced_filename) optimizer.zero_grad() shared.sd_model.cond_stage_model.to(devices.device) @@ -381,7 +386,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log if image is not None: shared.state.current_image = image - image.save(last_saved_image) + last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename) last_saved_image += f", prompt: {preview_text}" shared.state.job_no = hypernetwork.step From 6014fb8afbe05c8d02fffe7a36a2e48128713bd2 Mon Sep 17 00:00:00 2001 From: timntorres Date: Wed, 19 Oct 2022 12:22:23 -0700 Subject: [PATCH 62/82] Do nothing if image file already exists. --- modules/images.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/modules/images.py b/modules/images.py index b9589563a..550e53ae7 100644 --- a/modules/images.py +++ b/modules/images.py @@ -416,7 +416,11 @@ def save_image(image, path, basename, seed=None, prompt=None, extension='png', i dirname = apply_filename_pattern(opts.directories_filename_pattern or "[prompt_words]", p, seed, prompt).strip('\\ /') path = os.path.join(path, dirname) - os.makedirs(path, exist_ok=True) + try: + os.makedirs(path, exist_ok=True) + except FileExistsError: + # If the file already exists, continue and allow said file to be overwritten. + pass if forced_filename is None: basecount = get_next_sequence_number(path, basename) From 4ff274e1e35bb642687253ce744d2cfa738ab293 Mon Sep 17 00:00:00 2001 From: timntorres Date: Wed, 19 Oct 2022 12:32:22 -0700 Subject: [PATCH 63/82] Revise comments. --- modules/hypernetworks/hypernetwork.py | 2 +- modules/images.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 68c8f26da..3f96361c3 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -256,7 +256,7 @@ def stack_conds(conds): def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): - # images is required here to give training previews their infotext. Importing this at the very top causes a circular dependency. + # images allows training previews to have infotext. Importing it at the top causes a circular import problem. from modules import images assert hypernetwork_name, 'hypernetwork not selected' diff --git a/modules/images.py b/modules/images.py index 550e53ae7..b8834e3c3 100644 --- a/modules/images.py +++ b/modules/images.py @@ -419,7 +419,7 @@ def save_image(image, path, basename, seed=None, prompt=None, extension='png', i try: os.makedirs(path, exist_ok=True) except FileExistsError: - # If the file already exists, continue and allow said file to be overwritten. + # If the file already exists, allow said file to be overwritten. pass if forced_filename is None: From 2273e752fb3e578f1047f6d38b96330b07bf61a9 Mon Sep 17 00:00:00 2001 From: timntorres Date: Wed, 19 Oct 2022 14:23:48 -0700 Subject: [PATCH 64/82] Remove redundant try/except. --- modules/images.py | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-) diff --git a/modules/images.py b/modules/images.py index b8834e3c3..b9589563a 100644 --- a/modules/images.py +++ b/modules/images.py @@ -416,11 +416,7 @@ def save_image(image, path, basename, seed=None, prompt=None, extension='png', i dirname = apply_filename_pattern(opts.directories_filename_pattern or "[prompt_words]", p, seed, prompt).strip('\\ /') path = os.path.join(path, dirname) - try: - os.makedirs(path, exist_ok=True) - except FileExistsError: - # If the file already exists, allow said file to be overwritten. - pass + os.makedirs(path, exist_ok=True) if forced_filename is None: basecount = get_next_sequence_number(path, basename) From 03a1e288c4973dd2dff57a97469b40f146b6fccf Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 21 Oct 2022 10:13:24 +0300 Subject: [PATCH 65/82] turns out LayerNorm also has weight and bias and needs to be pre-multiplied and trained for hypernets --- modules/hypernetworks/hypernetwork.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 3274a8027..b1a5d0c73 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -52,7 +52,7 @@ class HypernetworkModule(torch.nn.Module): self.load_state_dict(state_dict) else: for layer in self.linear: - if type(layer) == torch.nn.Linear: + if type(layer) == torch.nn.Linear or type(layer) == torch.nn.LayerNorm: layer.weight.data.normal_(mean=0.0, std=0.01) layer.bias.data.zero_() @@ -80,7 +80,7 @@ class HypernetworkModule(torch.nn.Module): def trainables(self): layer_structure = [] for layer in self.linear: - if type(layer) == torch.nn.Linear: + if type(layer) == torch.nn.Linear or type(layer) == torch.nn.LayerNorm: layer_structure += [layer.weight, layer.bias] return layer_structure From bf30673f5132c8f28357b31224c54331e788d3e7 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 21 Oct 2022 10:19:25 +0300 Subject: [PATCH 66/82] Fix Hypernet infotext string split bug for PR #3283 --- modules/processing.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/processing.py b/modules/processing.py index 217869685..d1deffa94 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -304,7 +304,7 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration "Size": f"{p.width}x{p.height}", "Model hash": getattr(p, 'sd_model_hash', None if not opts.add_model_hash_to_info or not shared.sd_model.sd_model_hash else shared.sd_model.sd_model_hash), "Model": (None if not opts.add_model_name_to_info or not shared.sd_model.sd_checkpoint_info.model_name else shared.sd_model.sd_checkpoint_info.model_name.replace(',', '').replace(':', '')), - "Hypernet": (None if shared.loaded_hypernetwork is None else shared.loaded_hypernetwork.filename.split('\\')[-1].split('.')[0]), + "Hypernet": (None if shared.loaded_hypernetwork is None else os.path.splitext(os.path.basename(shared.loaded_hypernetwork.filename))[0]), "Batch size": (None if p.batch_size < 2 else p.batch_size), "Batch pos": (None if p.batch_size < 2 else position_in_batch), "Variation seed": (None if p.subseed_strength == 0 else all_subseeds[index]), From 1ed227b3b57f06f4152be3bfc9f83b0a839a2604 Mon Sep 17 00:00:00 2001 From: Leo Mozoloa Date: Fri, 21 Oct 2022 10:57:40 +0200 Subject: [PATCH 67/82] wtf is happening --- .github/ISSUE_TEMPLATE/config.yml | 5 +++++ 1 file changed, 5 insertions(+) create mode 100644 .github/ISSUE_TEMPLATE/config.yml diff --git a/.github/ISSUE_TEMPLATE/config.yml b/.github/ISSUE_TEMPLATE/config.yml new file mode 100644 index 000000000..f58c94a9b --- /dev/null +++ b/.github/ISSUE_TEMPLATE/config.yml @@ -0,0 +1,5 @@ +blank_issues_enabled: false +contact_links: + - name: WebUI Community Support + url: https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions + about: Please ask and answer questions here. From 003d2c7fe427edde299274c9e0d5fa59734e7f7e Mon Sep 17 00:00:00 2001 From: ClashSAN <98228077+ClashSAN@users.noreply.github.com> Date: Fri, 21 Oct 2022 11:40:37 +0000 Subject: [PATCH 68/82] Update README.md --- README.md | 16 ++++++++++++---- 1 file changed, 12 insertions(+), 4 deletions(-) diff --git a/README.md b/README.md index a89593bf1..348aaf87c 100644 --- a/README.md +++ b/README.md @@ -24,6 +24,7 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web - have as many embeddings as you want and use any names you like for them - use multiple embeddings with different numbers of vectors per token - works with half precision floating point numbers + - train embeddings on 8GB (also reports of 6GB working) - Extras tab with: - GFPGAN, neural network that fixes faces - CodeFormer, face restoration tool as an alternative to GFPGAN @@ -60,7 +61,7 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web - CLIP interrogator, a button that tries to guess prompt from an image - Prompt Editing, a way to change prompt mid-generation, say to start making a watermelon and switch to anime girl midway - Batch Processing, process a group of files using img2img -- Img2img Alternative +- Img2img Alternative, reverse Euler method of cross attention control - Highres Fix, a convenience option to produce high resolution pictures in one click without usual distortions - Reloading checkpoints on the fly - Checkpoint Merger, a tab that allows you to merge up to 3 checkpoints into one @@ -73,15 +74,22 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web - [xformers](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Xformers), major speed increase for select cards: (add --xformers to commandline args) - History tab: view, direct and delete images conveniently within the UI - Generate forever option -- Training Tab -- Preprocessing Image Datasets: cropping, mirroring, autotagging using BLIP or deepdanbooru (for anime) +- Training tab + - hypernetworks and embeddings options + - Preprocessing images: cropping, mirroring, autotagging using BLIP or deepdanbooru (for anime) +- Clip skip +- Use Hypernetworks +- Use VAEs +- Estimated completion time in progress bar +- API +- Support for dedicated inpainting model by RunwayML. ## Installation and Running Make sure the required [dependencies](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Dependencies) are met and follow the instructions available for both [NVidia](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs) (recommended) and [AMD](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs) GPUs. -Alternatively, use online services(like Google Colab): +Alternatively, use online services (like Google Colab): - [List of Online Services](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Online-Services) From df5706409386cc2e88718bd9101045587c39f8bb Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 21 Oct 2022 16:10:51 +0300 Subject: [PATCH 69/82] do not load aesthetic clip model until it's needed add refresh button for aesthetic embeddings add aesthetic params to images' infotext --- modules/aesthetic_clip.py | 40 ++++++++++++++--- modules/generation_parameters_copypaste.py | 18 +++++++- modules/img2img.py | 5 +-- modules/processing.py | 4 +- modules/sd_models.py | 3 -- modules/txt2img.py | 4 +- modules/ui.py | 52 +++++++++++++++------- style.css | 2 +- 8 files changed, 89 insertions(+), 39 deletions(-) diff --git a/modules/aesthetic_clip.py b/modules/aesthetic_clip.py index 34efa931b..8c828541f 100644 --- a/modules/aesthetic_clip.py +++ b/modules/aesthetic_clip.py @@ -40,6 +40,8 @@ def iter_to_batched(iterable, n=1): def create_ui(): + import modules.ui + with gr.Group(): with gr.Accordion("Open for Clip Aesthetic!", open=False): with gr.Row(): @@ -55,6 +57,8 @@ def create_ui(): label="Aesthetic imgs embedding", value="None") + modules.ui.create_refresh_button(aesthetic_imgs, shared.update_aesthetic_embeddings, lambda: {"choices": sorted(shared.aesthetic_embeddings.keys())}, "refresh_aesthetic_embeddings") + with gr.Row(): aesthetic_imgs_text = gr.Textbox(label='Aesthetic text for imgs', placeholder="This text is used to rotate the feature space of the imgs embs", @@ -66,11 +70,21 @@ def create_ui(): return aesthetic_weight, aesthetic_steps, aesthetic_lr, aesthetic_slerp, aesthetic_imgs, aesthetic_imgs_text, aesthetic_slerp_angle, aesthetic_text_negative +aesthetic_clip_model = None + + +def aesthetic_clip(): + global aesthetic_clip_model + + if aesthetic_clip_model is None or aesthetic_clip_model.name_or_path != shared.sd_model.cond_stage_model.wrapped.transformer.name_or_path: + aesthetic_clip_model = CLIPModel.from_pretrained(shared.sd_model.cond_stage_model.wrapped.transformer.name_or_path) + aesthetic_clip_model.cpu() + + return aesthetic_clip_model + + def generate_imgs_embd(name, folder, batch_size): - # clipModel = CLIPModel.from_pretrained( - # shared.sd_model.cond_stage_model.clipModel.name_or_path - # ) - model = shared.clip_model.to(device) + model = aesthetic_clip().to(device) processor = CLIPProcessor.from_pretrained(model.name_or_path) with torch.no_grad(): @@ -91,7 +105,7 @@ def generate_imgs_embd(name, folder, batch_size): path = str(Path(shared.cmd_opts.aesthetic_embeddings_dir) / f"{name}.pt") torch.save(embs, path) - model = model.cpu() + model.cpu() del processor del embs gc.collect() @@ -132,7 +146,7 @@ class AestheticCLIP: self.image_embs = None self.load_image_embs(None) - def set_aesthetic_params(self, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, image_embs_name=None, + def set_aesthetic_params(self, p, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, image_embs_name=None, aesthetic_slerp=True, aesthetic_imgs_text="", aesthetic_slerp_angle=0.15, aesthetic_text_negative=False): @@ -145,6 +159,18 @@ class AestheticCLIP: self.aesthetic_steps = aesthetic_steps self.load_image_embs(image_embs_name) + if self.image_embs_name is not None: + p.extra_generation_params.update({ + "Aesthetic LR": aesthetic_lr, + "Aesthetic weight": aesthetic_weight, + "Aesthetic steps": aesthetic_steps, + "Aesthetic embedding": self.image_embs_name, + "Aesthetic slerp": aesthetic_slerp, + "Aesthetic text": aesthetic_imgs_text, + "Aesthetic text negative": aesthetic_text_negative, + "Aesthetic slerp angle": aesthetic_slerp_angle, + }) + def set_skip(self, skip): self.skip = skip @@ -168,7 +194,7 @@ class AestheticCLIP: tokens = torch.asarray(remade_batch_tokens).to(device) - model = copy.deepcopy(shared.clip_model).to(device) + model = copy.deepcopy(aesthetic_clip()).to(device) model.requires_grad_(True) if self.aesthetic_imgs_text is not None and len(self.aesthetic_imgs_text) > 0: text_embs_2 = model.get_text_features( diff --git a/modules/generation_parameters_copypaste.py b/modules/generation_parameters_copypaste.py index 0f0414499..f73647dab 100644 --- a/modules/generation_parameters_copypaste.py +++ b/modules/generation_parameters_copypaste.py @@ -4,13 +4,22 @@ import gradio as gr from modules.shared import script_path from modules import shared -re_param_code = r"\s*([\w ]+):\s*([^,]+)(?:,|$)" +re_param_code = r'\s*([\w ]+):\s*("(?:\\|\"|[^\"])+"|[^,]*)(?:,|$)' re_param = re.compile(re_param_code) re_params = re.compile(r"^(?:" + re_param_code + "){3,}$") re_imagesize = re.compile(r"^(\d+)x(\d+)$") type_of_gr_update = type(gr.update()) +def quote(text): + if ',' not in str(text): + return text + + text = str(text) + text = text.replace('\\', '\\\\') + text = text.replace('"', '\\"') + return f'"{text}"' + def parse_generation_parameters(x: str): """parses generation parameters string, the one you see in text field under the picture in UI: ``` @@ -83,7 +92,12 @@ def connect_paste(button, paste_fields, input_comp, js=None): else: try: valtype = type(output.value) - val = valtype(v) + + if valtype == bool and v == "False": + val = False + else: + val = valtype(v) + res.append(gr.update(value=val)) except Exception: res.append(gr.update()) diff --git a/modules/img2img.py b/modules/img2img.py index bc7c66bc5..eea5199b9 100644 --- a/modules/img2img.py +++ b/modules/img2img.py @@ -109,10 +109,7 @@ def img2img(mode: int, prompt: str, negative_prompt: str, prompt_style: str, pro inpainting_mask_invert=inpainting_mask_invert, ) - shared.aesthetic_clip.set_aesthetic_params(float(aesthetic_lr), float(aesthetic_weight), int(aesthetic_steps), - aesthetic_imgs, aesthetic_slerp, aesthetic_imgs_text, - aesthetic_slerp_angle, - aesthetic_text_negative) + shared.aesthetic_clip.set_aesthetic_params(p, float(aesthetic_lr), float(aesthetic_weight), int(aesthetic_steps), aesthetic_imgs, aesthetic_slerp, aesthetic_imgs_text, aesthetic_slerp_angle, aesthetic_text_negative) if shared.cmd_opts.enable_console_prompts: print(f"\nimg2img: {prompt}", file=shared.progress_print_out) diff --git a/modules/processing.py b/modules/processing.py index d1deffa94..f0852cd55 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -12,7 +12,7 @@ from skimage import exposure from typing import Any, Dict, List, Optional import modules.sd_hijack -from modules import devices, prompt_parser, masking, sd_samplers, lowvram +from modules import devices, prompt_parser, masking, sd_samplers, lowvram, generation_parameters_copypaste from modules.sd_hijack import model_hijack from modules.shared import opts, cmd_opts, state import modules.shared as shared @@ -318,7 +318,7 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration generation_params.update(p.extra_generation_params) - generation_params_text = ", ".join([k if k == v else f'{k}: {v}' for k, v in generation_params.items() if v is not None]) + generation_params_text = ", ".join([k if k == v else f'{k}: {generation_parameters_copypaste.quote(v)}' for k, v in generation_params.items() if v is not None]) negative_prompt_text = "\nNegative prompt: " + p.negative_prompt if p.negative_prompt else "" diff --git a/modules/sd_models.py b/modules/sd_models.py index 05a1df280..b1c91b0db 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -234,9 +234,6 @@ def load_model(checkpoint_info=None): sd_hijack.model_hijack.hijack(sd_model) - if shared.clip_model is None or shared.clip_model.transformer.name_or_path != sd_model.cond_stage_model.wrapped.transformer.name_or_path: - shared.clip_model = CLIPModel.from_pretrained(sd_model.cond_stage_model.wrapped.transformer.name_or_path) - sd_model.eval() print(f"Model loaded.") diff --git a/modules/txt2img.py b/modules/txt2img.py index 32ed1d8d0..1761cfa2c 100644 --- a/modules/txt2img.py +++ b/modules/txt2img.py @@ -36,9 +36,7 @@ def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: firstphase_height=firstphase_height if enable_hr else None, ) - shared.aesthetic_clip.set_aesthetic_params(float(aesthetic_lr), float(aesthetic_weight), int(aesthetic_steps), - aesthetic_imgs, aesthetic_slerp, aesthetic_imgs_text, aesthetic_slerp_angle, - aesthetic_text_negative) + shared.aesthetic_clip.set_aesthetic_params(p, float(aesthetic_lr), float(aesthetic_weight), int(aesthetic_steps), aesthetic_imgs, aesthetic_slerp, aesthetic_imgs_text, aesthetic_slerp_angle, aesthetic_text_negative) if cmd_opts.enable_console_prompts: print(f"\ntxt2img: {prompt}", file=shared.progress_print_out) diff --git a/modules/ui.py b/modules/ui.py index 381ca925b..0d020de6a 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -597,27 +597,29 @@ def apply_setting(key, value): return value +def create_refresh_button(refresh_component, refresh_method, refreshed_args, elem_id): + def refresh(): + refresh_method() + args = refreshed_args() if callable(refreshed_args) else refreshed_args + + for k, v in args.items(): + setattr(refresh_component, k, v) + + return gr.update(**(args or {})) + + refresh_button = gr.Button(value=refresh_symbol, elem_id=elem_id) + refresh_button.click( + fn=refresh, + inputs=[], + outputs=[refresh_component] + ) + return refresh_button + + def create_ui(wrap_gradio_gpu_call): import modules.img2img import modules.txt2img - def create_refresh_button(refresh_component, refresh_method, refreshed_args, elem_id): - def refresh(): - refresh_method() - args = refreshed_args() if callable(refreshed_args) else refreshed_args - - for k, v in args.items(): - setattr(refresh_component, k, v) - - return gr.update(**(args or {})) - - refresh_button = gr.Button(value=refresh_symbol, elem_id=elem_id) - refresh_button.click( - fn = refresh, - inputs = [], - outputs = [refresh_component] - ) - return refresh_button with gr.Blocks(analytics_enabled=False) as txt2img_interface: txt2img_prompt, roll, txt2img_prompt_style, txt2img_negative_prompt, txt2img_prompt_style2, submit, _, _, txt2img_prompt_style_apply, txt2img_save_style, txt2img_paste, token_counter, token_button = create_toprow(is_img2img=False) @@ -802,6 +804,14 @@ def create_ui(wrap_gradio_gpu_call): (hr_options, lambda d: gr.Row.update(visible="Denoising strength" in d)), (firstphase_width, "First pass size-1"), (firstphase_height, "First pass size-2"), + (aesthetic_lr, "Aesthetic LR"), + (aesthetic_weight, "Aesthetic weight"), + (aesthetic_steps, "Aesthetic steps"), + (aesthetic_imgs, "Aesthetic embedding"), + (aesthetic_slerp, "Aesthetic slerp"), + (aesthetic_imgs_text, "Aesthetic text"), + (aesthetic_text_negative, "Aesthetic text negative"), + (aesthetic_slerp_angle, "Aesthetic slerp angle"), ] txt2img_preview_params = [ @@ -1077,6 +1087,14 @@ def create_ui(wrap_gradio_gpu_call): (seed_resize_from_w, "Seed resize from-1"), (seed_resize_from_h, "Seed resize from-2"), (denoising_strength, "Denoising strength"), + (aesthetic_lr_im, "Aesthetic LR"), + (aesthetic_weight_im, "Aesthetic weight"), + (aesthetic_steps_im, "Aesthetic steps"), + (aesthetic_imgs_im, "Aesthetic embedding"), + (aesthetic_slerp_im, "Aesthetic slerp"), + (aesthetic_imgs_text_im, "Aesthetic text"), + (aesthetic_text_negative_im, "Aesthetic text negative"), + (aesthetic_slerp_angle_im, "Aesthetic slerp angle"), ] token_button.click(fn=update_token_counter, inputs=[img2img_prompt, steps], outputs=[token_counter]) diff --git a/style.css b/style.css index 26ae36a5c..5d2bacc99 100644 --- a/style.css +++ b/style.css @@ -477,7 +477,7 @@ input[type="range"]{ padding: 0; } -#refresh_sd_model_checkpoint, #refresh_sd_hypernetwork, #refresh_train_hypernetwork_name, #refresh_train_embedding_name, #refresh_localization{ +#refresh_sd_model_checkpoint, #refresh_sd_hypernetwork, #refresh_train_hypernetwork_name, #refresh_train_embedding_name, #refresh_localization, #refresh_aesthetic_embeddings{ max-width: 2.5em; min-width: 2.5em; height: 2.4em; From 9286fe53de2eef91f13cc3ad5938ddf67ecc8413 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 21 Oct 2022 16:38:06 +0300 Subject: [PATCH 70/82] make aestetic embedding ciompatible with prompts longer than 75 tokens --- modules/sd_hijack.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py index 36198a3ce..1f8587d15 100644 --- a/modules/sd_hijack.py +++ b/modules/sd_hijack.py @@ -332,8 +332,8 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): multipliers.append([1.0] * 75) z1 = self.process_tokens(tokens, multipliers) + z1 = shared.aesthetic_clip(z1, remade_batch_tokens) z = z1 if z is None else torch.cat((z, z1), axis=-2) - z = shared.aesthetic_clip(z, remade_batch_tokens) remade_batch_tokens = rem_tokens batch_multipliers = rem_multipliers From d0ea471b0cdaede163c6e7f6fae8535f5c3cd226 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Fri, 21 Oct 2022 14:04:41 +0100 Subject: [PATCH 71/82] Use opts in textual_inversion image_embedding.py for dynamic fonts --- modules/textual_inversion/image_embedding.py | 1 + 1 file changed, 1 insertion(+) diff --git a/modules/textual_inversion/image_embedding.py b/modules/textual_inversion/image_embedding.py index 898ce3b3b..c50b1e7b3 100644 --- a/modules/textual_inversion/image_embedding.py +++ b/modules/textual_inversion/image_embedding.py @@ -5,6 +5,7 @@ import zlib from PIL import Image, PngImagePlugin, ImageDraw, ImageFont from fonts.ttf import Roboto import torch +from modules.shared import opts class EmbeddingEncoder(json.JSONEncoder): From 306e2ff6ab8f4c7e94ab55f4f08ab8f94d73d287 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Fri, 21 Oct 2022 14:47:21 +0100 Subject: [PATCH 72/82] Update image_embedding.py --- modules/textual_inversion/image_embedding.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/modules/textual_inversion/image_embedding.py b/modules/textual_inversion/image_embedding.py index c50b1e7b3..ea653806f 100644 --- a/modules/textual_inversion/image_embedding.py +++ b/modules/textual_inversion/image_embedding.py @@ -134,7 +134,7 @@ def caption_image_overlay(srcimage, title, footerLeft, footerMid, footerRight, t from math import cos image = srcimage.copy() - + fontsize = 32 if textfont is None: try: textfont = ImageFont.truetype(opts.font or Roboto, fontsize) @@ -151,7 +151,7 @@ def caption_image_overlay(srcimage, title, footerLeft, footerMid, footerRight, t image = Image.alpha_composite(image.convert('RGBA'), gradient.resize(image.size)) draw = ImageDraw.Draw(image) - fontsize = 32 + font = ImageFont.truetype(textfont, fontsize) padding = 10 From 85cb5918ee7c97cafafe4149880ed82f1933d919 Mon Sep 17 00:00:00 2001 From: parsec501 <105080989+parsec501@users.noreply.github.com> Date: Fri, 21 Oct 2022 14:43:23 +0200 Subject: [PATCH 73/82] Make commit hash mandatory field --- .github/ISSUE_TEMPLATE/bug_report.yml | 2 ++ 1 file changed, 2 insertions(+) diff --git a/.github/ISSUE_TEMPLATE/bug_report.yml b/.github/ISSUE_TEMPLATE/bug_report.yml index 35802a53c..9c2ff3138 100644 --- a/.github/ISSUE_TEMPLATE/bug_report.yml +++ b/.github/ISSUE_TEMPLATE/bug_report.yml @@ -45,6 +45,8 @@ body: attributes: label: Commit where the problem happens description: Which commit are you running ? (copy the **Commit hash** shown in the cmd/terminal when you launch the UI) + validations: + required: true - type: dropdown id: platforms attributes: From 51e3dc9ccad157d7161b697a246e26c868d46a7c Mon Sep 17 00:00:00 2001 From: timntorres Date: Fri, 21 Oct 2022 02:11:12 -0700 Subject: [PATCH 74/82] Sanitize hypernet name input. --- modules/hypernetworks/ui.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py index 266f04f60..e6f50a1f8 100644 --- a/modules/hypernetworks/ui.py +++ b/modules/hypernetworks/ui.py @@ -11,6 +11,9 @@ from modules.hypernetworks import hypernetwork def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None, add_layer_norm=False, activation_func=None): + # Remove illegal characters from name. + name = "".join( x for x in name if (x.isalnum() or x in "._- ")) + fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt") if not overwrite_old: assert not os.path.exists(fn), f"file {fn} already exists" From 19818f023cfafc472c6c241cab0b72896a168481 Mon Sep 17 00:00:00 2001 From: timntorres Date: Fri, 21 Oct 2022 02:14:02 -0700 Subject: [PATCH 75/82] Match hypernet name with filename in all cases. --- modules/hypernetworks/hypernetwork.py | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index b1a5d0c73..6d392be4d 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -340,7 +340,10 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log pbar.set_description(f"loss: {mean_loss:.7f}") if hypernetwork.step > 0 and hypernetwork_dir is not None and hypernetwork.step % save_hypernetwork_every == 0: - last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork_name}-{hypernetwork.step}.pt') + temp = hypernetwork.name + # Before saving, change name to match current checkpoint. + hypernetwork.name = f'{hypernetwork_name}-{hypernetwork.step}' + last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork.name}.pt') hypernetwork.save(last_saved_file) textual_inversion.write_loss(log_directory, "hypernetwork_loss.csv", hypernetwork.step, len(ds), { @@ -405,6 +408,9 @@ Last saved image: {html.escape(last_saved_image)}
hypernetwork.sd_checkpoint = checkpoint.hash hypernetwork.sd_checkpoint_name = checkpoint.model_name + # Before saving for the last time, change name back to the base name (as opposed to the save_hypernetwork_every step-suffixed naming convention). + hypernetwork.name = hypernetwork_name + filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork.name}.pt') hypernetwork.save(filename) return hypernetwork, filename From fccad18a59e3c2c33fefbbb1763c6a87a3a68eba Mon Sep 17 00:00:00 2001 From: timntorres Date: Fri, 21 Oct 2022 02:17:26 -0700 Subject: [PATCH 76/82] Refer to Hypernet's name, sensibly, by its name variable. --- modules/processing.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/processing.py b/modules/processing.py index f0852cd55..ff1ec4c9b 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -304,7 +304,7 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration "Size": f"{p.width}x{p.height}", "Model hash": getattr(p, 'sd_model_hash', None if not opts.add_model_hash_to_info or not shared.sd_model.sd_model_hash else shared.sd_model.sd_model_hash), "Model": (None if not opts.add_model_name_to_info or not shared.sd_model.sd_checkpoint_info.model_name else shared.sd_model.sd_checkpoint_info.model_name.replace(',', '').replace(':', '')), - "Hypernet": (None if shared.loaded_hypernetwork is None else os.path.splitext(os.path.basename(shared.loaded_hypernetwork.filename))[0]), + "Hypernet": (None if shared.loaded_hypernetwork is None else shared.loaded_hypernetwork.name), "Batch size": (None if p.batch_size < 2 else p.batch_size), "Batch pos": (None if p.batch_size < 2 else position_in_batch), "Variation seed": (None if p.subseed_strength == 0 else all_subseeds[index]), From 272fa527bbe93143668ffc16838107b7dca35b40 Mon Sep 17 00:00:00 2001 From: timntorres Date: Fri, 21 Oct 2022 02:41:55 -0700 Subject: [PATCH 77/82] Remove unused variable. --- modules/hypernetworks/hypernetwork.py | 1 - 1 file changed, 1 deletion(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 6d392be4d..47d91ea58 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -340,7 +340,6 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log pbar.set_description(f"loss: {mean_loss:.7f}") if hypernetwork.step > 0 and hypernetwork_dir is not None and hypernetwork.step % save_hypernetwork_every == 0: - temp = hypernetwork.name # Before saving, change name to match current checkpoint. hypernetwork.name = f'{hypernetwork_name}-{hypernetwork.step}' last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork.name}.pt') From 02e4d4694dd9254a6ca9f05c2eb7b01ea508abc7 Mon Sep 17 00:00:00 2001 From: Rcmcpe Date: Fri, 21 Oct 2022 15:53:35 +0800 Subject: [PATCH 78/82] Change option description of unload_models_when_training --- modules/shared.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/shared.py b/modules/shared.py index 5c675b805..41d7f08e1 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -266,7 +266,7 @@ options_templates.update(options_section(('system', "System"), { })) options_templates.update(options_section(('training', "Training"), { - "unload_models_when_training": OptionInfo(False, "Unload VAE and CLIP from VRAM when training"), + "unload_models_when_training": OptionInfo(False, "Move VAE and CLIP to RAM when training hypernetwork. Saves VRAM."), "dataset_filename_word_regex": OptionInfo("", "Filename word regex"), "dataset_filename_join_string": OptionInfo(" ", "Filename join string"), "training_image_repeats_per_epoch": OptionInfo(1, "Number of repeats for a single input image per epoch; used only for displaying epoch number", gr.Number, {"precision": 0}), From 704036ff07b71bf86cadcbbff2bcfeebdd1ed3a6 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 21 Oct 2022 17:11:42 +0300 Subject: [PATCH 79/82] make aspect ratio overlay work regardless of selected localization --- javascript/aspectRatioOverlay.js | 36 +++++++++++++++----------------- javascript/dragdrop.js | 2 +- modules/ui.py | 4 ++-- 3 files changed, 20 insertions(+), 22 deletions(-) diff --git a/javascript/aspectRatioOverlay.js b/javascript/aspectRatioOverlay.js index 96f1c00d0..d3ca2781b 100644 --- a/javascript/aspectRatioOverlay.js +++ b/javascript/aspectRatioOverlay.js @@ -3,12 +3,12 @@ let currentWidth = null; let currentHeight = null; let arFrameTimeout = setTimeout(function(){},0); -function dimensionChange(e,dimname){ +function dimensionChange(e, is_width, is_height){ - if(dimname == 'Width'){ + if(is_width){ currentWidth = e.target.value*1.0 } - if(dimname == 'Height'){ + if(is_height){ currentHeight = e.target.value*1.0 } @@ -98,22 +98,20 @@ onUiUpdate(function(){ var inImg2img = Boolean(gradioApp().querySelector("button.rounded-t-lg.border-gray-200")) if(inImg2img){ let inputs = gradioApp().querySelectorAll('input'); - inputs.forEach(function(e){ - let parentLabel = e.parentElement.querySelector('label') - if(parentLabel && parentLabel.innerText){ - if(!e.classList.contains('scrollwatch')){ - if(parentLabel.innerText == 'Width' || parentLabel.innerText == 'Height'){ - e.addEventListener('input', function(e){dimensionChange(e,parentLabel.innerText)} ) - e.classList.add('scrollwatch') - } - if(parentLabel.innerText == 'Width'){ - currentWidth = e.value*1.0 - } - if(parentLabel.innerText == 'Height'){ - currentHeight = e.value*1.0 - } - } - } + inputs.forEach(function(e){ + var is_width = e.parentElement.id == "img2img_width" + var is_height = e.parentElement.id == "img2img_height" + + if((is_width || is_height) && !e.classList.contains('scrollwatch')){ + e.addEventListener('input', function(e){dimensionChange(e, is_width, is_height)} ) + e.classList.add('scrollwatch') + } + if(is_width){ + currentWidth = e.value*1.0 + } + if(is_height){ + currentHeight = e.value*1.0 + } }) } }); diff --git a/javascript/dragdrop.js b/javascript/dragdrop.js index 070cf255f..3ed1cb3c6 100644 --- a/javascript/dragdrop.js +++ b/javascript/dragdrop.js @@ -43,7 +43,7 @@ function dropReplaceImage( imgWrap, files ) { window.document.addEventListener('dragover', e => { const target = e.composedPath()[0]; const imgWrap = target.closest('[data-testid="image"]'); - if ( !imgWrap && target.placeholder.indexOf("Prompt") == -1) { + if ( !imgWrap && target.placeholder && target.placeholder.indexOf("Prompt") == -1) { return; } e.stopPropagation(); diff --git a/modules/ui.py b/modules/ui.py index 0d020de6a..85f957923 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -879,8 +879,8 @@ def create_ui(wrap_gradio_gpu_call): sampler_index = gr.Radio(label='Sampling method', choices=[x.name for x in samplers_for_img2img], value=samplers_for_img2img[0].name, type="index") with gr.Group(): - width = gr.Slider(minimum=64, maximum=2048, step=64, label="Width", value=512) - height = gr.Slider(minimum=64, maximum=2048, step=64, label="Height", value=512) + width = gr.Slider(minimum=64, maximum=2048, step=64, label="Width", value=512, elem_id="img2img_width") + height = gr.Slider(minimum=64, maximum=2048, step=64, label="Height", value=512, elem_id="img2img_height") with gr.Row(): restore_faces = gr.Checkbox(label='Restore faces', value=False, visible=len(shared.face_restorers) > 1) From 3d898044e5e55dca1698e9b5b7d3558b5b78675a Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 21 Oct 2022 17:26:30 +0300 Subject: [PATCH 80/82] batch_size does not affect job count --- scripts/outpainting_mk_2.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/outpainting_mk_2.py b/scripts/outpainting_mk_2.py index 633dc1198..2afd4aa51 100644 --- a/scripts/outpainting_mk_2.py +++ b/scripts/outpainting_mk_2.py @@ -246,7 +246,7 @@ class Script(scripts.Script): batch_count = p.n_iter batch_size = p.batch_size p.n_iter = 1 - state.job_count = batch_count * batch_size * ((1 if left > 0 else 0) + (1 if right > 0 else 0) + (1 if up > 0 else 0) + (1 if down > 0 else 0)) + state.job_count = batch_count * ((1 if left > 0 else 0) + (1 if right > 0 else 0) + (1 if up > 0 else 0) + (1 if down > 0 else 0)) all_processed_images = [] for i in range(batch_count): From ac0aa2b18efeeb9220a5994c8dd54c7cdda7cc40 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 21 Oct 2022 17:35:51 +0300 Subject: [PATCH 81/82] loading SD VAE, see PR #3303 --- modules/sd_models.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/modules/sd_models.py b/modules/sd_models.py index b1c91b0db..d99dbce87 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -155,6 +155,9 @@ def get_state_dict_from_checkpoint(pl_sd): return pl_sd +vae_ignore_keys = {"model_ema.decay", "model_ema.num_updates"} + + def load_model_weights(model, checkpoint_info): checkpoint_file = checkpoint_info.filename sd_model_hash = checkpoint_info.hash @@ -186,7 +189,7 @@ def load_model_weights(model, checkpoint_info): if os.path.exists(vae_file): print(f"Loading VAE weights from: {vae_file}") vae_ckpt = torch.load(vae_file, map_location=shared.weight_load_location) - vae_dict = {k: v for k, v in vae_ckpt["state_dict"].items() if k[0:4] != "loss"} + vae_dict = {k: v for k, v in vae_ckpt["state_dict"].items() if k[0:4] != "loss" and k not in vae_ignore_keys} model.first_stage_model.load_state_dict(vae_dict) model.first_stage_model.to(devices.dtype_vae) From 24ce67a13bd74202d298cd8e2a306d90214980d8 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 21 Oct 2022 17:41:47 +0300 Subject: [PATCH 82/82] make aspect ratio overlay work regardless of selected localization, pt2 --- javascript/aspectRatioOverlay.js | 19 +++++-------------- 1 file changed, 5 insertions(+), 14 deletions(-) diff --git a/javascript/aspectRatioOverlay.js b/javascript/aspectRatioOverlay.js index d3ca2781b..66f26a22a 100644 --- a/javascript/aspectRatioOverlay.js +++ b/javascript/aspectRatioOverlay.js @@ -18,22 +18,13 @@ function dimensionChange(e, is_width, is_height){ return; } - var img2imgMode = gradioApp().querySelector('#mode_img2img.tabs > div > button.rounded-t-lg.border-gray-200') - if(img2imgMode){ - img2imgMode=img2imgMode.innerText - }else{ - return; - } - - var redrawImage = gradioApp().querySelector('div[data-testid=image] img'); - var inpaintImage = gradioApp().querySelector('#img2maskimg div[data-testid=image] img') - var targetElement = null; - if(img2imgMode=='img2img' && redrawImage){ - targetElement = redrawImage; - }else if(img2imgMode=='Inpaint' && inpaintImage){ - targetElement = inpaintImage; + var tabIndex = get_tab_index('mode_img2img') + if(tabIndex == 0){ + targetElement = gradioApp().querySelector('div[data-testid=image] img'); + } else if(tabIndex == 1){ + targetElement = gradioApp().querySelector('#img2maskimg div[data-testid=image] img'); } if(targetElement){