move some settings to the new Optimization page

add slider for token merging for img2img
rework StableDiffusionProcessing to have the token_merging_ratio field
fix a bug with applying png optimizations for live previews when they shouldn't be applied

modules/processing.py

@@ -29,12 +29,6 @@ from ldm.models.diffusion.ddpm import LatentDepth2ImageDiffusion
 
 from einops import repeat, rearrange
 from blendmodes.blend import blendLayers, BlendType
-import tomesd
-
-# add a logger for the processing module
-logger = logging.getLogger(__name__)
-# manually set output level here since there is no option to do so yet through launch options
-# logging.basicConfig(level=logging.DEBUG, format='%(asctime)s %(levelname)s %(name)s %(message)s')
 
 
 # some of those options should not be changed at all because they would break the model, so I removed them from options.
@@ -156,6 +150,8 @@ class StableDiffusionProcessing:
         self.override_settings_restore_afterwards = override_settings_restore_afterwards
         self.is_using_inpainting_conditioning = False
         self.disable_extra_networks = False
+        self.token_merging_ratio = 0
+        self.token_merging_ratio_hr = 0
 
         if not seed_enable_extras:
             self.subseed = -1
@@ -171,6 +167,7 @@ class StableDiffusionProcessing:
         self.all_subseeds = None
         self.iteration = 0
         self.is_hr_pass = False
+        self.sampler = None
 
 
     @property
@@ -280,6 +277,12 @@ class StableDiffusionProcessing:
     def close(self):
         self.sampler = None
 
+    def get_token_merging_ratio(self, for_hr=False):
+        if for_hr:
+            return self.token_merging_ratio_hr or opts.token_merging_ratio_hr or self.token_merging_ratio or opts.token_merging_ratio
+
+        return self.token_merging_ratio or opts.token_merging_ratio
+
 
 class Processed:
     def __init__(self, p: StableDiffusionProcessing, images_list, seed=-1, info="", subseed=None, all_prompts=None, all_negative_prompts=None, all_seeds=None, all_subseeds=None, index_of_first_image=0, infotexts=None, comments=""):
@@ -309,6 +312,8 @@ class Processed:
         self.styles = p.styles
         self.job_timestamp = state.job_timestamp
         self.clip_skip = opts.CLIP_stop_at_last_layers
+        self.token_merging_ratio = p.token_merging_ratio
+        self.token_merging_ratio_hr = p.token_merging_ratio_hr
 
         self.eta = p.eta
         self.ddim_discretize = p.ddim_discretize
@@ -367,6 +372,9 @@ class Processed:
     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 get_token_merging_ratio(self, for_hr=False):
+        return self.token_merging_ratio_hr if for_hr else self.token_merging_ratio
+
 
 # from https://discuss.pytorch.org/t/help-regarding-slerp-function-for-generative-model-sampling/32475/3
 def slerp(val, low, high):
@@ -480,6 +488,8 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments=None, iter
 
     clip_skip = getattr(p, 'clip_skip', opts.CLIP_stop_at_last_layers)
     enable_hr = getattr(p, 'enable_hr', False)
+    token_merging_ratio = p.get_token_merging_ratio()
+    token_merging_ratio_hr = p.get_token_merging_ratio(for_hr=True)
 
     uses_ensd = opts.eta_noise_seed_delta != 0
     if uses_ensd:
@@ -502,8 +512,8 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments=None, iter
         "Conditional mask weight": getattr(p, "inpainting_mask_weight", shared.opts.inpainting_mask_weight) if p.is_using_inpainting_conditioning else None,
         "Clip skip": None if clip_skip <= 1 else clip_skip,
         "ENSD": opts.eta_noise_seed_delta if uses_ensd else None,
-        "Token merging ratio": None if opts.token_merging_ratio == 0 else opts.token_merging_ratio,
+        "Token merging ratio": None if token_merging_ratio == 0 else token_merging_ratio,
-        "Token merging ratio hr": None if not enable_hr or opts.token_merging_ratio_hr == 0 else opts.token_merging_ratio_hr,
+        "Token merging ratio hr": None if not enable_hr or token_merging_ratio_hr == 0 else token_merging_ratio_hr,
         "Init image hash": getattr(p, 'init_img_hash', None),
         "RNG": opts.randn_source if opts.randn_source != "GPU" else None,
         "NGMS": None if p.s_min_uncond == 0 else p.s_min_uncond,
@@ -536,17 +546,12 @@ def process_images(p: StableDiffusionProcessing) -> Processed:
             if k == 'sd_vae':
                 sd_vae.reload_vae_weights()
 
-        if opts.token_merging_ratio > 0:
+        sd_models.apply_token_merging(p.sd_model, p.get_token_merging_ratio())
-            sd_models.apply_token_merging(sd_model=p.sd_model, hr=False)
-            logger.debug(f"Token merging applied to first pass. Ratio: '{opts.token_merging_ratio}'")
 
         res = process_images_inner(p)
 
     finally:
-        # undo model optimizations made by tomesd
+        sd_models.apply_token_merging(p.sd_model, 0)
-        if opts.token_merging_ratio > 0:
-            tomesd.remove_patch(p.sd_model)
-            logger.debug('Token merging model optimizations removed')
 
         # restore opts to original state
         if p.override_settings_restore_afterwards:
@@ -996,21 +1001,11 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing):
         x = None
         devices.torch_gc()
 
-        # apply token merging optimizations from tomesd for high-res pass
+        sd_models.apply_token_merging(self.sd_model, self.get_token_merging_ratio(for_hr=True))
-        if opts.token_merging_ratio_hr > 0:
-            # in case the user has used separate merge ratios
-            if opts.token_merging_ratio > 0:
-                tomesd.remove_patch(self.sd_model)
-                logger.debug('Adjusting token merging ratio for high-res pass')
-
-            sd_models.apply_token_merging(sd_model=self.sd_model, hr=True)
-            logger.debug(f"Applied token merging for high-res pass. Ratio: '{opts.token_merging_ratio_hr}'")
 
         samples = self.sampler.sample_img2img(self, samples, noise, conditioning, unconditional_conditioning, steps=self.hr_second_pass_steps or self.steps, image_conditioning=image_conditioning)
 
-        if opts.token_merging_ratio_hr > 0 or opts.token_merging_ratio > 0:
+        sd_models.apply_token_merging(self.sd_model, self.get_token_merging_ratio())
-            tomesd.remove_patch(self.sd_model)
-            logger.debug('Removed token merging optimizations from model')
 
         self.is_hr_pass = False
 
@@ -1173,3 +1168,6 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing):
         devices.torch_gc()
 
         return samples
+
+    def get_token_merging_ratio(self, for_hr=False):
+        return self.token_merging_ratio or ("token_merging_ratio" in self.override_settings and opts.token_merging_ratio) or opts.token_merging_ratio_img2img or opts.token_merging_ratio

modules/progress.py

@@ -98,7 +98,11 @@ def progressapi(req: ProgressRequest):
 
             if opts.live_previews_image_format == "png":
                 # using optimize for large images takes an enormous amount of time
-                save_kwargs = {"optimize": max(*image.size) > 256}
+                if max(*image.size) <= 256:
+                    save_kwargs = {"optimize": True}
+                else:
+                    save_kwargs = {"optimize": False, "compress_level": 1}
+
             else:
                 save_kwargs = {}
 

modules/sd_models.py

@@ -583,23 +583,27 @@ def unload_model_weights(sd_model=None, info=None):
     return sd_model
 
 
-def apply_token_merging(sd_model, hr: bool):
+def apply_token_merging(sd_model, token_merging_ratio):
     """
     Applies speed and memory optimizations from tomesd.
-
-    Args:
-        hr (bool): True if called in the context of a high-res pass
     """
 
-    ratio = shared.opts.token_merging_ratio
+    current_token_merging_ratio = getattr(sd_model, 'applied_token_merged_ratio', 0)
-    if hr:
-        ratio = shared.opts.token_merging_ratio_hr
 
-    tomesd.apply_patch(
+    if current_token_merging_ratio == token_merging_ratio:
-        sd_model,
+        return
-        ratio=ratio,
+
-        use_rand=False,  # can cause issues with some samplers
+    if current_token_merging_ratio > 0:
-        merge_attn=True,
+        tomesd.remove_patch(sd_model)
-        merge_crossattn=False,
+
-        merge_mlp=False
+    if token_merging_ratio > 0:
-    )
+        tomesd.apply_patch(
+            sd_model,
+            ratio=token_merging_ratio,
+            use_rand=False,  # can cause issues with some samplers
+            merge_attn=True,
+            merge_crossattn=False,
+            merge_mlp=False
+        )
+
+    sd_model.applied_token_merged_ratio = token_merging_ratio

modules/shared.py

@@ -413,8 +413,13 @@ options_templates.update(options_section(('sd', "Stable Diffusion"), {
     "CLIP_stop_at_last_layers": OptionInfo(1, "Clip skip", gr.Slider, {"minimum": 1, "maximum": 12, "step": 1}).link("wiki", "https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#clip-skip").info("ignore last layers of CLIP nrtwork; 1 ignores none, 2 ignores one layer"),
     "upcast_attn": OptionInfo(False, "Upcast cross attention layer to float32"),
     "randn_source": OptionInfo("GPU", "Random number generator source.", gr.Radio, {"choices": ["GPU", "CPU"]}).info("changes seeds drastically; use CPU to produce the same picture across different vidocard vendors"),
+}))
+
+options_templates.update(options_section(('optimizations', "Optimizations"), {
+    "s_min_uncond": OptionInfo(0, "Negative Guidance minimum sigma", gr.Slider, {"minimum": 0.0, "maximum": 4.0, "step": 0.01}).link("PR", "https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/9177").info("skip negative prompt for some steps when the image is almost ready; 0=disable, higher=faster"),
     "token_merging_ratio": OptionInfo(0.0, "Token merging ratio", gr.Slider, {"minimum": 0.0, "maximum": 0.9, "step": 0.1}).link("PR", "https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/9256").info("0=disable, higher=faster"),
-    "token_merging_ratio_hr": OptionInfo(0.0, "Togen merging ratio for high-res pass", gr.Slider, {"minimum": 0.0, "maximum": 0.9, "step": 0.1}),
+    "token_merging_ratio_img2img": OptionInfo(0.0, "Token merging ratio for img2img", gr.Slider, {"minimum": 0.0, "maximum": 0.9, "step": 0.1}).info("only applies if non-zero and overrides above"),
+    "token_merging_ratio_hr": OptionInfo(0.0, "Token merging ratio for high-res pass", gr.Slider, {"minimum": 0.0, "maximum": 0.9, "step": 0.1}).info("only applies if non-zero and overrides above"),
 }))
 
 options_templates.update(options_section(('compatibility', "Compatibility"), {
@@ -498,7 +503,6 @@ options_templates.update(options_section(('sampler-params', "Sampler parameters"
     "eta_ddim": OptionInfo(0.0, "Eta for DDIM", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}).info("noise multiplier; higher = more unperdictable results"),
     "eta_ancestral": OptionInfo(1.0, "Eta for ancestral samplers", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}).info("noise multiplier; applies to Euler a and other samplers that have a in them"),
     "ddim_discretize": OptionInfo('uniform', "img2img DDIM discretize", gr.Radio, {"choices": ['uniform', 'quad']}),
-    's_min_uncond': OptionInfo(0, "Negative Guidance minimum sigma", gr.Slider, {"minimum": 0.0, "maximum": 4.0, "step": 0.01}).link("PR", "https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/9177").info("skip negative prompt for some steps when the image is almost ready; 0=disable, higher=faster"),
     's_churn': OptionInfo(0.0, "sigma churn", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}),
     's_tmin':  OptionInfo(0.0, "sigma tmin",  gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}),
     's_noise': OptionInfo(1.0, "sigma noise", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}),