diff --git a/Custom-Scripts.md b/Custom-Scripts.md index 8b92777..67313f0 100644 --- a/Custom-Scripts.md +++ b/Custom-Scripts.md @@ -320,19 +320,18 @@ Steps: 20, Sampler: DDIM, CFG scale: 7, Seed: 4288895889, Size: 512x512, Model h -## Multiple Hypernetworks -https://github.com/antis0007/sd-webui-multiple-hypernetworks +## sd_script-skippy-the-step-kangaroo +https://github.com/FartyPants/sd_script_skippy -Adds the ability to apply multiple hypernetworks at once. Overrides the hijack, optimization and CrossAttention forward functions in order to apply multiple hypernetworks sequentially, with different weights. +Script for Automatic1111 The ultimate solution for those pesky last steps that ruin your SD generation images -## Hypernetwork Structure(.hns)/Variable Dropout/ Monkey Patches -https://github.com/aria1th/Hypernetwork-MonkeyPatch-Extension +Imagine this: So you're watching the SD generation in your browser - image is good, good, good then on the last few steps, BOOM, it's ugly. -Adds the ability to apply Hypernetwork Structure, as defining it in .hns file. see [here](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/4334) for detailed information. +This handy script is the remedy. It enables you to snatch incomplete generation by skipping last few steps. Works great when you have overtrained ckpt -Adds the ability to use proper variable dropout rate, like 0.05. Also fixes issues with using hypernetwork right after training. +Example: you put euler a on 25 steps, but now you can snatch image at step 19, like a ... I don't know, a puppy? The new, early image will be sent to the interface -Adds creating beta-hypernetwork(dropout), and beta-training which allows automatic cosine annealing, and no-crop usage of original images. +Here is how you test this contraption: Set Steps at 20 and move the Skip steps to 16 (that means: 20-16 = 4) - you should see a blurry image in interface that was snatched at step 4 as we skipped last 16 steps. ## Saving steps of the sampling process