I've made an Extension "Anti Burn (average smoothing of last steps images)"!

[aleksusklim]

Here is the repo: https://github.com/klimaleksus/stable-diffusion-webui-anti-burn

Sometimes, very last step of generation is corrupting the image, and there were a few userscripts that aimed to prevent this by discarding some steps at the end.

Also, there are extensions and userscripts that dump all of intermediate samples to .png files – so it is possible to see, which one was the best.

I thought: what if I take a short chain of images near to the end of generation – and try to average them by simple arithmetic mean? Will this improve overall quality by smoothing out random corrupted details?

Turns out it did! (And even better, if I'll merge the result again with "median" value for each channel).

Then I realized that averaging already rendered images – is worse than averaging their float values right after VAE step. So I've came up with the idea of this extension.

Install it, and you will get a new section in WebUI in txt2img/img2img tabs:

It allows you to:

• Do nothing and work as usual.
• Automatically skip a fixed amount of last steps (post-factum after they're done).
• Average some steps before that and replace the returned image (latents will be stored in RAM during generation).
• Create all combinations of skipping and averaging (and dump them to a separate folder), so you'll be able to pick the best settings specific for current image.
• Reliably dump all intermediate steps of generation (without relying on "Live previews" feature), but this mode is slow.
• Quickly compare averaged and original image by returning a combined checked pattern with replaced parts.

Example of averaged result, sampler – PLMS with 32 steps, model – DreamShaper 2.52

Original:

Skipping 3 steps and then averaging 3 steps:

More examples are posted on my extension page.

[chaewai]

Nice, will give it a try. Nothing more frustrating than seeing a good looking image on the live preview and then the final result seeming worse.

[w-e-w]

added to the index

[fractal-fumbler]

hello :)

tried yesterday your extension and results are nice.
should it work with highress fix in webui?

[aleksusklim]

It is working only for second pass, because I'm not sure how to intercept in-between two passes (to cut latents before the second pass takes them).
So, currently I drop anything stored so far, if I detect a new pass during generation.

[fractal-fumbler]

hhhmm, sorry for bothering, it's working as it should: removing overprocessed (burn) effect

i should generate less at late night :D
thanks for your extension :)

[aleksusklim]

Actually, I've changed my mind! I implemented "Stop" slider, which interrupts first-pass at the specified step (counting from the beginning, not from the end), which could be used to simulate at least "Skip" option! (No way to average together at first-pass though).

A great side effect of this – greatly saving time by directly skipping over last steps for draft generations without highres-fix (for example, at 24 steps of DPM++ 2M Karras, a descent draft image can be extracted from step 16; which is different from actually asking 16 steps).
Also it plays nice with averaging, skipping or brute-forcing.

I will push a new release after I'll test everything locally.

[aleksusklim]

Done!

[aleksusklim]

I've pushed an update addressing this bug:
klimaleksus/stable-diffusion-webui-anti-burn#2

Basically, for some samplers my extension detected two half-steps for each actual step, which means that averaging lied (so you had to set twice the needed count).
Now the UI looks like this:

If somebody had used "slow" samplers (like Heun or SDE), now you will need to lower your Count and Skip values to get result nearly-equal to previous version (or set the compatibility checkbox).

Also I've added new "Stop" slider, which aborts the generation right away, but only at the first pass of highres.fix! You can use it to either simulate Skip for lowres pass, or to save time making draft generations (but ensure they will look similar when you'll disable Stop), which especially useful to ancestral unstable samplers.

This is how Stop might be used, prompt a cat BREAK (the dog:1.05) BREAK (child girl:0.8) for suzumehachi model:

Steps: 10, Sampler: Heun, CFG scale: 7, Seed: 164395240, Size: 512x512, Model hash: 58c5c27858, Model: suzumehachi_V10, Denoising strength: 0.5, Hires upscale: 2, Hires steps: 20, Hires upscaler: Lanczos

The second image was utilizing Stop=9 to drop the last step of first pass. This is how lowres versions look like:

You will need to update my extension from WebUI to see new slider.

P.S. If somebody will help me at #7907, I'll make it so loading an image made with enabled Anti Burn – will populate correct settings from the generation info.