add CFG Rescale (thank you, anonymous user from 4chan)

[AUTOMATIC1111]

Proposed by anon based on this: https://arxiv.org/abs/2305.08891

Adds the setting, xyz plot support and infotext. The setting is at the bottom Sampler Parameters page.

[AUTOMATIC1111]

To also put this in context, this - https://github.com/mcmonkeyprojects/sd-dynamic-thresholding - extension presumably does the same thing but better.

[AUTOMATIC1111]

Although from usability perspective having one slider that you can just set and forget may be more preferable to what dynamic thresholding requires.

[ljleb]

I would like there to be an extension point for combine_denoised so that extensions can add alternative keywords similar to AND to the prompt. Is that already possible without stepping on other extension code?

If that is not the case, I can take a bite at the code. If it is simple to implement, it would be nice to add it to this PR.

[AUTOMATIC1111]

Generally, you can just replace combine_denoised with your function and call the original editing inputs/outputs. If the other extension does the same, and you both properly put the original function back on on_unload, scripts should work just fine together. If you do something that does not involve calling the original function, I don't think any extension point will prevent scripts from stepping on each others' toes.

[ljleb]

I found a good enough solution in the end, which involves recomputing the composable result without cfg std rescale. Thanks anyways!

It's possible the extension code goes out of sync with the repo if other stuff ends up being added. For now, I should not have any trouble when this is merged at least.

[drhead]

I think I have gotten it to work much more closely to what the paper authors demonstrated:

The desaturation is actually far less of an issue if you are using a model trained with zero terminal SNR and v-loss as the paper authors described -- this grid was generated on a SD1.5 model trained on zero terminal SNR for quite a while and on V-loss for a very short time, which is currently somewhat underbaked. For other prompts I've tested, overexposed images look much more reasonable with a value of 0.7, which also as demonstrated makes the saturation consistent at any CFG scale.

I have not had the opportunity to test dynamic thresholding on this model yet to see if it actually does the same thing, but based on my past experiences with that plugin CFG rescale is far easier to use in any case.

[ljleb]

I just want to bring to your attention that the greek letter is causing problems with extensions like civitai. A user reported it in my implementation: ljleb/sd-webui-neutral-prompt#18

Replacing the greek letter with "phi" would prevent metadata issues. It would also make old generations that used my extension for the rescaling work with future webui generations when using paste params.

[6DammK9]

I am deeply regret to not try to preserve this feature, or at least made a viable migration guide *sob
Here is my research: https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch01/dynamic_cfg.md
Long things short (I just started playing sd-dynamic-thresholding), here is the closest parameter to replicate this original feature:

• Mimic CFG Scale: 1 CFG ONE. Hence the awful green-ish effect on high phi.
• Top percentile of latents to clamp: 100
• Interpolate Phi: Same phi
• Mimic Scale Scheduler: Constant
• CFG Scale Scheduler: Constant
• Separate Feature Channels: ON OFF is better. It disables some green-ish effect in some area.
• Scaling Startpoint: MEAN
• Variability Measure: AD I was expecting STD, but AD looks closer

The image previews

• Original:

parameters

(aesthetic:0), (quality:0), (solo:0), (1boy:0), [astolfo], [[ducati]]
Negative prompt: (worst:0), (low:0), (bad:0), (exceptional:0), (masterpiece:0), (comic:0), (extra:0), (lowres:0), (breasts:0.5)
Steps: 48, Sampler: Euler, CFG scale: 4.5, Seed: 4226915900, Size: 768x768, Model hash: ba7e3203e2, Model: marble_v141.fp16, Denoising strength: 0.7, Clip skip: 2, Hires upscale: 1.5, Hires upscaler: Latent, Version: v1.5.1-5-g819dedb7

• Phi 0.3 (cfg-rescale)
  
• mimic 1, phi 0.3, ZERO, STD
  
• mimic 1, phi 0.3, MEAN, AD
  
• mimic 1, phi 0.3, MEAN, AD, Separate Feature Channels OFF