Skip to content

Conversation

AmericanPresidentJimmyCarter

Skip layer guidance can dramatically enhance the quality of output videos.

output_combined.mp4

@AmericanPresidentJimmyCarter
Copy link
Author

@deepbeepmeep Unfortunately I don't have time to add it to the gradio, but it should be pretty simple to do so. Maybe you can merge and take care of that?

kijai added a commit to kijai/ComfyUI-WanVideoWrapper that referenced this pull request Mar 13, 2025
Credits to AmericanPresidentJimmyCarter: deepbeepmeep/Wan2GP#61
@AmericanPresidentJimmyCarter
Copy link
Author

AmericanPresidentJimmyCarter commented Mar 13, 2025

output_sm.mp4

Other layer skips, since people were asking for it. Most of them are detrimental aside from 9 or 10.

@jffu
Copy link

jffu commented Mar 14, 2025

Hi @AmericanPresidentJimmyCarter, very amazing work! I did simple tests and the results indeed improved. Could you kindly explain the reason behind this?

@FurkanGozukara
Copy link

If anyone else testing with more examples can post here would be nice

I wonder if worth to implement or not

@AmericanPresidentJimmyCarter
Copy link
Author

Hi @AmericanPresidentJimmyCarter, very amazing work! I did simple tests and the results indeed improved. Could you kindly explain the reason behind this?

There's a lot of discussion in the reddit thread.

https://old.reddit.com/r/StableDiffusion/comments/1jac3wm/dramatically_enhance_the_quality_of_wan_21_using/

@deepbeepmeep
Copy link
Owner

Many thanks for your contribution. Do you have more examples that illustrate the quality gain ?

@AmericanPresidentJimmyCarter
Copy link
Author

output_anime_combined.mp4

Here is one yesterday with skipping layer 10, with anime style. You can see that the skipped layer version is less soupy, has more consistent lines, and smoother animation.. I've been meaning to play with it more but haven't had the chance. Sometimes you need to set it to not do the first 10% or so of timesteps, because it can severely disrupt the planning stage of denoising. Below is one where I had layer 10 skipped throughout 100% of the steps.

output_slg_anime_0_1.mp4

@deepbeepmeep deepbeepmeep merged commit 5f07a70 into deepbeepmeep:main Mar 14, 2025
@deepbeepmeep
Copy link
Owner

Thanks, slg, is now part of the gradio webserver of Wan2GP as it seems to improve the quality (it is a placebo ?). I did commit new features as well for Wan2GP. Hopefully I didnt break your command line version.

@NonaBuhtig
Copy link

NonaBuhtig commented Mar 15, 2025

@FurkanGozukara

I wonder if worth to implement or not

Of course you'll do it, won't you? I'm curious, have you ever submitted a single useful pull request, or do you simply make money by profiting from the labor of others?

@FurkanGozukara
Copy link

@FurkanGozukara

I wonder if worth to implement or not

Of course you'll do it, won't you? I'm curious, have you ever submitted a single useful pull request, or do you simply make money by profiting from the labor of others?

modelscope/DiffSynth-Studio#442

modelscope/DiffSynth-Studio#445

image

@NonaBuhtig
Copy link

NonaBuhtig commented Mar 15, 2025

I must admit, I'm impressed by the fact you've made a few PRs. 🫨
Maybe, things are changing ...

I'm not sure what you're trying to achieve by showing my profile. I'm not a dev, and I don't make money by profiting from the work of others like you do.

@AmericanPresidentJimmyCarter
Copy link
Author

AmericanPresidentJimmyCarter commented Mar 16, 2025

Having messed around extensively with this now, it definitely works. The effects are best if you're using teacache 0.15-0.25, which can degrade the model output significantly.

@AmericanPresidentJimmyCarter
Copy link
Author

Oh, someone published a paper on this too. I hadn't read it before implementing it, but it may convince some people of the potential utility.

https://arxiv.org/pdf/2411.18664

@vTuanpham
Copy link

Mr president @AmericanPresidentJimmyCarter, any plan to implement this in ComfyUI ?

@vTuanpham
Copy link

peteromallet pushed a commit to peteromallet/Headless-Wan2GP that referenced this pull request Jun 4, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

6 participants