search

AUTOMATIC1111 / stable-diffusion-webui

Stable Diffusion web UI
GNU Affero General Public License v3.0
131.71k stars 25.28k forks source link

[Feature Request]: #15624

Closed risedangel closed 3 weeks ago

risedangel commented 3 weeks ago

Is there an existing issue for this?

What would your feature do ?

NVIDIA align your steps sampler support ?

Proposed workflow

  1. Image generation should allow the user to select NVIDIA align your steps sampler

Additional information

You may find the details here : https://research.nvidia.com/labs/toronto-ai/AlignYourSteps/ Also it seems comfui already supports this: https://github.com/comfyanonymous/ComfyUI/commit/644a3ae58d426ffbbc02ef4104034c98e8fc6513

AG-w commented 3 weeks ago

...have you checked pull request?

risedangel commented 3 weeks ago

I believe i did. But maybe i am being mistaken?

nizo303 commented 3 weeks ago

Check pull request https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/15608. (KL Optimal scheduler)

risedangel commented 3 weeks ago

İ didnt know that "kl scheduler" included align your steps. The issue can be closed. Thank you

drhead commented 3 weeks ago

İ didnt know that "kl scheduler" included align your steps. The issue can be closed. Thank you

To be clear, this PR doesn't actually include proper Align Your Steps support -- it includes the theoretically better derived scheduler, but doesn't include the option to use a trained timestep schedule which is the actual focus of the paper. I'm still experimenting with that part of the paper -- I don't really see a point in implementing it until I am able to actually reproduce the training code to create that schedule and thus reproduce the paper's findings, because until that happens it's hard to say how much AYS generalizes to different finetunes.