LMS Karras introduces noise when done

[guzuligo]

Describe the bug
When using LMS Karras with high sampling number, it produces noise on the image right when it is done, ruining the image it produced.

To Reproduce
Try using the following prompt.

banana AND car
Steps: 70, Sampler: LMS Karras, CFG scale: 7, Seed: 1745286955, Size: 576x320

[HanJammer]

Yes, I noticed it yesterday too when testing different samplers (with 50 steps).

[dfaker]

The extended sampler settings using Karras in the UI are best approximations of good default parameters.
If you'd like to experiment with https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts#alternate-noise-schedules you may find better parameters.

[Adil3tr]

I've had the same issue. I was making images with LMS at 20 CFG and 50 steps and having no issues for weeks, but today I make ones at 12 CFG and 60 steps and I'm still getting artifacts. There is definitely something wrong.

[Adil3tr]

The extended sampler settings using Karras in the UI are best approximations of good default parameters. If you'd like to experiment with https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts#alternate-noise-schedules you may find better parameters.

This is a real issue, LMS should not be doing this and didn't until very recently.

[dfaker]

This is a real issue, LMS should not be doing this and didn't until very recently.

That's an important distinction, I'll try to hunt down some old reference generations.

[aeloi]

I've also had this issue as of late...

[quasiblob]

This seems to be happing still, and I have pulled the latest version.

[Adil3tr]

If it's not fixable, they should maybe remove LMS from Automatic, because it has really serious artifacting most of the time right now.

[TheOnlyHolyMoly]

Does the issue persist in the current version? Can anyone provide a commit-hash from a version where this worked (plus image generation parameter)? The commit should be recorded in the image's meta data

#2856 looks like there are multiple filed issues around that topic anyhow. #2856 also proposes the solution that Invoke has used for the matter.

[catboxanon]

This is related to k-diffusion specifically. The root issue is the k-diffusion implementation, as ComfyUI has the same issue, so this is not webui related.

Read the comment I have posted there for further info. Closing.