LMS Karras introduces strange artifacts looking like a mess of colors

[Starcatcher2009]

28 steps, 2 last layer/clip ignore thingy, cfg 11, seed -1. Uses Intel (R) Iris (R) iGPU. I cannot describe further since the image is nsfw, however I may privately send you.

[Starcatcher2009]

Which is a shame since it seems like k_lms is so much better than Euler/Euler a

[kalkal11]

Not all samplers work well with low steps, having said that, it seems to be fine for me with these settings which is as far as we can test with the given information without the prompt

[Starcatcher2009]

i don't think 28 is low
plus it seems like LMS Karras (or k_lms, don't know if that's similar) is the default used in NovelAI along with 28 steps

[Starcatcher2009]

Not all samplers work well with low steps, having said that, it seems to be fine for me with these settings which is as far as we can test with the given information without the prompt

Also would you like to check my nsfw picture along with the prompt?

[Starcatcher2009]

Or it may be that the AI suddenly decides to be concerningly creative and decides to spit out random blocks of colors, but that's unlikely

[kalkal11]

Not all samplers work well with low steps, having said that, it seems to be fine for me with these settings which is as far as we can test with the given information without the prompt

Also would you like to check my nsfw picture along with the prompt?

Not really but if it's really an issue, you should be able to produce an alternative sample/samples

[khanonnie]

k_lms and its variants are prone to these types of blocky/mosaic-like artifacts when steps are too low for the selected CFG scale setting. Not a bug, k_lms just requires higher step counts than the other samplers when CFG is higher than ~9.

[Starcatcher2009]

k_lms and its variants are prone to these types of blocky/mosaic-like artifacts when steps are too low for the selected CFG scale setting. Not a bug, k_lms just requires higher step counts than the other samplers when CFG is higher than ~9.

It also seems like running CPU-only makes the steps less refined, and so might require possibly even more steps. So I might take it to like 40-50+ steps

[lockmatrix]

k_lms and its variants are prone to these types of blocky/mosaic-like artifacts when steps are too low for the selected CFG scale setting. Not a bug, k_lms just requires higher step counts than the other samplers when CFG is higher than ~9.

Maybe should add a hint contain these info to the sampler on ui

[Starcatcher2009]

Also where's the "disable vae2.pt" in settings? Like, which category?

[Starcatcher2009]

Also, is it just me or each time I write something anal in the prompt it usually shows up 2 of them? A big one and a much smaller one near the vagina.

[Starcatcher2009]

And does smaller resolutions results in worse pictures?

[Chryseus]

k_lms and its variants are prone to these types of blocky/mosaic-like artifacts when steps are too low for the selected CFG scale setting. Not a bug, k_lms just requires higher step counts than the other samplers when CFG is higher than ~9.

That doesn't appear to be the case, it's getting worse as the steps go up, something is wrong with k_lms, I never had this issue a few weeks back.

50 steps

100 steps

150 steps

Prompt: star trek enterprise ship, 4k, highly detailed, space
Negative: text. not in frame, blurry, bad geometry, ugly
CFG Scale: 14
Seed: 2297975078
Resolution: 512x512

[Starcatcher2009]

k_lms and its variants are prone to these types of blocky/mosaic-like artifacts when steps are too low for the selected CFG scale setting. Not a bug, k_lms just requires higher step counts than the other samplers when CFG is higher than ~9.

That doesn't appear to be the case, it's getting worse as the steps go up, something is wrong with k_lms, I never had this issue a few weeks back.

50 steps 100 steps 150 steps

Prompt: star trek enterprise ship, 4k, highly detailed, space Negative: text. not in frame, blurry, bad geometry, ugly CFG Scale: 14 Seed: 2297975078 Resolution: 512x512

At which point the picture becomes best? Or is it actually in reverse and is most beautiful at 1 step?

[Chryseus]

At which point the picture becomes best? Or is it actually in reverse and is most beautiful at 1 step?

Somewhere between 20-30 steps looks best, it really starts to be noticeable above 40 steps, if I have it show previews however it looks fine as it progresses even with 150 steps, it only becomes corrupt when making the final image.
I've used 100+ steps in the past without this issue appearing.

[Chryseus]

I found the issue, running it with --no-half appears to solve the problem which is a bit weird as the other samplers seem to work fine without it.
I'm now using --medvram --no-half

[jmp909]

i noticed if you set eg Show image creation progress every N sampling steps to 5 and watch it build up it looks fine, but on saving the image gets corrupted

used 100 steps here, and preview every 5

last few steps:

final output saved to disk:

[ryakr]

Issue is still occurring, even on high end cards with --no-half. My 3090 just cooks it throughout the whole process on LMS which is very strange. Im using the settings I have always used and found to work on my model but now it explodes on even a clean install of auto. You can see it just dying in the preview generation. Multiple users in the Furry Diffusion discord have also reported this issue.

Steps 70, cfg 10, normal LMS/ LMS Karras both encounter same issue.

XLfcFIGpFM.mp4

[MadMinstrel]

InvokeAI had the same problem, and they fixed it here: invoke-ai/InvokeAI#1086

[TheOnlyHolyMoly]

@vladmandic not top of the list, but are you aware of this issue and the solution?

[vladmandic]

I wasn't aware.
Fix seems trivial, but only if done upstream in k_diffusors. Doing it downstream is really messy and I don't want to create a kdiffusion fork that I'd have to maintain.

I have another proposal right now that I need to followup on - to make Karras in general a checkbox and remove Karras schedulers in general as in theory any scheduler can work in Karras and standard mode.

[TheOnlyHolyMoly]

yes, have seen that looks really charming.

[catboxanon]

Duplicate of #1973

[ani917]

I found that the problem is that the image always has color artifacts, but when using LMS it becomes 35 times more noticeable