[bug]: Deformed Images, Noise in final image when using karras samplers and high steps.

[WhatsHacking]

Is there an existing issue for this?

• [X] I have searched the existing issues

OS

Linux

GPU

cuda

VRAM

8GB

What happened?

After completing the instalation It seems to me there might be a problem regarding the denoising schedulers, and images come out with lots of noise when using k diffusion and high steps (<50) I tried copying the Horse Image from the toolkit example page, and the result was this atrocity... I copied the entire prompt, seed, image size and cfg, but instead of a pony I got a albino chupakabra. Also I tried asking for help in the discord but a bot kept flagging me for explicit language or imagery, so thats why im posting this here.

Screenshots

No response

Additional context

No response

Contact Details

No response

[lstein]

Which k sampler are you using?

[lstein]

I think that what's happened here is that between the current release and the last one we did indeed change the k-diffusion noise schedule to allow the samplers to work better at low (<30) step counts. It looks like for this specific prompt the k_euler_a sampler creates a monstrosity. Either increasing the step count modestly, or switching to another k-sampler returns a more acceptable image. If you are using the CLI, you can restore the original noise schedule by passing --karras_max=0 at the invoke> line. Currently there is no easy way to do this on the Web side, but this was an oversight and I'm fixing it now.