Is 8step model higher quality than the 4step one ?

#4
by NielsGx - opened

If running both at their intended number of steps, is the 8step one better at being close to the original model ?

Or maybe I should try running the 4step model at 8steps ?

NielsGx changed discussion title from Is 8step model high quality than the 4step one ? to Is 8step model higher quality than the 4step one ?

Technically, the 4-step model should not be used in 8-step inference. But the test shows it can be used in 8-step setting.

We are not confident which model you should choose. Howerver, using 8-step model for 8step inference and 4-step model for 4step inference is a good choice.

We are going to release a new 8-step model, it should be better.

Thank you.

I thought using more steps would make same things as using more steps in base model (add more detaisl and get even sharper) ?
But I guess it's different here...

Also hopefully the new model will work with torch.compile (see my other post)
Or it's comfyUI issue

Also the 4step lora has dak results in comfy? Also I am not getting what you mean by "Howerver, using 8-step model for 8step inference and 4-step model for 4step inference is a good choice."

the 4 step model is fast at generating very static and very identical images. Early days, i know.

Sign up or log in to comment