You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Lightning models are fixed at 4-8 steps of the loras baked into them, and in advanced ksamplers the denoise is regulated by the start_at_steps, so with low denoise the quality drops noticeably. It would be better to use a regular ksampler with a constant number of steps. Even for other models.
The text was updated successfully, but these errors were encountered:
I don't know a good solution to make img2img work well with Lightning models. What you're suggesting is not different from increasing total steps. For example running 8-step lightning model at 50% denoise with 16 total steps -> results in 8 actual steps. Doesn't actually look good though in my experience, the noise schedule is too diferent from what Lightning was trained on.
It helps a bit to use a uniform scheduler (not Karras).
I use the lightening models at 45 and they look pretty good especially after a couple of passes but literally go back to that noisy hazy quality at 40 or 50 or any other level.
Lightning models are fixed at 4-8 steps of the loras baked into them, and in advanced ksamplers the denoise is regulated by the start_at_steps, so with low denoise the quality drops noticeably. It would be better to use a regular ksampler with a constant number of steps. Even for other models.
The text was updated successfully, but these errors were encountered: