search

deep-floyd / IF

Other
7.63k stars 495 forks source link

Variance Scheduler #144

Open sai-advaith opened 9 months ago

sai-advaith commented 9 months ago

I am using DeepFloyd locally with Stage I medium, T5 medium, and stage II medium. I noticed that if I were to modify the variance scheduler for sampling to linear from cosine, the sample results are much worse than with a cosine scheduler.

Is there a reason why this happens? And, is there a way DeepFloyd can work well with a linear scheduler too?

ShaochengShen commented 6 months ago

i discover the same question too.Do u have any progress on this issue?

bghira commented 4 months ago

DeepFloyd was trained with a cosine schedule.