Variance Scheduler

[sai-advaith]

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]

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

[bghira]

DeepFloyd was trained with a cosine schedule.