todo

[rosikand]

• try with hugging face u-net
• try the artbench or cifar data with the HF pipeline to see how well that data can be learned and with how many samples needed.

[rosikand]

Debugging:

• [x] normalize to [-1, 1] instead of /255.
• gradient accumulation

note: forward diffusion working... cross-tested with lucidrains implementation.

[rosikand]

• [x] add more variance scheduler options:

[rosikand]

things to try:

• use the same noise vector sample for each sample in the batch