1.23.3

fix min snr logic, thanks to @justinlovelace

1.23.2

just force train_step to only accept kwargs

1.23.1

add token shift along time for feedforwards in transformer blocks in …

1.23.0

substitute p2 loss weight with min snr loss weight

1.22.4

allow for disabling memory efficient way of keeping only one unet in …

1.22.3

allow for disabling memory efficient way of keeping only one unet in …

1.22.2

cleanup

1.22.1

allow for using Lion optimizer

1.22.0

adopt a new promising attention stabilizing trick out of Google Brain

1.21.6

automatically temporally downsample the conditioning video frames, ei…