diff_loss_mid_down_blocks
This is model is a demonstration of an alternative loss function for video training, diff_loss. Unlike the typical frame by frame reconstruction loss, diff_loss ensures the frame by frame differences of the target and predicted noise are the same.
This model is a fine-tuned version of the AD 1.5 v2 model. It was trainined for 3600 steps with LR of 9e-7 and a batch size of 16. Only the down and mid blocks were trained.
For more details about this model see the associated blog post here