About the loss curves.

[striveAgain]

Hi! Sorry to trouble you again.

Could you please share the loss curves for all stages? I found the curves for ODE and CD in the closed issues, but Stage 1 and Stage 3 are currently missing.

By the way, I initialized the casual model directly with the bidirectional ckpt and run inference without any training. But the results are terribly bad, almost all noise. Is this normal?

[zhuhz22]

Stage 1:

Stage 3:

As for your observation I initialized the casual model directly with the bidirectional ckpt and run inference without any training. But the results are terribly bad, this is quite normal, and is exactly why we need AR diffusion distillation.

[WJohnnyW]

can you show the curve of critic loss and grad of stage 3, I always find the generator becomes better at first, them occurs grid-like, stripe-like artifacts gradually along the training steps (artifacts fist at the end of rollout, then the mid and end, finally full rollout, you can see #41 for example). After debugging, I think the problem may be the critic can't catch up the generator, I try to increase the learning rate of critic same with generator. the phenomenon alleviates but still occurs after longer training. Can you provide some insight for this problem? @zhuhz22

[zhuhz22]

@WJohnnyW