Training Details

[yzhoulv]

Thanks for your great work. I have a question: in Stage 1, there is no high-quality video used during training, so the model does not have the ability to repair videos. However, in Stage 3, the DMD loss is used to distill the model into a one-step model, where the training target is the latent features, and the decoder is not trainable. The gradient from the LPIPS loss will also affect the latents, which is inconsistent with the standard DMD setting.
What confuses me is whether high-quality videos are used in Stage 1 and Stage 2.