About VAE distill training detail

[Leoooo333]

Thanks for the lightx2v's great work! I want to reimplement taew2_2 from scratch, and just saw lightx2v have a better lighttaew2_2 then their original ckpt. Currently I am training under the same architecture, would you mind to share the training details or any suggestion how to improve it? I am currently train it without GAN first at very low resolution, and then finetune with GAN loss. But seems still not as good as them.

[gushiqiao]

From my experience, LPIPS loss is crucial.In addition, increasing the resolution and num_frames also helps improve performance.