ysz2022 / nvedit Goto Github PK

I am wondering how you implement the IP2P+, the mask derived from the difference between conditional and unconditional after first_stage_decode or at the latent space before first_stage_decode, could you please figure it out in detail?

Hello, great work! I'd like to understand whether the final video frames inferred by NVF are of the same length as the original video. If they are, does it mean we still need to generate each frame using the diffusion model? Can I understand it this way: first, use the I2T edit model to edit each frame of the video individually, and then NVF acts as a fine-tuner for the generated edited video to make it more coherent.

Hi! I love your idea that improved ip2p in the second stage, but I don't really understand why you need to use the edited results to optimize the former Neural Video Field, and can you further explain that? Did you do the ablations about the the additional optimization of the NVF in the second stage?

Looking forward to your answer.