Question about interior conditions and masking

[tung-nd]

In the frame_prediction_cond function, you specify both cond_frames and cond_latent_frames. However, from the paper, cond_latent_frames should not be specified manually, and whether a latent token is considered part of interior conditions depends on whether the supervoxel of that token contains the conditioning pixels. Can you please explain this? For a video with 16 frames, if I want to use different numbers of frames (from 1 to 15) as conditions, how should I set the value of cond_latent_frames? Thank you so much! @Lijun-Yu @roadjiang