Thanks a lot for your great work! I’m trying to reproduce this method and ran into a question.
Wan2.2 backbone uses latent chunks of shape [9, 44, 80, 48]. 9 stands for latent time steps, 44 and 80 are compressed latent height and width, while 48 is the channel number.
However, latents read from latent space memory are calculated per actual RGB frame, resulting in a [33, 44, 80, 48] tensor. This creates a temporal dimension mismatch with the backbone input.
The paper doesn’t mention how to align the 33-frame memory latent to the 9-time-step format required by the backbone.
Do we need to add a temporal aggregation operation for this? Thanks!
Thanks a lot for your great work! I’m trying to reproduce this method and ran into a question.
Wan2.2 backbone uses latent chunks of shape [9, 44, 80, 48]. 9 stands for latent time steps, 44 and 80 are compressed latent height and width, while 48 is the channel number.
However, latents read from latent space memory are calculated per actual RGB frame, resulting in a [33, 44, 80, 48] tensor. This creates a temporal dimension mismatch with the backbone input.
The paper doesn’t mention how to align the 33-frame memory latent to the 9-time-step format required by the backbone.
Do we need to add a temporal aggregation operation for this? Thanks!