This series began with a perplexing body-flipped video…

VideoCrafter2 Extended by FIFO-Diffusion
"a person swimming in ocean, high quality, 4K resolution."
Inconsistancy in FIFO-Diffusion (frame 85-95)

How can I fix visual inconsistency in long video generation in FIFO-Diffusion? I rolled up my sleeves and started digging around..

After some exploring, I uncovered a few tricks that can help (no cats were harmed because I don’t have one.):

  1. Seeding the initial latent frame
  2. Weighted Q-caches
  3. Extending the Latent Uniformly

Code Implemention

Updated:

Comments