Open yja1 opened 1 month ago
In next verision, we will release a new vae to figure out this problem.
In next verision, we will release a new vae to figure out this problem. (What time is it approximately) the reason for this is conv3d input size too big.
In next verision, we will release a new vae to figure out this problem.
can open-sora-plan train one-step generation like ADD or LCM
1 gpu ,generate 1 prompt, 10 step,24FPS,29frames 480p transformer part 38s but vae decode 13min(one of ResnetBlock3d cost 126s ; one upsample cost 96s)