Prism: image + text → video with audio

Model · Code · Paper. Checkpoint: preview_alpha. Runs on one ZeroGPU xlarge (RTX Pro 6000, 96GB). xlarge uses quota at 2×, so keep clips short. Sound can be tagged in the prompt with <music>, <sfx>, <speech>.

Resolution
25 121
10 50
1 10
1 17
1 15
Block-sparse attention sparsity
Examples
Reference image (first frame) Prompt (video)