NVIDIA CMD — image to video

Autoregressive image-to-video with nvidia/cmd (Context-Matched Distillation): a 4-step causal student distilled from Cosmos-Predict2.5-2B, generating one latent frame at a time with a rolling KV cache.

Give it a first frame and a prompt describing the motion. Released chunk1-short checkpoint · 832×480 · 16 fps · up to 93 frames.

Code · Model card · Non-commercial use only (NVIDIA OneWay Noncommercial License).