Context-Matched Distillation¶

Development line: project:nvidia-cmd · thread context-matched-distillation
Last event: 2026-08-21 · 1 dated since 2026-08-21 · Researched: 2026-09-05 · confidence: high

What it is¶

Context-Matched Distillation (CMD) is NVIDIA’s causal distribution-matching distillation method and checkpoint release for autoregressive image-to-video generation.

• Chunk-1 or chunk-4 generation.
• Long rollouts.
• Camera-conditioned variants.

The released CMD code and derivatives are non-commercial research and education only. The underlying Cosmos base model also requires access approval. Use it to evaluate interactive causal image-to-video research, not as a drop-in commercial video stack.

Development line¶

• 2026-08-21 — Context-Matched Distillation resources published. On 2026-08-21, NVIDIA published the Context-Matched Distillation (CMD) project page, code repository, model page, and image-to-video demo. These links opened the project to public inspection.

What changed¶

• 2026-08-13 — CMD paper introduced causal teacher scoring, Prefix Scoring, and Prefix Corruption for few-step autoregressive video distillation.
• 2026-08-19 — The Hugging Face CMD repository was created.
• 2026-08-21 — CMD was documented with its project page, code repository, model repository, and an image-to-video demo link.
• 2026-08-23 — NVIDIA added the CMD model card and checkpoints, covering short, long, and camera-conditioned chunk-1/chunk-4 variants.
• 2026-08-24 — NVIDIA updated the model card with the arXiv citation.
• 2026-08-30 — The model repository added config.json.

How to use this¶

As of 2026-08-21, evaluate CMD by examining the linked code, model page, documentation, and demo together.

1. Request access to the Cosmos-Predict2.5 2B base model before running CMD. — https://huggingface.co/nvidia/Cosmos-Predict2.5-2B
2. Clone CMD, create the Python 3.10 conda environment, install requirements and FlashAttention, then install the package in development mode. — https://github.com/nv-tlabs/cmd
3. Download all CMD checkpoints, or download only the checkpoint needed for a chosen variant. — https://huggingface.co/nvidia/cmd
4. Match the checkpoint to its supplied configuration, image, prompt, block size, local-attention window, and requested output-frame count; the repository gives a chunk-4 short-video invocation. — https://github.com/nv-tlabs/cmd
5. Run the supplied example runner for one of the short, long, or camera-conditioned variants before replacing the example inputs. — https://github.com/nv-tlabs/cmd

Best practices¶

• Treat checkpoint and configuration pairing as an invariant: t denotes frame duration and l denotes local-attention length, while chunk-1 and chunk-4 use different generation granularity. — https://huggingface.co/nvidia/cmd
• Evaluate long rollouts on representative inputs: autoregressive errors accumulate, degrading temporal consistency, fine detail, faces, hands, text, and fast motion. — https://huggingface.co/nvidia/cmd
• Do not expect the public DL3DV setup to reproduce reported quality; it is only an end-to-end pipeline check, while the reported training data is not released. — https://github.com/nv-tlabs/cmd
• Keep CMD use and derivatives non-commercial unless licensing changes; redistribution must retain the license and notices. — https://github.com/nv-tlabs/cmd/blob/main/LICENSE

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• The Hugging Face commit history dates the initial CMD repository to 2026-08-19 and the checkpoint and model-card upload to 2026-08-23, while the 2026-08-21 event already carried the model URL. The link list may have been enriched after the dated event; this cannot be resolved from the available first-party history.
• The public demo page was reachable as a Hugging Face Space but exposed no reproducible runtime settings or availability guarantee.
• No first-party post-release benchmark or production-deployment evidence was found; paper results remain research evidence rather than a production SLA.

Sources¶

Agent brief¶

• Subject: project:nvidia-cmd, thread context-matched-distillation, 1 dated events 2026-08-21 → 2026-08-21.
• Practical note: As of 2026-08-21, practitioners evaluating CMD should consult the linked code repository, model page, project documentation, and image-to-video demo together rather than treating the announcement as an unsupported research reference.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.