nvidia/cmd

Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillation

Hmrishav Bandyopadhyay^1,2, Xuanchi Ren¹, Zijian Huang¹, Jay Zhangjie Wu¹, Tianshi Cao¹, Ruilong Li¹, Bryan Chu¹, Sanja Fidler¹, Yi-Zhe Song², Zian Wang¹

¹NVIDIA
²SketchX, CVSSP, University of Surrey

Project Page | Code

Model Overview

Description

Context-Matched Distillation (CMD) is a method for autoregressive video distillation using causal teachers. This repository provides CMD checkpoints based on Cosmos-Predict2.5 2B for causal image-to-video generation.

The release includes chunk-1 and chunk-4 models for short- and long-duration generation, together with camera-control variants. The models use local temporal attention and generate video autoregressively in causal chunks.

This model is for research and development only.

License and Terms of Use

This model is released under the NVIDIA OneWay Noncommercial License. It may only be used for non-commercial research or educational purposes.

Deployment Geography

Global

Intended Use

Primary users include researchers and developers studying:

• Autoregressive video generation.
• Causal video distillation.
• Long-duration video generation.
• Interactive world models.
• Camera-controlled video generation.

The checkpoints are not intended for safety-critical applications or uses prohibited by the model license.

Available Checkpoints

Checkpoint	Configuration	Intended use
chunk1_short_t24_l21.safetensors	t24_l21	Chunk-1 short-duration generation
chunk4_short_t21_l16.safetensors	t21_l16	Chunk-4 short-duration generation
chunk1_long_t126_l21.safetensors	t126_l21	Chunk-1 long-duration generation
chunk4_long_t121_l16.safetensors	t121_l16	Chunk-4 long-duration generation
chunk1_camera_control_t32_l21.safetensors	t32_l21	Chunk-1 generation with camera control
chunk4_camera_control_t29_l24.safetensors	t29_l24	Chunk-4 generation with camera control

In the configuration names, t denotes frame duration and l denotes local attention. Checkpoint names do not include dataset names or training iteration information.

Quick Start

Install the reference implementation:

git clone https://github.com/nv-tlabs/cmd.git
cd cmd

conda create -n causal-cosmos python=3.10 -y
conda activate causal-cosmos
python -m pip install -r requirements.txt
python -m pip install flash-attn --no-build-isolation
python setup.py develop

Download the checkpoints:

hf download nvidia/cmd --local-dir checkpoints

To download one model only:

hf download nvidia/cmd \
  chunk1_short_t24_l21.safetensors \
  --local-dir checkpoints

Follow the inference instructions in the CMD repository to select the matching model configuration and run generation.

Model Architecture

Architecture type: Causal video diffusion transformer

Base model: Cosmos-Predict2.5 2B

Generation strategy: Block-autoregressive video generation

CMD uses causal temporal attention to generate video sequentially. Chunk-1 and chunk-4 variants provide different autoregressive generation granularities, while the camera-control variants add camera-motion conditioning.

Inputs

Standard checkpoints accept:

• An initial image.
• A text prompt.

Camera-control checkpoints additionally accept a camera trajectory in the format expected by the CMD inference pipeline.

Outputs

The model produces an RGB video conditioned on the initial image, text prompt, and, when applicable, camera trajectory. Output duration is determined by the selected t configuration.

Software Integration

• Framework: PyTorch
• Checkpoint format: Safetensors
• Reference implementation: nv-tlabs/cmd
• Preferred operating system: Linux
• Recommended hardware: NVIDIA CUDA-capable GPU

The checkpoints must be used with the corresponding model architecture and configuration supplied by the reference implementation.

Training Method

CMD uses a two-stage procedure:

1. Causal teacher pretraining.
2. Context-matched distillation.

See the CMD repository and project page for the method and evaluation details.

Limitations

• Autoregressive errors may accumulate during long rollouts.
• Visual quality and temporal consistency may degrade over time.
• Generated motion may be physically implausible.
• Fine details, text, faces, hands, and rapid motion may contain artifacts.
• Prompt adherence may vary with scene complexity.
• Camera-control outputs may deviate from the requested trajectory.
• Generated content may reflect biases present in the training data.

Users should evaluate the model on data representative of their intended application before deployment.

Ethical Considerations

Generated video may be mistaken for authentic footage. Users should clearly identify synthetic content where appropriate and must not use the model for deception, impersonation, harassment, misinformation, or violations of privacy, intellectual-property, or other applicable rights.

NVIDIA believes Trustworthy AI is a shared responsibility. Developers should ensure that their use of this model complies with applicable laws, policies, licenses, and safety requirements.

Please report model quality, risk, security vulnerabilities, or NVIDIA AI concerns through the NVIDIA support portal.

Citation

If you find this work useful, please cite:

@article{bandyopadhyay2026context,
  title   = {Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillation},
  author  = {Bandyopadhyay, Hmrishav and Ren, Xuanchi and Huang, Zijian
             and Wu, Jay Zhangjie and Cao, Tianshi and Li, Ruilong
             and Chu, Bryan and Fidler, Sanja and Song, Yi-Zhe
             and Wang, Zian},
  journal = {arXiv preprint arXiv:2608.13391},
  year    = {2026},
  eprint  = {2608.13391},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url     = {https://arxiv.org/abs/2608.13391}
}

Acknowledgements

This project builds on Self-Forcing and NVIDIA Cosmos-Predict2.5.