PhyWorld: Physics-Faithful Video World Model

Project page · Paper · Code and release status · PhyGround benchmark

PhyWorld is a physics-faithful video world model post-trained from Wan2.2-I2V-A14B. It is designed to continue an observed scene while improving temporal coherence and physical plausibility.

The model is trained in two stages:

1. Video-continuation flow matching to learn coherent continuation from an observed video prefix.
2. Physics-preference DPO using preference pairs derived from the PhyGround human-annotation pool.

Usage status: this repository contains the released checkpoint, but it is not a generic DiffusionPipeline and the public code repository does not yet contain a verified end-to-end PhyWorld inference command. Download and file inspection are supported below. Follow the GitHub release status for the official inference entry point rather than substituting a Wan2.2 or generic Diffusers command.

Model details

Download

Install the current Hugging Face CLI and download the snapshot:

pip install -U huggingface_hub

hf download NU-World-Model-Embodied-AI/phyworld \
  --local-dir ./phyworld

The complete snapshot is approximately 70 GB. To inspect only the model card and configs before downloading weights:

hf download NU-World-Model-Embodied-AI/phyworld \
  README.md configuration.json \
  high_noise_model/config.json low_noise_model/config.json \
  --local-dir ./phyworld-config

Checkpoint layout

phyworld/
├── high_noise_model/
│   ├── config.json
│   └── diffusion.safetensors
├── low_noise_model/
│   ├── config.json
│   └── diffusion.safetensors
├── Wan2.1_VAE.pth
├── models_t5_umt5-xxl-enc-bf16.pth
├── google/umt5-xxl/
└── configuration.json

The two denoiser directories follow the high-noise/low-noise expert layout used by Wan2.2. The snapshot also includes the VAE and UMT5 text-encoder artifacts required by the released checkpoint. File names are kept as released for compatibility.

Reported results

PhyGround physical faithfulness

Scores use the released PhyJudge-9B evaluation protocol on a 1–5 scale.

VBench general video quality

These are the authors' reported results under the paper's protocol. They are not a guarantee of performance on different prompts, resolutions, inference implementations, or downstream domains.

Training data

Stage 1 uses OpenVid-1M clips filtered by inter-frame CLIP similarity and optical-flow magnitude. Stage 2 uses 1,000 class-balanced preference pairs derived from PhyGround human annotations across seven event classes. See the paper for filtering thresholds, DPO settings, and ablations.

Intended use

PhyWorld is intended for:

• research on physically faithful video generation and world models;
• controlled comparison with Wan2.2 and other video-generation baselines;
• evaluation on PhyGround and general video-quality benchmarks;
• further research on physics-aware post-training.

This is a research checkpoint, not a validated physics simulator, safety model, or robot controller. Generated video should not be treated as evidence that an event is physically possible or as a substitute for simulation or real-world testing.

Limitations

• PhyWorld can still produce physically implausible motion, object changes, and long-horizon inconsistencies.
• The reported physical improvement is benchmark- and prompt-dependent.
• PhyGround covers 13 observable laws across solid-body mechanics, fluids, and optics; it does not cover all physical phenomena.
• The public repository does not yet provide a verified end-to-end inference script, a minimum-memory configuration, or a complete reproducibility recipe.
• Biases and coverage limitations in Wan2.2, OpenVid-1M, and the preference data may be inherited by this checkpoint.

License

This checkpoint is released under the Apache License 2.0. PhyWorld is derived from Wan2.2-I2V-A14B, which is also released under Apache-2.0. Users remain responsible for reviewing all applicable upstream terms and the terms of any data or software used with the model.

Citation

@misc{zhao2026phyworld,
  title         = {PhyWorld: Physics-Faithful World Model for Video Generation},
  author        = {Pu Zhao and Juyi Lin and Timothy Rupprecht and Arash Akbari and Chence Yang and Rahul Chowdhury and Elaheh Motamedi and Arman Akbari and Yumei He and Chen Wang and Geng Yuan and Weiwei Chen and Yanzhi Wang},
  year          = {2026},
  eprint        = {2605.19242},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2605.19242}
}