PhyWorld — Physics-Faithful World Model for Video Generation

PhyWorld is a video-generation world model designed to produce temporally coherent and physically faithful scene continuations. Starting from Wan2.2-I2V-A14B, we post-train in two stages: (1) flow-matching fine-tuning for video-to-video continuation, then (2) Direct Preference Optimization (DPO) over physics preference pairs sourced from the PhyGround human-annotation pool.

On the PhyGround physical-faithfulness benchmark PhyWorld reaches 3.09 overall vs. 2.99 for the strongest open baseline; on VBench it reaches 0.769 vs. 0.756 or below for SOTA baselines.

Companion artifact	Where
Live project page	nu-world-model-embodied-ai.github.io/PhyWorld
Model checkpoint	🤗 NU-World-Model-Embodied-AI/phyworld
Paper	arXiv:2605.19242
Sibling benchmark	PhyGround — the benchmark + PhyJudge-9B judge we evaluate against

---

What's in this repo

This repository currently hosts the public project page for PhyWorld. The page is served via GitHub Pages from main and includes the abstract, method overview, both leaderboards (VBench and PhyGround), a featured carousel, and a 4-model head-to-head video grid (PhyWorld vs. Cosmos, LTX, OmniWeave) across 17 shared prompts.

index.html        # the project page
static/css/       # Bulma + PhyWorld custom styles
static/js/        # carousel + FontAwesome loaders
static/videos/    # PhyWorld and baseline mp4 clips
static/images/    # favicons, teaser/framework figures (drop-in)

Training, inference, and evaluation code will land in this repo as the project ships. For now, the model checkpoint is available on Hugging Face and evaluations can be reproduced with the PhyGround pipeline and its released PhyJudge-9B judge.

---

Method

Stage 1 — Physical Consistency Enhancement. Flow-matching fine-tuning on a video-to-video continuation pipeline. The conditioning video is encoded by the Wan-VAE, a binary mask delimits preserved vs. synthesized frames, and a CLIP global-context embedding is injected via decoupled cross-attention. Training data is OpenVid-1M, filtered for inter-frame CLIP cosine similarity and per-clip UniMatch optical-flow magnitude.

Stage 2 — Physics Enforcement via DPO. A rank-16 LoRA adapter wraps the Wan2.2 denoiser's attention and feed-forward projections. Preference pairs are derived from the PhyGround human-annotation pool: within-prompt winner/loser pairs with score margin ≥ 1.0, organized into a 1,000-pair class-balanced trainset over seven physical-event classes (collision/rebound, destruction/deformation, fluids, shadow/reflection, chain, rolling/sliding, throwing/ballistic). DPO is restricted to the high-noise window t ∈ [901, 999] with β = 100 to suppress reward hacking. Training: 16 × H100, 2 epochs, AdamW lr 1e-5.

---

Results

VBench — generation quality

Model	Subject cons.	Background cons.	Motion smooth.	Dynamic	Aesthetic	Imaging	Avg.
PhyWorld (ours)	0.932	0.944	0.986	0.564	0.555	0.632	0.769
Wan2.2-I2V-A14B	0.912	0.928	0.977	0.554	0.543	0.622	0.756
Cosmos-14B	0.899	0.923	0.973	0.559	0.536	0.629	0.753
LTX-2.3-22B	0.894	0.918	0.982	0.549	0.532	0.626	0.751
OmniWeaving	0.903	0.907	0.972	0.556	0.541	0.621	0.750
Cosmos-2-2B	0.887	0.905	0.964	0.543	0.524	0.608	0.739

PhyGround — physical faithfulness (PhyJudge-9B, 1–5 scale)

Model	SA	PTV	Persist.	Solid-Body	Fluid	Optical	Overall
PhyWorld (ours)	2.78	3.07	3.23	2.84	3.04	3.57	3.09
Wan2.2-I2V-A14B	2.72	2.97	3.08	2.79	3.03	3.36	2.99
Cosmos-14B	2.60	2.73	3.07	2.72	2.92	3.53	2.80
OmniWeaving	2.68	2.73	2.92	2.71	2.99	3.13	2.78
LTX-2.3-22B	2.63	2.79	2.91	2.55	3.02	3.21	2.72
Wan2.2-TI2V-5B	2.48	2.70	2.76	2.61	3.01	3.45	2.68
LTX-2-19B	2.50	2.62	2.79	2.49	3.01	3.09	2.62

---

Local preview

The project page is static — clone and open it directly:

git clone https://github.com/NU-World-Model-Embodied-AI/PhyWorld.git
cd PhyWorld
python3 -m http.server 8000
# then visit http://localhost:8000

Pushing to main triggers a GitHub Pages rebuild within ~30–60 seconds.

---

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}
}

---

Acknowledgements

PhyWorld is post-trained from Wan2.2-I2V-A14B and evaluated with the PhyGround benchmark and its PhyJudge-9B judge. The project page template is adapted from OpenVLA, which is based on Nerfies.