WorldAtlas: A Unified Large-Scale Benchmark for
World Generation toward World Intelligence

OpenEnvision

Is your world model a consistent, controllable, and physically plausible scene renderer?

---

TL;DR — WorldAtlas evaluates the published world-model integrations on image-conditioned generation (image_static, image_dynamic, plus a closed-loop memory track at 60s and 30s) across 5 dimensions.

📢 News

• [2026/10/15] 🚀 WorldAtlas is now open-sourcing as a preview! Released the dataset, evaluation code, model weights, leaderboard, and dataset gallery. 🌐

✨ Contributions

• Unified Evaluation Protocol for World Generation. We introduce WorldAtlas, the largest unified benchmark to date for video, 3D, 4D, and interactive world generation, which formulates the task as next-state generation and maps a single world specification onto each system's native interface, comparing heterogeneous models on task semantics rather than on a shared representation.
• Large-scale Diagnostic Suite for World Generation. We curate 10,000 image and video evaluation instances spanning diverse scenes, viewpoint transformations, temporal dynamics, physical interactions, and action-conditioned transitions. The suite assesses whether a system can faithfully render a world, preserve its state across viewpoints and time, simulate plausible dynamics, and make targeted changes in response to interventions, enabling capability-specific diagnosis beyond aggregate perceptual scores.
• Unified Evaluation of World Generation Paradigms. We evaluate representative video, 3D, 4D, and interactive systems under a common capability framework, revealing a tradeoff between scalable visual synthesis and persistent world state. Visual fidelity does not establish a coherent, controllable world: state must persist across views and time, evolve plausibly, and change selectively in response to interventions.

🏆 Leaderboard

You're welcome to explore the latest results on our hosted leaderboard.

To submit your model, see the Submission guide (中文).

🚀 Quick Start

See the Getting Started guide for installation, data preparation, and runtime setup (中文).

Download the benchmark dataset from Hugging Face or ModelScope.

🎮 Evaluate Your Model

1. Complete the setup in Getting Started.
2. Generate videos and score their content-bound evaluation manifest:

# End-to-end (generate + evaluate)
worldatlas run-model \
  --config config/benchmark.yaml \
  --model-config config/models/lingbot_world/lingbot_world.yaml \
  --suites image_static

# Evaluate a completed WorldAtlas generation run
worldatlas evaluate \
  --config config/benchmark.yaml \
  --manifest artifacts/predictions/my-model/evaluation_manifest.jsonl \
  --predictions-root artifacts/predictions/my-model \
  --model-name my_custom_model

# Sharded multi-GPU evaluation
python scripts/run_eval_shards.py --help

Reports land in artifacts/benchmark/reports/<model>/ (per_sample_scores.jsonl, per_metric_summary.json, per_dimension_summary.json, leaderboard.csv).

Generation suites (what models produce and what the leaderboard scores)

Suite	Conditioning	Typical control
image_static	Single image	Camera path / pose, text prompt
image_dynamic	Image + mask / motion label	Dynamic prompt, object motion
memory_loop	Still image (interactive models)	Closed-loop 60s (*_memory.yaml) or 30s (*_memory_30s.yaml) rollout → long sequence

Reference video corpus

Collection	Role
video_static	Real-video reference corpus for jedi, fvmd, and trajan
video_dynamic	Real-video reference corpus for jedi, fvmd, and trajan

These clips live under data/video and are indexed as benchmark.reference_corpus. Image-conditioned predictions are scored against that corpus.

Dimensions (headline official metrics; full list on each dimension page)

Dimension	Representative metrics
Action control	prompt alignment, camera error, motion accuracy, keystroke action / trajectory
3D/4D consistency	MEt3R consistency, reprojection error, optical flow AEPE, geometric consistency, photometric consistency, FVMD, JEDi
Physics	motion magnitude, motion smoothness, physical FPS error, physical FPS percent error, intra-video FPS variation, TRAJAN
Quality	perceptual quality, style consistency, image quality, HPSv3 norm, background consistency, segment continuity
Long sequence / memory	long-sequence metrics, object permanence, scene memory F1, closed-loop memory family

experimental is a supplemental slice and does not enter formal headline aggregates.

🔌 Implement Your Model

WorldAtlas supports several control interfaces through the same YAML adapter contract:

Type	Input	Suites	Status
Image-conditioned	Reference image + prompt	image_static, image_dynamic	✅ Adapters in config/models/
Camera-conditioned	First frame + camera path / 6-DoF pose	image_static	✅ e.g. RealCam-I2V, LingBot-World
Action-conditioned	First frame + discrete action / keystroke	image_static	✅ e.g. MatrixGame, AlayaWorld
Interactive / memory	Closed-loop rollout from a still image	memory_loop	✅ *_memory.yaml (60s), *_memory_30s.yaml (30s)

Steps

1. Place upstream code under thirdparty/<Name>/ (or set repo_root).
2. Download weights to ckpt/<name>/.
3. Add config/models/<family>/<name>.yaml.
4. Implement or extend an adapter in worldatlas/models/adapters/ and register it in worldatlas/models/registry.py.

Minimal local-repo template:

model:
  name: my-world-model
  family: my_world_model
  repo_root: thirdparty/MyModel
  checkpoint_dir: ckpt/my-model
  python_bin: python          # or a dedicated env interpreter
  entrypoint: infer.py
  output_ext: .mp4
  supported_suites: [image_static, image_dynamic]
  prompt_mode: prompt_target
  reference_mode: first_frame
  generation: { num_frames: 81, fps: 24 }

HTTP API template: config/models/video_api/template_video_api.yaml.example. Adapter contracts: Getting Started.

📝 Citation

If you use WorldAtlas in your research, please cite:

@article{worldatlas2026,
  title   = {WorldAtlas: A Unified Large-Scale Benchmark for World Generation toward World Intelligence},
  author  = {OpenEnvision},
  year    = {2026},
  url     = {https://github.com/OpenEnvision/WorldAtlas}
}

🙏 Acknowledgement

This project builds upon the following excellent works:

• MEt3R — multi-view consistency
• Depth Anything 3 — geometric / photometric reconstruction
• FVMD — Fréchet video motion distance
• JEDi — video distributional similarity
• TRAJAN / BootsTAPIR — trajectory reconstruction
• MegaSAM / DROID-SLAM — camera and reprojection
• HPSv3 — human preference score
• SAM 2 / GroundingDINO — motion-region supervision
• RAFT / SEA-RAFT — optical flow
• TransNetV2 — scene-boundary detection
• AMT / VFIMamba — temporal interpolation
• … and the upstream world-model repositories listed under thirdparty/

📧 Contact

Open an Issue or Pull Request, or email jiadonghou3@gmail.com.

📄 License

• Code in this repository: MIT License
• Model weights: subject to each upstream project's license
• Benchmark media: subject to the dataset license under which you obtained the assets

---

WorldAtlas · Homepage · GitHub · Documentation