Infinite-World

Scaling Interactive World Models to 1000-Frame Horizons via Pose-Free Hierarchical Memory

Ruiqi Wu^1,2,3*, Xuanhua He^4,2*, Meng Cheng^2*, Tianyu Yang², Yong Zhang^2‡, Chunle Guo^1,3†, Chongyi Li^1,3, Ming-Ming Cheng^1,3

¹Nankai University   ²Meituan   ³NKIARI   ⁴HKUST

^*Equal Contribution   ^†Corresponding Author   ^‡Project Leader

---

Highlights

Infinite-World is a robust interactive world model with:

• Real-World Training — Trained on real-world videos without requiring perfect pose annotations or synthetic data
• 1000+ Frame Memory — Maintains coherent visual memory over 1000+ frames via Hierarchical Pose-free Memory Compressor (HPMC)
• Robust Action Control — Uncertainty-aware action labeling ensures accurate action-response learning from noisy trajectories

Installation

Environment: Python 3.10, CUDA 12.4 recommended.

1. Create conda environment

conda create -n infworld python=3.10
conda activate infworld

2. Install PyTorch with CUDA 12.4

Install from the official PyTorch index (no local whl):

pip install torch==2.6.0 torchvision==0.21.0 --index-url https://download.pytorch.org/whl/cu124

3. Install Python dependencies

pip install -r requirements.txt

---

Checkpoint Configuration

All model paths are configured in configs/infworld_config.yaml. Paths are relative to the project root unless absolute.

Download checkpoints

Download from Wan-AI/Wan2.1-T2V-1.3B and place files under checkpoints/:

• DiT checkpoint: Can be downloaded from TBD.

---

Upload to Hugging Face (including checkpoints)

To upload this repo to Hugging Face Hub (code + checkpoints/):

1. Login

   pip install huggingface_hub
   huggingface-cli login
   

   Use a token from https://huggingface.co/settings/tokens (need write permission).

2. Upload From the project root (infinite-world/):

   python scripts/upload_to_hf.py YOUR_USERNAME/infinite-world
   

   Or set the repo and run:

   export HF_REPO_ID=YOUR_USERNAME/infinite-world
   python scripts/upload_to_hf.py
   

   The script uploads the whole directory (including checkpoints/) and skips __pycache__, outputs, .git, etc. Large checkpoint files are uploaded via the Hub API; the first run may take a while depending on size and network.

3. Create repo manually (optional)
   You can create the model repo first at https://huggingface.co/new (type: Model), then run the script with that repo_id.

---

Results

Quantitative Comparison

Citation

If you find this work useful, please consider citing:

@article{wu2026infiniteworld,
  title={Infinite-World: Scaling Interactive World Models to 1000-Frame Horizons via Pose-Free Hierarchical Memory},
  author={Wu, Ruiqi and He, Xuanhua e.a.},
  journal={arXiv preprint arXiv:2602.02393},
  year={2026}
}

License

This project is released under the MIT License.