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‡, Zhuoliang Kang², Xunliang Cai², Xiaoming Wei², 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

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 checkpoints from https://huggingface.co/MeiGen-AI/Infinite-World and place files under checkpoints/:

File / directory	Config key	Description
models/Wan2.1_VAE.pth	vae_cfg.vae_pth	VAE weights
models/models_t5_umt5-xxl-enc-bf16.pth	text_encoder_cfg.checkpoint_path	T5 text encoder
models/google/umt5-xxl (folder)	text_encoder_cfg.tokenizer_path	T5 tokenizer
infinite_world_model.ckpt	checkpoint_path	DiT model weights

---

Results

Quantitative Comparison

Model	Mot. Smo.↑	Dyn. Deg.↑	Aes. Qual.↑	Img. Qual.↑	Avg. Score↑	Memory↓	Fidelity↓	Action↓	ELO Rating↑
Hunyuan-GameCraft	0.9855	0.9896	0.5380	0.6010	0.7785	2.67	2.49	2.56	1311
Matrix-Game 2.0	0.9788	1.0000	0.5267	0.7215	0.8068	2.98	2.91	1.78	1432
Yume 1.5	0.9861	0.9896	0.5840	0.6969	0.8141	2.43	1.91	2.47	1495
HY-World-1.5	0.9905	1.0000	0.5280	0.6611	0.7949	2.59	2.78	1.50	1542
Infinite-World	0.9876	1.0000	0.5440	0.7159	0.8119	1.92	1.67	1.54	1719

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 and Cheng, Meng and Yang, Tianyu and Zhang, Yong and Kang, Zhuoliang and Cai, Xunliang and Wei, Xiaoming and Guo, Chunle and Li, Chongyi and Cheng, Ming-Ming},
  journal={arXiv preprint arXiv:2602.02393},
  year={2026}
}

License

This project is released under the MIT License.