Multics: Generative Multi-Player World Model with Hierarchical Communication

Kaiyi Huang¹, Wangbo Yu^2†, Yukun Huang², Jiwen Yu¹, Kunhao Liu²,
Donglin Yang¹, Wenbo Hu², Ying Shan², Xihui Liu^1†

¹ The University of Hong Kong   ² ARC Lab, Tencent IEG

^† Corresponding authors.

Multics enables scalable multi-player video generation through hierarchical communication.

🎬 Demo video

Multics.mp4

💡 Overview

Multics organizes communication across player, group, and global levels, using dynamic grouping based on recent player positions and camera poses. This hierarchy improves cross-player consistency, action and camera controllability, and robustness to interference while reducing attention computation. Shared group-token parameters allow a model trained with 4 players to generalize to 6, 8, and 10 players without retraining.

📰 News

• 🔥 [2026/10] Paper, training and inference code released.

🚀 Get started

Installation

Requires Linux, Python 3.12, an NVIDIA GPU, the CUDA 12.8 toolkit, and a C++ compiler.

From the repository root, install dependencies with uv:

uv sync --locked
source .venv/bin/activate

Model Weights

Download the complete inference bundle from TencentARC/Multics, including the 4-player checkpoint, VAE, text encoder, and tokenizer:

hf download TencentARC/Multics --local-dir models

Inference

1. 4-player generation

Generate example 4_01 from its first frames, captions, and tick logs:

python inference.py --config configs/inference_demo.yaml \
  --checkpoint models/multics/4players/checkpoints/model.ckpt \
  --hparams models/multics/4players/hparams.yaml \
  --case 1

Outputs are saved to outputs/inference/4players/. Videos show only the generated player views. Use --cases 1 2 3 4 5 to run all five examples.

2. Scalability

The same 4-player checkpoint supports 6-, 8-, and 10-player generation without retraining:

python inference.py --config configs/inference_demo.yaml \
  --checkpoint models/multics/4players/checkpoints/model.ckpt \
  --hparams models/multics/4players/hparams.yaml \
  --num_players 8 --cases 1 2

Players	Examples	Arguments
6	6_01–6_03	--num_players 6 --cases 1 2 3
8	8_01–8_02	--num_players 8 --cases 1 2
10	10_01–10_02	--num_players 10 --cases 1 2

Outputs are saved to outputs/inference/4infer6/, 4infer8/, or 4infer10/. See example inputs for the data format and previews.

GPU memory

Single-GPU inference with configs/inference_demo.yaml: BF16, 640 × 352 per player, 165 frames, and 50 sampling steps (PyTorch 2.8.0, CUDA 12.8).

GPU memory grows approximately linearly with player count over the tested 4–10 player range:

Players	Peak allocated (GiB)	Peak reserved (GiB)
4	59.9	74.9
6	75.3	97.4
8	90.6	119.0
10	106.1	139.2

Measured PyTorch peaks for reference. Allocated measures tensor memory; reserved also includes allocator cache.

Training

To train the 4-player model, prepare your dataset in the precomputed input format. In addition to the model bundle above, download the original Wan2.2 backbone for initialization:

hf download Wan-AI/Wan2.2-TI2V-5B \
  --include 'diffusion_pytorch_model*.safetensors' \
  --local-dir models/Wan-AI/Wan2.2-TI2V-5B

uv sync --locked --extra distributed

NPROC_PER_NODE=8 bash scripts/train.sh \
  --config configs/train.yaml \
  --dataset_path /path/to/data \
  --metadata_file_name train.csv

Set NPROC_PER_NODE to your GPU count. Training defaults are provided in configs/train.yaml.

🤗 Acknowledgements

This project is built on Wan and DiffSynth-Studio. We thank the authors for their open-source contributions.

We also thank the creators of OpenCS2 for making their dataset publicly available.

The code is released under the Apache-2.0 license. Model weights and example content retain their upstream terms.

📝 Citation

If you find this work useful, please cite:

@misc{huang2026multics,
  title  = {Multics: Generative Multi-Player World Model with Hierarchical Communication},
  author = {Huang, Kaiyi and Yu, Wangbo and Huang, Yukun and Yu, Jiwen and Liu, Kunhao and Yang, Donglin and Hu, Wenbo and Shan, Ying and Liu, Xihui},
  year   = {2026},
  eprint = {2610.17201},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/2610.17201}
}