TencentARC / Multics

# *Multics*: Generative Multi-Player World Model with Hierarchical Communication
Kaiyi Huang 1 ,
Wangbo Yu 2† ,
Yukun Huang 2 ,
Jiwen Yu 1 ,
Kunhao Liu 2 ,
Donglin Yang 1 ,
Wenbo Hu 2 ,
Ying Shan 2 ,
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
https://github.com/user-attachments/assets/ba220953-b57f-47a5-bbb8-9b440b55ed11
## 💡 Overview
**Multics** organizes communication across player, group, and global levels, using dynamic grouping based on recent camera trajectories. 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:
 [code] 
### Model Weights
Download the complete inference bundle from TencentARC/Multics, including the 4-player checkpoint, VAE, text encoder, and tokenizer:
 [code] 
### Inference
#### 1. 4-player generation
Generate example `4_01` from its first frames, captions, and tick logs:
 [code] 
Outputs are saved to `outputs/inference/4players/`. 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:
 [code] 
| 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 and download the original Wan2.2 backbone for initialization:
 [code] 
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 maki