Blame for README.md

---
license: apache-2.0
language:
- en
base_model: MiniMaxAI/MiniMax-H3
base_model_relation: adapter
datasets:
- iapple1/CoCode-World-Data
pipeline_tag: image-to-video
tags:
- lora
- minimax-h3
- video-generation
- multi-view
- world-model
---

# CoCode-World

**Shared 3D Proxy Representation for Multi-Player World Models**

[Project Page](https://iapple233.github.io/CCW/) · [Code](https://github.com/IApple233/CoCode-World) · [Dataset](https://huggingface.co/datasets/iapple1/CoCode-World-Data) · [Examples](https://github.com/IApple233/CoCode-World/tree/main/examples/evaluation)

## Model overview

CoCode-World generates synchronized video from two player perspectives using a shared 3D proxy representation. A reference image defines the visual appearance, a proxy video provides scene structure and motion, and a text prompt describes the scene.

This repository provides the official CoCode-World LoRA checkpoint for **MiniMax-H3 Ref2VA**. The recommended EMA weights are used in the released evaluation examples.

| Property | Description |
|---|---|
| Base model | MiniMax-H3 Ref2VA |
| Adapter | Rank-64 LoRA |
| Checkpoint | Step 4000 |
| Inputs | Reference image, synchronized proxy video, and text prompt |
| Output | Two player views arranged side by side |
| Resolution | 2560 × 720 at 24 FPS |

## Download

```bash
pip install -U huggingface_hub
hf download iapple1/CoCode-World-Model --local-dir checkpoints/step-004000
```

The repository includes `ema.safetensors` for inference, `lora.safetensors` for the task adapter, and training state and configuration files. Use the checkpoint loader in the official code repository. Download the MiniMax-H3 base model separately.

## Usage

Follow the [installation guide](https://github.com/IApple233/CoCode-World/blob/main/docs/SETUP.md), then download the required models from the code repository:

```bash
python scripts/download_models.py base
python scripts/download_models.py cocode
```

The [evaluation guide](https://github.com/IApple233/CoCode-World/blob/main/docs/EVALUATION.md) explains input preparation and generation. Eight ready-to-use examples include reference images, prompts, proxy videos, and generated results.

## Training

Training starts from the official [MiniMax-H3 weights](https://huggingface.co/MiniMaxAI/MiniMax-H3) and [DiffSynth-Studio's Ref2VA DeCFG LoRA](https://www.modelscope.cn/models/DiffSynth-Studio/MiniMax-H3-TrainingAdapter). The DeCFG adapter is used as a frozen training preset, followed by training a separate CoCode-World task LoRA. Inference uses the original MiniMax-H3 base and the CoCode-World EMA LoRA.

Training code and instructions are available in the [training guide](https://github.com/IApple233/CoCode-World/blob/main/docs/TRAINING.md). The public dataset is a selected subset of the original training collection.

## Intended use and limitations

The model supports research on multi-player world models, synchronized multi-view generation, and proxy-conditioned video synthesis. The first-frame reference guides appearance. Generated videos may contain visual artifacts or temporal inconsistencies, and the selected examples do not constitute a quantitative benchmark.

## License and acknowledgments

The task adapter is released under Apache 2.0. The MiniMax-H3 base model is governed by the [MiniMax-H3 Community License Agreement](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE). The DeCFG training adapter retains its upstream terms.

CoCode-World builds on [MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3), [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio), [Diffusers](https://github.com/huggingface/diffusers), and [Code World Model](https://github.com/buaacyw/code-world-model).