README.md

---
license: apache-2.0
tags:
  - stereo-video
  - world-model
  - diffusion
---

# StereoWorld

Official model weights for **[StereoWorld: Camera-Guided Stereo Video Generation](https://arxiv.org/abs/2603.17375)**.

[![GitHub](https://img.shields.io/badge/GitHub-Code-black?logo=github)](https://github.com/VAST-AI-Research/StereoWorld)
[![Paper](https://img.shields.io/badge/arXiv-PDF-b31b1b)](https://arxiv.org/abs/2603.17375)

## Models

This repository contains two models:

| Directory | Model | Description |
| --- | --- | --- |
| `StereoWorldModel/` | Fixed-Baseline Stereo | Generates side-by-side stereo video with a configurable but fixed stereo baseline. Use `--use_raymap` during inference. |
| `StereoWorldFlexModel/` | Flexible Stereo | Independent left/right camera control with multiple right-camera modes (converging, horizontal offset, depth offset, height offset). |

### Fixed-Baseline Stereo (`StereoWorldModel/`)

A binocular teacher model that produces left-right stereo pairs with a fixed baseline between cameras. Best for standard stereo video generation where consistent disparity is desired.

### Flexible Stereo (`StereoWorldFlexModel/`)

A more general multi-view world model that allows independent camera trajectories for left and right views. Supports four right-camera modes for creative stereo effects.

## Download

### Download everything

```bash
huggingface-cli download Yang-Tian/StereoWorld --local-dir weights
```

### Download only Fixed-Baseline Stereo

```bash
huggingface-cli download Yang-Tian/StereoWorld \
  --include "StereoWorldModel/*" \
  --local-dir weights
```

### Download only Flexible Stereo

```bash
huggingface-cli download Yang-Tian/StereoWorld \
  --include "StereoWorldFlexModel/*" \
  --local-dir weights
```

## Directory Structure

```text
StereoWorld/
├── StereoWorldModel/
│   ├── transformer/
│   ├── vae/
│   ├── tokenizer/
│   ├── text_encoder/
│   └── scheduler/
└── StereoWorldFlexModel/
    ├── transformer/
    ├── vae/
    ├── tokenizer/
    ├── text_encoder/
    └── scheduler/
```

## Usage

See the [GitHub repository](https://github.com/VAST-AI-Research/StereoWorld) for installation and inference instructions.

```bash
# Fixed-Baseline Stereo
python3 inference.py \
  --pipeline_dir weights/StereoWorldModel \
  --use_raymap \
  --eval_json ExpData/demo_custom_eval.json

# Flexible Stereo
python3 inference_flex.py \
  --pipeline_dir weights/StereoWorldFlexModel \
  --eval_json ExpData/flex_demo_custom_eval.json
```

## Citation

```bibtex
@article{sun2026stereo,
  title={Stereo World Model: Camera-Guided Stereo Video Generation},
  author={Sun Yang-Tian and Huang Zehuan and Niu Yifan and Ma Lin and Cao Yan-Pei and Ma Yuewen and Qi Xiaojuan},
  journal={arXiv preprint arXiv:2603.17375},
  year={2026}
}
```