---
tags:
  - world-model
  - video-generation
  - camera-control
  - long-horizon-video
  - wan2.2
  - ltx-2.5
  - minimax-h3
library_name: pytorch
---

<div align="center">

# SolarWM

### Open Data and Scalable Training for Long-Horizon Video World Models

[Project Page](https://junchao-cs.github.io/SolarWM-Web/) · [GitHub](https://github.com/Junchao-cs/SolarWM/tree/main) · [Model Collection](https://huggingface.co/collections/junchaoh-cs/solarwm-6a992c9d76d85c29637c9111) · [Dataset](https://huggingface.co/datasets/junchaoh-cs/SolarWM-Data) · [Dataset Access Form](https://docs.google.com/forms/d/e/1FAIpQLSfS-SLOiSRVDWwZ2kPl9ywN27aB6QplN0jpdKaBu-gG8aNsvQ/viewform)

</div>

SolarWM is an open foundation for building interactive video world models, covering data preparation, scalable training, and long-horizon inference. The framework supports multiple backbone families while preserving each backbone's native representation and training objective.

SolarWM uses a shared three-stage recipe:

1. **Stage0.5 — Bidirectional Flow Matching:** adapts the backbone to camera-conditioned full-clip generation.
2. **Stage1 — Teacher Forcing + AnyFlow:** trains autoregressive denoising and finite-step flow maps using clean history.
3. **Stage2 — Self-Gradient Forcing (SGF):** distills a few-step causal student on its own autoregressive rollouts.

Models trained only on short clips can be rolled out autoregressively for long-horizon interactive video generation.

## Family-specific repositories

For inference, download only the backbone family you need. Each repository contains its base runtime assets and all released SolarWM checkpoints for that family.

| Repository | Backbone | Released stages |
|---|---|---|
| [`SolarWM-Wan2.2-5B`](https://huggingface.co/junchaoh-cs/SolarWM-Wan2.2-5B) | Wan2.2 TI2V-5B | Base, Stage0.5, Stage1, Stage2 |
| [`SolarWM-Wan2.2-14B`](https://huggingface.co/junchaoh-cs/SolarWM-Wan2.2-14B) | Wan2.2 I2V-A14B | Base, Stage0.5 |
| [`SolarWM-LTX-22B`](https://huggingface.co/junchaoh-cs/SolarWM-LTX-22B) | LTX-2.5 22B | Base, Stage0.5 |
| [`SolarWM-H3-33B`](https://huggingface.co/junchaoh-cs/SolarWM-H3-33B) | MiniMax-H3 33B | Base, Stage0.5 |

Browse the complete [SolarWM collection](https://huggingface.co/collections/junchaoh-cs/solarwm-6a992c9d76d85c29637c9111) for models, data, and the paper. This original repository remains available as the complete all-in-one bundle for backward compatibility.

## Complete bundle contents

| Directory | Backbone | Stage | Frames | Release contents |
|---|---|---:|---:|---|
| `SolarWM-5B-base` | Wan2.2 TI2V-5B | Base | — | Base model and runtime assets |
| `SolarWM-5B-bid-stage0p5-81f` | Wan2.2 TI2V-5B | Stage0.5 | 81 | Bidirectional camera-conditioned checkpoint |
| `SolarWM-5B-bid-stage0p5-153f` | Wan2.2 TI2V-5B | Stage0.5 | 153 | Longer-context bidirectional checkpoint |
| `SolarWM-5B-tf-stage1-81f` | Wan2.2 TI2V-5B | Stage1 | 81 | Teacher-forced AnyFlow checkpoint |
| `SolarWM-5B-sgf-stage2-81f` | Wan2.2 TI2V-5B | Stage2 | 81 | Few-step causal SGF checkpoint |
| `SolarWM-14B-base-high` | Wan2.2 I2V-A14B | Base | — | High-noise expert and runtime assets |
| `SolarWM-14B-bid-stage0p5-81f` | Wan2.2 I2V-A14B | Stage0.5 | 81 | Bidirectional camera-conditioned checkpoint |
| `SolarWM-ltx-22B-base` | LTX-2.5 22B | Base | — | Base model and runtime assets |
| `SolarWM-ltx-22B-bid-stage0p5-153f` | LTX-2.5 22B | Stage0.5 | 153 | EMA LoRA adapter |
| `SolarWM-h3-33B-base` | MiniMax-H3 33B | Base | — | Base model and runtime assets |
| `SolarWM-h3-33B-bid-stage0p5-158f` | MiniMax-H3 33B | Stage0.5 | 158 | Live LoRA adapter |

## Download

Download one family without fetching the other model sizes:

```bash
hf download junchaoh-cs/SolarWM-Wan2.2-5B --local-dir ./SolarWM-ckpt
```

The complete all-in-one release remains available from this repository:

```bash
hf download junchaoh-cs/SolarWM --local-dir ./SolarWM-ckpt
```

Download a single checkpoint:

```bash
hf download junchaoh-cs/SolarWM \
  --include "SolarWM-5B-sgf-stage2-81f/*" \
  --local-dir ./SolarWM-ckpt
```

Keep the directory layout unchanged. SolarWM example configurations resolve the base and staged model paths from these release directories. See the [GitHub repository](https://github.com/Junchao-cs/SolarWM/tree/main) for installation, backend environments, training configurations, and inference commands.

## Related resources

- [SolarWM project page](https://junchao-cs.github.io/SolarWM-Web/)
- [SolarWM source code](https://github.com/Junchao-cs/SolarWM/tree/main)
- [SolarWM dataset](https://huggingface.co/datasets/junchaoh-cs/SolarWM-Data)
- [Wan2.2](https://github.com/Wan-Video/Wan2.2)
- [LTX-2](https://github.com/Lightricks/LTX-2)
- [MiniMax-H3](https://github.com/MiniMax-AI/MiniMax-H3)

## License

SolarWM code is released under Apache-2.0. Model weights and bundled backbone assets remain subject to the license and notices included in each release directory. In particular, LTX-2.5 packages are governed by the LTX-2.x Community License, and MiniMax-H3 packages are governed by the MiniMax-H3 Community License. These packages are not relicensed under Apache-2.0. Review the applicable package license before use or redistribution.

## Citation

**If you use SolarWM-Data, the data engine, or the released models in your research,
please cite our paper.**

Paper: https://arxiv.org/abs/2609.02886

```bibtex
@misc{huang2026solarwmopendatascalable,
      title={SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models}, 
      author={Junchao Huang and Guian Fang and Shengju Qian and Xianghao Kong and Zhuoran Zhao and Wei Huang and Yihua Du and Zixin Zhang and Justin Cui and Yuchao Gu and Yukang Chen and Xinting Hu and Tianyu He and Shaoshuai Shi and Zhuotao Tian and Xin Wang and Mike Zheng Shou and Li Jiang},
      year={2026},
      eprint={2609.02886},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.02886}, 
}
```