SolarWM

Solar is an open framework for interactive world models, spanning open data, scalable training, long-horizon video generation, and real-time deployment. This repository contains code for SolarWM and SolarRenderer. See Install for setup, Data for datasets, and Documentation for training, inference, and deployment guides.

🎬 Demo videos

📰 News

• November 5, 2026: We release SolarWM-H3 Memory (Blog · Weights · Usage), improving long-term scene consistency and preserving visual details across revisits.

• October 22, 2026: We release SolarWM-H3 Realtime (Blog), enabling low-latency interactive video generation from an image and a prompt, with real-time camera control (Code).

• October 16, 2026: We release H3 Streaming VAE Decoder (Blog), enabling low-latency streaming video decoding that produces frames as latents arrive. It supports both the official MiniMax-H3 VAE and NVIDIA’s distilled StreamRender-H3 decoder (Code).

• October 16, 2026: We release SolarRenderer (Project · Paper), bringing four-step, structure-conditioned video rendering with bidirectional and block-causal training and inference on MiniMax-H3. Follow the SolarRenderer tutorial to render from a structure video, an optional first frame, and a prompt.

• September 13, 2026: We release SolarWM-H3 (Guide), the first MiniMax-H3-based world model for streaming, few-step interaction. Training and inference code, model weights, and pre-encoded training data are available for all three stages.

• September 3, 2026: We open-source SolarWM (Project · Paper), releasing 1.43M video clips, the data pipeline, and training and inference code with model weights for four video backbones spanning 5B–33B.

💡 Introduction

We introduce Solar, an open framework for interactive world modeling and generative video rendering:

• SolarWM: Open data and scalable training for long-horizon, camera-controlled video world models, backed by 1.43M clips (25 TB) and an open data-processing pipeline. Supports four 5B–33B models across Wan2.2, LTX-2.5, and MiniMax-H3, with a shared training workflow and real-time interaction with SolarWM-H3.
• SolarRenderer: Four-step, structure-conditioned video rendering on MiniMax-H3, supporting bidirectional and block-causal generation from a structure video, an optional first frame, and a prompt. Trained on 136K clips from eight sources, with released training data and generalization benchmarks.

🛠️ Install

Set up the environment for your model or deployment using the guides below. Each guide specifies the required Python, PyTorch, and CUDA versions.

Use case	Installation guide
SolarWM · Wan2.2, LTX-2.5, or MiniMax-H3	Runtime environments
SolarRenderer · training and inference	SolarRenderer setup
SolarWM-H3 · realtime interaction	Realtime deployment

Activate your environment, then install Solar from the repository root:

python -m pip install -e .
solar --help

🗂️ Data

SolarWM-Data

1.43M clips · 25 TB · 14 sources, with camera annotations, captions, and quality metadata. The public release includes annotations, reconstruction tools, and training recipes. Prepared raw video data can be requested through the access form; preencoded latents are distributed in separate releases.

• Download: Hugging Face · ModelScope · Download & access
• Data preparation: Dataset details · Data pipeline
• Training & evaluation: Preencoded latents · Test set

SolarRenderer-Data

The training corpus contains 136K clips from eight sources. The public release provides 57,738 training clips paired with structure videos, first frames, and prompts, plus a 208-case generalization benchmark and 53 held-out clips with ground truth.

• Download: Hugging Face
• Data preparation: Training data preparation · Data pipeline
• Evaluation: Benchmark evaluation

📚 Documentation

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

• Getting started: Quickstart · All guides
• Models: Wan2.2-5B · Wan2.2-14B · LTX-2.5 · MiniMax-H3
• Deployment: H3 realtime interaction

SolarRenderer: Few-Step Generative Reference Rendering

• Getting started: Inference · Training · All guides
• Deployment: ComfyUI · Reactor
• Tools: Prompts and reference frames · Evaluation

Contributing

🤗 Acknowledgements

We thank the teams behind Wan2.2, LTX-2.5, and MiniMax-H3 for their open models and code. Additional dataset and code attribution is recorded in NOTICE.

License and attribution

The Solar codebase is licensed under Apache-2.0. See LICENSE and NOTICE. Model weights and backbone packages may carry their own licenses; review the license and model card in the corresponding release package before use or redistribution. In particular, LTX-2.5 derivatives are subject to the LTX-2.x Community License, and the MiniMax H3 Community License contains territory restrictions. Those packages are not relicensed under the Solar code license.

📖 BibTeX

If you use SolarWM, please cite the paper below. Citations for SolarRenderer and the technical blogs are available in the expandable sections.

@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},
}

@article{fang2026solarrenderer,
  title  = {SolarRenderer: Few-Step Generative Reference Rendering},
  author = {Fang, Guian and Qian, Shengju and Lu, Dongyue and
            Huang, Junchao and Jiang, Lutao and Wang, Xin and
            Song, Yiren and Jiang, Li and Shou, Mike Zheng},
  year   = {2026},
  url    = {https://arxiv.org/abs/2610.28630}
}

@misc{du2026acceleratingsolarwmh3,
  title  = {SolarWM H3 Realtime Interaction},
  author = {Du, Yihua and Huang, Junchao and Chen, Bohong and
            Liu, Yaoyang and Zhang, Zixin and Kong, Xianghao and
            Yan, Dongyu and Zhao, Zhuoran and Fang, Guian and
            Chen, Yajing and Qian, Shengju},
  year   = {2026},
  month  = {oct},
  note   = {SolarX-Team technical blog},
  url    = {https://solarx-team.github.io/SolarWM-H3-Realtime/}
}

@misc{chen2026h3streaming,
  title        = {H3 Streaming VAE Decoder: Training-Free Streaming without Future Latents},
  author       = {Chen, Bohong and Chen, Yajing and Du, Yihua and Huang, Junchao and Fang, Guian and Qian, Shengju},
  year         = {2026},
  month        = {oct},
  howpublished = {\url{https://solarx-team.github.io/H3-Streaming-VAE-Decoder/}},
  note         = {SolarX-Team blog post}
}

@misc{zhang2026solarwmmemory,
  title  = {{SolarWM H3 + Memory}: Training Memory For Video World Model Is Easier Than You Think},
  author = {Zhang, Zixin and Huang, Junchao and Du, Yihua and
            Lu, Dongyue and Zhao, Zhuoran and Kong, Xianghao and
            Chen, Bohong and Fang, Guian and Yan, Dongyu and Liu, Yaoyang and Qian, Shengju},
  year   = {2026},
  month  = nov,
  note   = {SolarX-Team technical blog},
  url    = {https://solarx-team.github.io/SolarWM-H3-Memory/}
}