Enderfga/SolarRenderer

SolarRenderer: Few-Step Generative Reference Rendering

🌐 Project Page • 📄 Paper • 🤗 Weights • 🤗 Data • 💻 Data code

SolarRenderer turns a structure video from a game engine or an agent (grey box, semantic map, skeletons, depth, body mesh, or a sketch it never saw in training) into a finished video, with an optional first frame and a prompt deciding the look. It runs on MiniMax-H3 with two LoRAs and renders a clip in four steps. A block-causal pair of the same LoRAs streams the clip block by block. SolarRenderer is built on SolarWM, the open framework for training and inference of video world models (paper).

News

• [2026-10] Code, weights, the shareable training data and the benchmark released.

Highlights

• 🧱 Any structure, one model: no extra branch and no condition-type label; the prompt names the structure.
• ⚡ Four steps: structure-guided post-training, 12.3× faster sampling than 50 steps.
• 🎨 Any look: the same structure renders as photoreal, anime, pixel art or ink, set by the first frame and prompt.
• 🌊 Streaming: a block-causal version of both LoRAs renders the video block by block, each block shown as soon as

• 📊 Open benchmark: 208 hard cases and 53 held-out clips, with scoring code, our renders and every judge verdict.

Gallery

Structure video (left) and SolarRenderer (right).

| Scene grey box | Semantic segmentation | |:---:|:---:| | | | | Object masks + skeletons | SMPL body mesh | | | | | Near-field shaded geometry | Pencil storyboard (unseen) | | | | | Point cloud (unseen) | Child's crayon drawing (unseen) | | | | | Voxels (unseen) | Wireframe (unseen) | | | |

More structure types, styles and comparisons are on the project page.

Quick Start

git clone https://github.com/Enderfga/SolarRenderer && cd SolarRenderer
pip install -e . && pip install flash-attn==2.7.4.post1 --no-build-isolation

python scripts/prepare_h3.py --dir ./MiniMax-H3
hf download Enderfga/SolarRenderer solarrenderer_control_lora.safetensors solarrenderer_4step_lora.safetensors \
    --local-dir ./weights

python scripts/render.py --h3 ./MiniMax-H3 \
    --control weights/solarrenderer_control_lora.safetensors \
    --fewstep weights/solarrenderer_4step_lora.safetensors \
    --structure examples/sem-highway_synthwave/structure.mp4 \
    --first examples/sem-highway_synthwave/first.png \
    --prompt examples/sem-highway_synthwave/prompt.txt --out render.mp4

The example is a synthwave highway run (semantic map in, neon render out). Add --causal with the two causal LoRAs to stream the clip block by block (details).

• ✍️ Prompts and first frames: give skills/solarrenderer-caption to your coding

• 🎬 Inference details: inputs, sizes and timing (guide).
• 🏋️ Training on one 8-GPU node (guide).
• 📊 Benchmark scoring (guide).

Citation

If you use SolarRenderer in your research, please cite SolarRenderer and SolarWM, the framework on which it is built.

@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://enderfga.cn/data/solarrenderer.pdf}
}

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

@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},
}

License

Code: Apache 2.0. Weights: MiniMax H3 Community License, including its territorial restrictions (EU, UK, Republic of Korea and US excluded) and Acceptable Use Policy; the four-step LoRA is fine-tuned from DMAD under the same license.