Enderfga/SolarRenderer

SolarRenderer is a tool that turns a rough "structure video" — like a grey-box mockup, a semantic map, skeletons, depth data, a body mesh, or even a sketch it has never seen before — into a finished, polished video. You can optionally give it a first frame and a text prompt to control the final look (photoreal, anime, pixel art, ink, etc.).

How it works: It builds on a video model called MiniMax-H3 and adds two small adapter modules (LoRAs). One adapter helps it understand the structure input; the other lets it produce a full clip in just four generation steps, which the README says is 12.3× faster than the usual 50-step approach. A special "block-causal" version of the same adapters streams the video block by block, showing each chunk as soon as it's ready.

Key highlights:

• One model, any structure — no separate branches or labels for different structure types; the prompt just names what the structure is.
• Fast — four steps instead of dozens.
• Flexible style — the same structure can be rendered in many visual styles, set by the first frame and prompt.
• Streaming — block-by-block output for long clips.
• Open benchmark — 208 hard test cases and 53 held-out clips, with scoring code and all judge verdicts.

How to run it: You can try it in a browser via the Hugging Face online demo (no install), use the reference Python script on one GPU, run it as ComfyUI node workflows (a low-memory version fits a 16 GB GPU like an RTX 4090), or deploy a streaming service on the Reactor platform (a community contribution, validated on four B200 GPUs).

Setup: Clone the repo, install dependencies, download the base model and the two LoRA weight files, then run a render script with your structure video, optional first frame, and prompt. The included example is a synthwave highway scene: semantic map in, neon render out.

License: The code is Apache 2.0. The model weights follow the MiniMax H3 Community License, which includes territorial restrictions (EU, UK, Republic of Korea, and US excluded) and an Acceptable Use Policy. The four-step LoRA is fine-tuned from another model (DMAD) under the same license.

Credits: SolarRenderer is built on the SolarWM framework for video world models, and acknowledges MiniMax-H3 and DMAD as foundations.