SolarRenderer adapts MiniMax-H3 for four-step structure-conditioned video
Original title: SolarRenderer: Few-Step Generative Reference Rendering
Two LoRAs keep a frozen 33B backbone and render about 12.3 times faster than 50-step sampling.
SolarRenderer is a structure-conditioned video renderer that adapts a frozen MiniMax-H3 33B backbone with two lightweight LoRAs. Structural videos enter the backbone's native reference pathway in plain text, and a control LoRA follows arbitrary structural representations. Starting from a pretrained four-step LoRA, structure-guided distribution matching renders in four steps; a block-causal variant streams 768p video in real time. Training uses 136K frame-aligned clips, and the authors report generalization to 11 unseen representations.
- Four-step rendering is 12.3 times faster than 50-step sampling.
- The corpus draws on game engines, simulators, and YouTube videos.
- A benchmark covers 16 structural representations.
- Unseen inputs include pose skeletons, drawings, and point clouds.
- No type identifier is added to the structural reference videos.
From the original article
1 min read at huggingface.co
Game engines and coding agents can specify the layout, motion, and camera of a scene, yet turning such coarse structure into realistic video requires a renderer that follows the structure and completes the appearance it leaves open.
By Guian Fang, Shengju Qian, Dongyue Lu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.