SolarRenderer Enables Reference Rendering Through a Few-Step Generative Process
SolarRenderer converts untextured structure passes from game engines, Blender, three.js, or other sources into finished, styled real-time video guided by a single reference first frame. Built on MiniMax-H3 with control and few-step distilled low-rank adaptation models, it preserves object positions, motion, and timing frame by frame without per-game training. On a held-out structure benchmark it ranks first among nine methods with an Elo score of 1313. Its block-causal student renders 4 steps per chunk, producing a 17-frame chunk every 0.70 seconds at 24 frames per second, with about 1.1 seconds of input delay and endless streaming enabled by a key-value cache. Paper, project page, code, and weights are available.
Key Takeaways
- SolarRenderer converts untextured engine structure passes into styled real-time video while preserving geometry, motion, and timing frame by frame.
- It builds on MiniMax-H3 with a control LoRA and a few-step distilled LoRA, using only a reference first frame and requiring no per-game training.
- It supports diverse inputs such as Blender, three.js, game engines, and phone-captured screens, and can stream endlessly with a KV cache.
- It ranks first on the structure benchmark with Elo 1313 and achieves low latency: 17-frame chunks every 0.70 seconds at 24 fps, about 1.1 seconds behind input.
Meta description: SolarRenderer converts untextured engine structure passes into styled real-time video using MiniMax-H3 LoRAs and reports Elo 1313 structure benchmark.
Tags: SolarRenderer, real-time video generation, structure guidance, MiniMax-H3, game rendering
Featured image: abstract technology style showing untextured geometry becoming a stylized video stream, no real people, no company logos, no text
Background: structure passes need visual style
An untextured structure pass from any engine may already define where objects are, how they move, and when each frame happens, yet it still lacks a finished, styled video look. SolarRenderer addresses that gap by converting such passes into finished, styled video guided by a reference first frame.
Development: guided by one first frame
The model is built on MiniMax-H3 with a control LoRA (low-rank adaptation) and a few-step distilled LoRA. Its output follows the input geometry, motion, and timing frame by frame, so every object remains where the engine placed it and the render keeps time with the game. It accepts any game, character, or scene: Blender, three.js browser games, game engines, even a phone filming a screen. It renders in the look set by the first frame, with no per-game training.
Latest progress: benchmark and latency numbers
On a held-out benchmark, SolarRenderer ranks first on structure with an Elo score of 1313 against nine other methods. Its block-causal student renders 4 steps per chunk and produces a 17-frame chunk every 0.70 seconds at 24 frames per second, running about 1.1 seconds behind the input. With a key-value cache, or KV cache, it keeps going as long as the input does. The release frames it as a controller-driven, endless stream.
Resources are listed in the paper, project page, code, and weights, data, and benchmark collection.
Comments
No comments yet. Be the first to comment.