NVIDIA Labs' StreamRender-H3 renders a code engine with MiniMax-H3

NVIDIA Labs has posted StreamRender-H3 on its Sol-Engine site. The page carries no calendar date. Its BibTeX block sets year = {2026}. Treat the page as a research write-up to catch up on, not as an announcement with a day attached.

The split the page describes is simple. A code game engine keeps the parts of a world that code can state: maps, physics, rules, and the camera. It draws a semantic map, one flat color per object class, without texture or lighting. A streaming video model then paints the frame the player sees. The page says that model is fine-tuned from MiniMax-H3 Ref2VA, and that the deployed generator denoises each chunk in two steps. MiniMax-H3 is the base model the page names. The subject here is the renderer on the Sol-Engine page.

The page says the native MiniMax-H3 Ref2VA model generates a clip with 25 denoising steps. It says a 15-second clip made in one pass stops following the semantic video and the reference image closely after about five seconds, so the authors split each clip into three segments of about five seconds. The streaming generator is fine-tuned from that model. The page says teacher forcing and resampling forcing pre-train it as a multi-step model, and a later distillation stage trains it to produce each chunk in two steps.

What the page measures

These figures are printed on the project page for the racing demo. They are the authors’ measurements:

• 218 ms for the diffusion render of one streaming step on 8 GB200 GPUs, down from 799.7 ms.
• 3.67× as the speedup of that diffusion render from the PyTorch baseline.
• 23.8 ms for a causal H3 VAE decode of 2 latent frames at 720p on one GB200.

The page says the 218 ms step and the eight-GPU setup are how the streaming render runs. It does not present those numbers as a third-party benchmark.

Racing data, and eight GPUs

The limitations section says everything reported comes from the racing demo. The renderer, the two-step student, and the critics are trained only on racing data, with a third-person camera behind the player’s car. Test drives come from the same game. The page says the interface itself is not tied to racing: any engine that can paint its scene with one flat color per class can drive the renderer. Covering more games, and measuring how much new data each one needs, is listed as later work.

The same section says the streaming render still uses the full MiniMax-H3 diffusion transformer on eight GPUs. Cutting the parameter count, lowering the cost of each streaming step, and improving visual quality are the other directions the page names.

Who signed it

Core contributors on the page are Yanzuo Lu, Tian Ye, Yitong Li, Shuchen Xue, Haozhe Liu, Song Han, and Enze Xie. Song Han and Enze Xie also appear on the Sol-Engine BibTeX entry the page asks readers to cite alongside the project. The credits line for MiniMax calls MiniMax-H3 Ref2VA open weights and code the base of the streaming renderer and of the H3 VAE.

The page links a source tree under NVlabs/Sana and the MiniMax-H3 weights. Demo clips stay on the project page.