NVIDIA research system where a code game engine drives a two-step streaming MiniMax-H3 video model for real-time rendering.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
StreamRender-H3 is a research project from NVIDIA's Sol-Engine / Sana team that separates world logic from visual generation. A code game engine (a three.js browser racing game in the demo) keeps maps, physics, rules and state, and outputs a semantic map in which each pixel is a flat colour identifying an object class. A streaming video diffusion model then renders each frame photorealistically from that map.
The renderer is built on MiniMax-H3 Ref2VA, which generates video and audio from a prompt plus reference images and videos. The semantic stream enters as a reference video aligned with the output latent frames. A reference image, produced by an image editing model from the first frame, and a prompt can change the style, materials, weather and lighting of the same play session. Generation uses two diffusion steps, and a persistent KV-cache layout brings diffusion render time to 218 ms per step.
The demo is trained only on racing, with 13 semantic classes, although the authors say the design is not specific to racing. According to creator coverage, it needs eight GB200 GPUs and over 200 GB of files, and the model and code are reported as released. I could not confirm repository, weights or licence URLs from the provided material.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Game code outputs a color-coded semantic map and a video model renders the frames in real time, allowing restyling with a reference image. Needs eight GB200 GPUs (~218 ms per step) and is trained only on racing; Claude Code helped build a faster decoder.
World model based on Minimax H3 that converts rough game-engine output into realistic visuals in real time, using two-step generation. Needs eight GPUs and over 200 GB of files; model and code released.
Closest neighbours by meaning, not just by category.
Model
Research framework that turns the Minimax H3 video generator into an interactive world model controlled by key presses.
Model · NVIDIA
Open few-step diffusion refiner from NVIDIA that regenerates low-resolution H3 video drafts at 768P to 4K, 5–15 seconds.
Model
Research code that streams video generation chunk by chunk from a rough 3D proxy, built on Code World Model.
Model · Hao AI Lab
Open-weight 4-step sparse-distilled version of MiniMax H3 for fast text-to-video-and-audio generation on Blackwell GPUs.
Research
Research framework where a coding agent maintains world state as code, rendered to video by a fine-tuned MiniMax-H3 model.
Model · NVIDIA
NVIDIA's 117M-parameter looped transformer for multi-view 3D reconstruction of camera poses and dense geometry.
