StreamRender-H3 renders interactive game worlds in real time with a 2-step diffusion model at 24 FPS
songhan_mit · x · 2026-10-12
The Sol-Engine team introduced StreamRender-H3, post-training MiniMax's H3 Ref2VA into a real-time streaming neural renderer where a code game engine handles world logic and a video model generates the visuals:
- Architecture: the engine maintains physics, object states and player interactions; changing the reference image and prompt renders the same interactive world in entirely different styles.
- Performance: 2 denoising steps, 24 FPS, 273ms per chunk on 8×GB200 GPUs; diffusion render cut from 799.7ms to 218ms per step (3.67× speedup); Fast Causal H3 VAE decodes two 720p latent frames in 23.8ms on one GB200.
- Built with streaming post-training, distillation, a causal H3 VAE and full-stack inference acceleration. Demo and code released.
More from Embodied
- Yeah Hand: open-source humanoid robotic hand kits start at just $217 — CyberRobooo · 2026-10-12
- Fly-by-Code: coding agents program a tiltrotor drone with an arm, sim-to-real — furongh · 2026-10-12
- Artisan: multi-agent pipeline generates physically grounded articulated 3D assets — nanyang-technological-university-singapore · 2026-10-12
- DexForce's UNITAS: a 1.7B 3D-native world action model tops LIBERO at 99.8%, beats 8B rivals — 机器之心 · 2026-10-12
- SuperNav: agentic navigation with pretrained multimodal models, no fine-tuning — rsasaki0109 · 2026-10-12
- Dad vibe-codes a physical MLB scoreboard for his son with Raspberry Pi and Claude Code — dtizzlenizzle · 2026-10-12