StreamRender-H3: Real-Time Code-Based Interactive Worlds via Streaming Diffusion

StreamRender-H3 is a framework that renders code-based interactive worlds into photorealistic video in real time. The architecture decouples World Logic (managed by a Code Game Engine) from Visual Generation (handled by a streaming diffusion model), connecting them via an Intermediate State Representation known as a Semantic Map.

Architecture and Performance

The system uses a MiniMax-H3 Ref2VA based video model that receives semantic streams to determine frame structure, while appearance is controlled via reference images and prompts. Key performance metrics from the racing demo include:

• Diffusion Rendering: 218ms per streaming step on an 8 GB200 GPU setup, achieving a 3.67x speedup over the 799.7ms PyTorch baseline.
• Decoding: The custom Fast Causal H3 VAE decodes 720p latent frames in 23.8ms on a single GB200.
• Denoising: The deployed generator performs 2 steps per chunk.

Fast Causal H3 VAE

To meet real-time constraints, the team developed a specialized decoder using an agent loop with Claude Code. The released H3 VAE was too slow for streaming (approx. 870ms per block at 720p). The new Fast Causal H3 VAE reduces parameters to 1.62B and employs causal attention with a sliding window of cached past frames. Optimization steps included removing spatial tiling, using FlashAttention-4, and kernel fusion, reducing latency from 146ms (naive PyTorch) to 23.8ms.

Training and Inference

The model was trained on 10,000 paired clips (approx. 15 seconds each) from 5,100 recorded races, with reference images generated by GPT-Image-2.5 and prompts by Claude Opus 5.5. Inference acceleration was achieved through multi-GPU parallel inference, static-context KV caching, and attention optimization, bringing the total diffusion render time down to 218ms. The system currently supports third-person camera scenarios and is designed to allow independent replacement of the engine and renderer.