Explorable 3D world generation is becoming a foundational capability for spatial and embodied intelligence. Training agents that can navigate, reason, and interact with environments requires not just static datasets, but rich, immersive worlds that support free-view exploration and consistent geometry. Recent works like Matrix3D1 have pushed this frontier forward by combining panoramic generation with explicit 3D reconstruction, enabling higher-quality and more coherent environments than prior video-only methods.
In this blog, we describe how we deployed Matrix3D on AMD GPUs (Instinct™ MI250 GPU and Instinct™ MI300 GPU). With several targeted modifications and optimizations, we made the framework both more efficient and more portable. The end-to-end generation time for a single world is reduced from 2887s to 1306s on a single MI250 GPU, and from 972s to 482s on a MI300 GPU.
We first provide a brief introduction to Matrix3D. We then present the optimizations we implemented to improve efficiency, including replacing CUDA-specific rendering kernels with more portable Triton kernels, accelerating mesh optimization with more efficient solvers, using gsplat2 for faster 3DGS fitting, and refactoring parts of the pipeline to remove overhead from repeated model loading and unnecessary I/O.