Microsoft Unveils TRELLIS.2: A 4B Parameter 3D Generation Model

Microsoft has released TRELLIS.2, a large-scale 3D generation model with 4 billion parameters, featuring an innovative O-Voxel sparse voxel structure that supports high-quality image-to-3D conversion. The model can handle complex topologies, including open surfaces and internally enclosed spaces, and provides rich PBR material modeling, including base color, roughness, metallic, and opacity. It runs efficiently on NVIDIA H100 and A100 GPUs, generating 3D assets with resolutions from 512³ to 1536³ in just seconds to a minute. The pretrained model has been released on Hugging Face, with the code open-sourced, advancing the application of AI 3D generation technology in gaming, film, and design industries.