TRELLIS.2

Most image-to-3D pipelines squeeze geometry through implicit fields and lose thin structures along the way; TRELLIS.2 instead generates a field-free sparse voxel format called O-Voxel that stores arbitrary surface attributes directly, so open surfaces, non-manifold joints, and fine topology survive intact. A sparse 3D VAE with 16x spatial downsampling feeds a plain diffusion transformer at 4B parameters, and every output carries full PBR materials covering base color, roughness, metallic, and opacity, ready for game engines and DCC tools. Generation is fast for the class: about 3 seconds at 512^3 resolution and 17 seconds at 1024^3 on an H100. The repo ships inference and training code, a data preprocessing toolkit, web demos, and the 4B weights on Hugging Face; install is a conda-driven script wanting an NVIDIA GPU with roughly 24GB VRAM and CUDA 12.4. MIT licensing on Microsoft's successor to the original TRELLIS has made it the default open choice for production 3D asset generation, reflected in over 10k GitHub stars.