TRELLIS.2: state-of-the-art large 3D generative model (4B)
3 days ago
- #3D-generation
- #computer-graphics
- #AI-model
- TRELLIS.2 is a 4B-parameter large 3D generative model for high-fidelity image-to-3D generation.
- It uses a novel 'field-free' sparse voxel structure called O-Voxel for complex topologies and sharp features.
- The model supports full PBR materials, including Base Color, Roughness, Metallic, and Opacity.
- Generates high-resolution textured assets efficiently with vanilla DiTs and a Sparse 3D VAE.
- O-Voxel representation handles open surfaces, non-manifold geometry, and internal enclosed structures.
- Data processing allows instant conversions (textured mesh ↔ O-Voxel) in under 10s on CPU and 100ms on CUDA.
- Upcoming releases include inference code, pretrained checkpoints, and Hugging Face demo.
- Requires Linux, NVIDIA GPU (24GB+ memory), CUDA Toolkit 12.4, Python 3.8+, and Conda for dependencies.
- Includes specialized packages: O-Voxel, FlexGEMM, and CuMesh for high-performance processing.
- Released under MIT License, with some dependencies under their own licenses.