TRELLIS.2 is a 4-billion-parameter large generative model developed by Microsoft for high-fidelity image-to-3D asset generation. The repository contains the complete implementation, inference code, training framework, and supporting tools for this state-of-the-art 3D generation system. The model is written in Python and represents a significant advancement in converting 2D images into fully textured 3D assets with complex geometric and material properties.
The core innovation of TRELLIS.2 is its "field-free" sparse voxel representation called O-Voxel, which enables the model to handle arbitrary topology in 3D geometry. Unlike traditional iso-surface field approaches, O-Voxel robustly reconstructs open surfaces such as clothing and leaves, non-manifold geometry, and internal enclosed structures without lossy conversion. This representation is encoded through a Sparse 3D VAE with 16× spatial downsampling, creating a compact latent space that allows efficient generation using vanilla Diffusion Transformers.
The model generates high-resolution fully textured 3D assets with exceptional speed and quality. On an NVIDIA H100 GPU, it produces 512-cubed resolution assets in approximately 3 seconds, 1024-cubed resolution in 17 seconds, and 1536-cubed resolution in 60 seconds. Beyond basic color information, TRELLIS.2 models rich surface attributes including base color, roughness, metallic properties, and opacity, enabling photorealistic rendering and transparency support in the generated assets.
The repository includes comprehensive inference capabilities for image-to-3D generation and shape-conditioned texture generation. Data processing is streamlined with rendering-free and optimization-free conversions that complete in under 10 seconds on a single CPU for converting textured meshes to O-Voxel format, and under 100 milliseconds on CUDA for the reverse conversion. The system exports generated assets in GLB format with PBR-ready materials suitable for immediate use in 3D applications.
Installation requires Linux, an NVIDIA GPU with at least 24GB of memory, CUDA Toolkit 12.4, and Python 3.8 or higher. The repository provides setup scripts for conda environment configuration and dependency management. Pretrained weights for the TRELLIS.2-4B model are available on Hugging Face, along with a web demo interface for interactive asset generation.
The repository includes full training infrastructure enabling users to train TRELLIS.2 from scratch or fine-tune on custom datasets. Training components include the shape SC-VAE, texture SC-VAE, and multi-stage flow models for sparse-structure, shape, and texture generation. The training pipeline accepts detailed configuration files and supports distributed training across multiple nodes and GPUs. Data preparation tools are provided for converting raw 3D assets into O-Voxel representations, with comprehensive documentation in the data_toolkit directory.
The project is released under the MIT license and includes links to the research paper on arXiv, a project website with full-quality demonstrations, and Hugging Face model and demo spaces. The repository represents a complete production-ready system for 3D generative modeling with both inference and training capabilities fully documented and implemented.
