microsoft/TRELLIS.2-4B - 数据集下载 - GITATTRIBUTES/JSON/MD/SAFE…

TRELLIS.2: Native and Compact Structured Latents for 3D Generation

Model Name: TRELLIS.2-4B

Paper: https://arxiv.org/abs/2512.14692

Repository: https://github.com/microsoft/TRELLIS.2

Project Page: https://microsoft.github.io/trellis.2

Introduction

TRELLIS.2 is a state-of-the-art large 3D generative model designed for high-fidelity image-to-3D generation. It leverages a novel "field-free" sparse voxel structure termed O-Voxel and a large-scale flow-matching transformer (4 Billion parameters).

Unlike previous methods that rely on iso-surface fields (e.g., SDF, Flexicubes) which struggle with open surfaces or non-manifold geometry, TRELLIS can reconstruct and generate arbitrary 3D assets with complex topologies, sharp features, and full Physical-Based Rendering (PBR) materials—including transparency/translucency.

Model Details

Developed by: Jianfeng Xiang, Xiaoxue Chen, Sicheng Xu, Ruicheng Wang, Zelong Lv, Yu Deng, Hongyuan Zhu, Yue Dong, Hao Zhao, Nicholas Jing Yuan, Jiaolong Yang
Model Type: Flow-Matching Transformers with Sparse Voxel based 3D VAE
Parameters: 4 Billion
Input: Single Image
Output: 3D Asset (Mesh with PBR Materials)
Resolution: Varies from 512³ to 1536³ (Voxel Grid Resolution)

Key Features

O-Voxel Representation: An omni-voxel structure that encodes both geometry and appearance. It supports:
- Arbitrary Topology: Handles open surfaces, non-manifold geometry, and fully-enclosed structures without lossy conversion.
- Rich Appearance: Captures PBR attributes (including opacity for translucent surfaces) aligned with geometry.
- Efficiency: Instant optimization-free bidirectional conversion between meshes and O-Voxels (ms to seconds).
High-Resolution Generation: The model is trained to generate fully textured assets at up to 1536³ resolution.
High-Fidelity while Compact Latent Space: Utilizes a Sparse 3D VAE with 16× spatial downsampling, encoding a 1024³ asset into only ~9.6K latent tokens with negligible perceptual degradation.
Shape-conditioned Texture Generation: Generates textures for input 3D meshes and reference images.
State-of-the-Art Speed: Inference is highly efficient; see table below.

Inference Speed (NVIDIA H100 GPU)

Resolution	Time
512³	~3 seconds
1024³	~17 seconds
1536³	~60 seconds

Requirements

System: The model is currently tested only on Linux.
Hardware: An NVIDIA GPU with at least 24GB of memory is necessary. The code has been verified on NVIDIA A100 and H100 GPUs.
Software:
- The CUDA Toolkit is needed to compile certain packages. Recommended version is 12.4.
- Conda is recommended for managing dependencies.
- Python version 3.8 or higher is required.

Known Limitations

Geometric Artifacts (Small Holes): While O-Voxels handle complex topology well, the generated raw meshes may occasionally contain small holes or minor topological discontinuities. For applications requiring strictly watertight geometry (e.g., 3D printing), we provide accompanying mesh post-processing scripts, such as hole-filling algorithms.
Base Model w/o Alignment: TRELLIS.2-4B is a pre-trained foundation model. It has not been aligned with human preferences (e.g., via RLHF) or fine-tuned for specific aesthetic standards. Consequently, the outputs reflect the distribution of the training data and may vary in style; users may need to experiment with inputs to achieve the desired artistic result.

We are actively working on improving the model and addressing these limitations.

Usage

*Note: Please r

microsoft/TRELLIS.2-4B 已完整同步