TRELLIS.2: Microsoft's open-source image-to-3D model generates PBR assets in seconds
Microsoft Research released TRELLIS.2, a 4-billion-parameter image-to-3D model that generates textured 3D assets with full PBR materials. It's MIT licensed, open source, and fast: 3 seconds at 512 cubed resolution, 17 seconds at 1024 cubed, and 60 seconds at 1536 cubed on an NVIDIA H100.
Microsoft's TRELLIS turning 2D images into fully textured 3D models
What it generates
Unlike earlier image-to-3D models that output basic meshes with baked vertex colors, TRELLIS.2 produces complete PBR material maps: albedo, metallic, roughness, normal, and opacity. These are the same maps that game engines and rendering pipelines expect, which means generated assets can go straight into Blender, Unreal Engine, or Unity without manual material setup.
The model uses flow-matching transformers and a novel O-Voxel representation. O-Voxel is a "field-free" sparse voxel structure with 16x spatial compression that handles arbitrary topology. That includes open surfaces, non-manifold geometry, and internal structures that previous iso-surface methods couldn't represent.
Speed and resolution
The generation speed scales with resolution:
| Resolution | Time (H100) |
|---|---|
| 512³ | ~3 seconds |
| 1024³ | ~17 seconds |
| 1536³ | ~60 seconds |
Mesh-to-O-Voxel conversion takes under 10 seconds on CPU. O-Voxel back to mesh takes under 100 milliseconds with CUDA acceleration.
How it compares to TRELLIS 1
TRELLIS.2 doubles the parameter count from 2 billion to 4 billion. The original TRELLIS could generate decent geometry but lacked proper PBR materials. Version 2 adds the full material pipeline, handles more complex topology, and runs at higher resolutions.
Tutorial: creating 3D models from 2D images with TRELLIS
How to use it
There are several ways to run TRELLIS.2:
Hugging Face demo. A free web demo lets you upload an image and get a 3D model back without any local setup.
Local deployment. The GitHub repository includes Docker images for local installation. You'll need an NVIDIA GPU with at least 24GB of VRAM for comfortable operation.
ComfyUI integration. Community-built ComfyUI nodes let you incorporate TRELLIS.2 into existing generation workflows alongside other AI tools.
Why this matters
Open-source image-to-3D has been improving steadily, but most models produce either decent geometry with poor textures or good-looking results that fall apart when you inspect the topology. TRELLIS.2 addresses both by generating clean geometry with proper PBR materials at production-relevant resolutions. The MIT license means you can use it commercially without restrictions.
For game developers and 3D artists, this is a useful prototyping tool. Drop in a concept art image, get a textured 3D model in seconds, and refine from there.