TRELLIS.2 at a glance
TRELLIS.2 generates high-resolution textured meshes from one reference image. Its field-free sparse O-Voxel representation is designed for complex topology, open surfaces, internal structures, and PBR attributes.
- Parameters
- 4B — Official TRELLIS.2-4B checkpoint
- Generation resolution
- 512³ to 1536³ — Options listed by the official project
- Representation
- O-Voxel — Sparse, field-free structured representation
- Official local baseline
- Linux, NVIDIA GPU, 24 GB+ VRAM — Community workflows may differ
- License
- MIT — Dependencies may carry separate terms
Where it stands out
- Native geometry and PBR material generation
- High-resolution structured output
- Representation supports open and non-manifold surfaces
- Official inference, export, and training code
Limits to plan around
- The official local setup is demanding and currently documented for Linux
- A single image cannot fully specify hidden surfaces
- Generated meshes still require inspection for holes, density, scale, and downstream topology needs
- Alpha can be stored in an export without being enabled by the receiving material
From image evidence to a structured 3D latent
Image features condition sparse generative stages for structure, geometry, and material attributes. O-Voxel provides the representation used to move between textured meshes and compact structured latents.
What the output contains
The official pipeline can export a PBR-ready GLB with geometry and surface attributes. Inspect each channel and the mesh itself before treating the asset as production-ready.
Local deployment starts from the official baseline
The upstream repository documents a Linux environment, CUDA build dependencies, and at least 24 GB of NVIDIA GPU memory. Lower-memory and Windows routes are community adaptations, not equivalent official support.