TRELLIS.2 is a Microsoft Research image-to-3D model based on a sparse, field-free representation called O-Voxel.
- Input
- Single image
- Checkpoint
- 4B parameters
- Output
- Mesh + PBR attributes
- Representation
- O-Voxel
The compact view
An image encoder supplies visual conditioning, sparse generative stages predict structured geometry and material latents, and the O-Voxel toolchain converts the result into a textured mesh.
Generation is not asset certification
Output can contain rich geometry and PBR data while still needing checks for unseen surfaces, mesh density, watertightness, scale, UVs, and target-engine constraints.