TRELLIS at a glance
TRELLIS introduced Structured LATents (SLAT) as a flexible foundation for generating several 3D representations, including meshes, radiance fields, and 3D Gaussians.
- Image model
- 1.2B parameters — TRELLIS-image-large
- Text variants
- 342M to 2.0B parameters
- Representation
- Structured LATent (SLAT)
- License
- MIT for models and most code — Review submodule terms
Where it stands out
- Multiple output representations
- Image- and text-conditioned variants
- Published training code and dataset documentation
Limits to plan around
- TRELLIS and TRELLIS.2 use different core representations and should not be treated as the same model
- The upstream project recommends image conditioning when detail and creativity matter
- Output readiness depends on the selected representation and downstream task
Why structured latents matter
SLAT stores local latent features at active spatial locations. Decoders can turn that common representation into different 3D asset types.
TRELLIS is not shorthand for TRELLIS.2
The first generation centers on SLAT and multiple 3D representations. TRELLIS.2 centers on the field-free O-Voxel representation and native PBR mesh generation.
This is an independent editorial profile. Trellis 3D is not affiliated with or endorsed by the model developer.