TRELLIS.2 is notable because it combines high output resolution, a native O-Voxel representation, and PBR material attributes in one open-source image-to-3D system. That makes it worth evaluating for production—but it does not make every output production-ready by default.
Earlier versions of this article included invented user stories and unsupported side-by-side claims about TripoSR and Instant3D. Those claims have been removed.
Published capabilities
Microsoft documents:
- a 4B-parameter image-to-3D model
- 512³ to 1536³ output resolution
- a Sparse 3D VAE with 16× spatial downsampling
- open-surface, non-manifold, and internal-structure handling
- Base Color, Roughness, Metallic, and Opacity
- GLB export in the reference code
- MIT licensing for the model and code, with separate dependency licenses
Primary sources: project page and repository.
Why the representation matters
O-Voxel is designed to express topology that can be awkward for an iso-surface field, including open sheets and enclosed internal structures. Geometry and appearance are represented together, supporting a PBR-oriented workflow.
This is a capability claim about the representation. It is not proof that every generated fence, garment, glass object, or hidden surface will be correct.
Performance: use the right denominator
The official H100 measurements are about 3 seconds at 512³, 17 seconds at 1024³, and 60 seconds at 1536³ for shape plus material generation.
A hosted workflow also includes queue time, input processing, upload, export, storage, download, and possible retries. Production planning should measure end-to-end time and cost per accepted asset rather than quoting inference time alone.
Production acceptance checklist
Evaluate a representative batch and record:
| Area | What to inspect |
|---|---|
| Silhouette | visible outline and proportions |
| Hidden geometry | back, underside, occluded regions |
| Topology | non-manifold edges, intersections, density |
| Materials | map availability, alignment, plausible response |
| UVs | overlaps, seams, texel density |
| Delivery | GLB size, import warnings, scale, axes |
| Reliability | failures, retries, queue time |
| Labor | cleanup minutes per accepted asset |
Where TRELLIS 2 fits
It is a strong candidate when:
- the input is a clear single object
- PBR attributes matter
- open or complex structures appear in the source
- you want resolution, texture, remesh, decimation, and sampling controls
- GLB is part of the delivery path
Use extra review when the object has precise mechanical dimensions, strict brand geometry, heavy occlusion, multiple interacting objects, or regulated use.
Compare without marketing shortcuts
To compare TRELLIS 2 with another model, use the same images, document every setting, export original files, and publish failures as well as successes. Do not infer mesh quality from a front-facing screenshot.
- TRELLIS 2 vs Meshy workflow comparison
- TRELLIS 2 vs Hunyuan3D model comparison
- TRELLIS 2 vs Tripo evaluation framework
- TRELLIS 2 vs CSM AI evaluation framework
- Best TRELLIS 2 alternatives
Try and inspect
Generate a model, download the GLB, and follow the Blender export and validation guide.