Feature matrix
| Criterion | Model A | Model B |
|---|---|---|
| Primary input | Single image | Single 512 × 512 object image |
| Representation | O-Voxel | Feed-forward mesh reconstruction |
| Asset focus | Complex topology and high-resolution PBR | Relatively low-polygon UV-unwrapped mesh |
| Material handling | Base color, roughness, metallic, opacity | Texture, delighting, roughness, metallic |
| Local platform note | Official Linux + NVIDIA baseline | CUDA path plus experimental MPS and Windows notes |
| License | MIT, with dependency terms | Stability AI Community License |
Choose by workflow
Complex generated forms and detailed PBR asset studies
Conventional UV-textured mesh reconstruction and compact-asset experiments
Practical verdict
Test TRELLIS.2 when topology complexity and high-resolution PBR generation are priorities. Test Stable Fast 3D when a compact UV-unwrapped reconstruction pipeline is closer to the delivery target. License terms differ and belong in the decision.
Generation scope versus reconstruction scope
The models emphasize different asset paths. Evaluate whether the task rewards inferred structural detail or a compact reconstruction aligned with a conventional UV pipeline.
Test the destination asset
Import both exports into the same renderer, use the same lighting and scale, then measure file size, decoded resources, draw calls, and visible detail at the intended camera distance.
Method and limits
Normalize the input framing, then compare mesh shape, UV layout, texture lighting, material parameters, transfer size, and target-engine behavior. Do not reuse an upstream runtime claim as a promise for different consumer hardware.