Feature matrix
| Criterion | Model A | Model B |
|---|---|---|
| Primary input | Single image | Single image |
| Conditioning | Image-conditioned sparse generation | Explicit pixel feature back-projection |
| Current relationship | Upstream TRELLIS.2 system | Main branch builds on TRELLIS.2 |
| Output focus | High-resolution mesh + PBR attributes | Visible-detail geometry + PBR textures |
| Version caveat | Checkpoint and resolution must be recorded | Main and paper branches must be distinguished |
| License | MIT, with dependency terms | MIT, with third-party terms |
Choose by workflow
A documented upstream O-Voxel workflow and baseline comparisons
Experiments focused on visible-pixel fidelity and alignment
Practical verdict
Use TRELLIS.2 as the clean upstream baseline. Evaluate Pixal3D when preserving visible-image correspondence is central, and always record whether the main or paper branch produced the result.
Understand the model relationship
The current Pixal3D main implementation extends a TRELLIS.2 foundation, so results should be framed as an architectural or conditioning comparison rather than two wholly independent systems.
Measure fidelity from more than one view
Compare the conditioned view, side silhouette, inferred back, thin structures, topology, and material response. A close front projection can coexist with weak hidden geometry.
Method and limits
Pin repository commit, branch, checkpoint, preprocessing, seed, resolution, and post-processing. Compare visible and hidden regions separately; pixel alignment cannot be fairly scored only from the conditioned front view.