Pixal3D at a glance
Pixal3D is designed to preserve visible-image evidence through direct pixel-to-3D correspondence. The current main implementation builds on a TRELLIS.2 backbone; the paper branch preserves the reported SIGGRAPH 2026 method.
- Conditioning
- Pixel-aligned feature back-projection
- Current implementation
- TRELLIS.2-based — The paper branch uses the original reported implementation
- Venue
- SIGGRAPH 2026
- License
- MIT — Third-party components retain their own terms
Where it stands out
- Explicit visible-pixel correspondence
- Geometry and PBR texture output
- Inference and training code are available
Limits to plan around
- Results from the main and paper branches should not be mixed in one benchmark
- Pixel alignment improves use of visible evidence but cannot reveal genuinely unseen surfaces
- As a research release, setup and interfaces may change
Pixel-aligned conditioning
Instead of relying only on global or attention-injected image features, Pixal3D back-projects features into 3D to establish more direct correspondence with visible pixels.
Record the implementation you tested
The official repository distinguishes the latest TRELLIS.2-based main branch from the original paper implementation. A credible comparison should name the branch, checkpoint, and settings.
This is an independent editorial profile. Trellis 3D is not affiliated with or endorsed by the model developer.