TRELLIS.2 on a 24 GB RTX 5090 Laptop: Local 3D Experiments
This is a local, offline image-to-3D evaluation of Microsoft's TRELLIS.2. It produced four verified static PBR GLB outputs marked geometry_reference_only: an official sample, a standing footballer, a footballer in a power-shot pose, and an elder alone. It did not create a character animation: each MP4 below is a four-second camera orbit around one static reconstruction.
Tests ran locally in WSL2 on a 24 GB RTX 5090 Laptop GPU, fixed at source revision 75fbf0183001ed9876c8dbb35de6b68552ee08bd, with no external inference.
Results
Each successful GLB was parsed locally and had nonzero geometry. Every video below is a four-second orbit of one static reconstruction.
| Input | Result | Verified geometry |
|---|---|---|
| Official TRELLIS.2 sample | GLB passed | 421,412 vertices / 484,799 faces |
| Standing footballer | GLB passed | 343,320 vertices / 497,550 faces |
| Footballer power-shot pose | GLB passed | 349,518 vertices / 494,414 faces |
| Elder in front of a house | Elder GLB passed; house removed | 350,133 vertices / 484,438 faces |
| House by itself | Render preview passed; GLB export failed | — |
Footballer: a static power-shot pose
One clean action reference produced one static power-shot mesh. It is a pose reference, not a rigged or animated kick.
The elder-and-house scene: why the building vanished
The source had no alpha channel, so foreground preprocessing retained the elder and removed the house. The result is an elder-only GLB, not a complete scene.
The house: reconstructed separately
We removed the elder and used the house as the only subject. Inference, texture generation, and the 60-frame render completed; final GLB mesh simplification ran out of VRAM. The video is therefore a house preview, not a delivered GLB.
Practical input rule
Use one isolated foreground asset per input. A person in front of a building should be split into separate character and building inputs, then composed later.