Evaluation Data: TRELLIS vs. TRELLIS.2-4B geometry metrics in 2,307-model census

Hi TRELLIS team,

I executed an automated geometry quality evaluation across 2,307 assets from 23 generators (sampled from the 3D Arena dataset).

Comparing both TRELLIS generations in the dataset yielded the following metrics:

• TRELLIS (n=100): 60.0% non-watertight | 0.0% non-manifold | 4.0% > 30k tris
• TRELLIS.2-4B (n=101): 91.1% non-watertight | 14.9% non-manifold | 100% > 30k tris

The higher face density in TRELLIS.2-4B accounts for the triangle budget, but the non-manifold edge rise from 0% to 14.9% stood out during automated QA.

The evaluation runner, manifest files, and raw aggregate reports are published and reproducible via pip install 3dqa:
https://github.com/alza123123/3dqa-benchmark

Thought this data might be useful for your ongoing model evaluation.

Start with the published 3dqa-benchmark repository, its manifest files, and raw aggregate reports; reproduce the evaluation with pip install 3dqa. Compare the TRELLIS and TRELLIS.2-4B geometry metrics and verify the reported sample counts. Done would require an agreed investigation or remediation plan, which the issue does not specify.