The Context: I recently ran a repair pass on a mesh and got an interesting result:
10 detached clusters (2,861 faces) floating above the hull, marked in red by the detector, the same tool meant to delete them.
Repair stripped the debris, leaving the mesh completely watertight.
Catch: It wasn’t watertight because the hole-fill succeeded. The fill was attempted, but because it would have introduced 336 non-manifold edges, it reverted itself.
All 25 open boundary edges just happened to live on the floating debris that got removed.
The tool ended up reporting the mesh as watertight on its own output report without claiming credit for “fixing” the holes. The side effect did the heavy lifting.
The Question for Discussion: When your repair tool can technically close a hole, but only by degrading the topology (e.g., introducing non-manifold geometry), do you actually want it to close the hole? Or would you prefer it fails and leaves the boundary open?
GitHub repo on more results: GitHub - alza123123/3dqa-benchmark: The State of AI-Generated 3D Assets — a public census over 2,307 real generative-AI exports. Reproducible with the published 3dqa package. · GitHub
