Datasets:
A geometry pass over the corpus, a bug in my own metric, and a question about the point clouds
I built a geometry linter and ran it over the 3D Arena corpus, topology only: watertightness, non-manifold edges, degenerate faces, disconnected components, UVs, poly budget. No aesthetic judgement in it anywhere. Everything below I pulled on 2026-08-01, 2,307 entries across 23 generators; the dataset has moved since, so that date governs the counts rather than the current row count.
Posting rather than just publishing, for three reasons.
A bug in my own metric. My watertight check asked the wrong question, its predicate was false for an edge shared by more than two faces as well as by exactly one, so a closed mesh with a non-manifold edge got reported as holed, while its own boundary-edge count read zero in the same record. Minimal repro is two closed cubes sharing an edge. Fixed. Blast radius 127 of the 2,004 meshes, and it moved a published number of mine toward fewer defects, not more. My own pre-fix reclassification said 128; the extra one was already scored correctly, so there was nothing to flip. I would rather show the disagreement than the tidy number.
The per-asset output is yours if it is useful. One row for every asset, raw counts, no score and no ranking, no verdict on any model. A different axis from human preference rather than a competing one, and cheap to regenerate as the corpus grows.
303 of the 2,307 entries are .ply point clouds, so a mesh-topology check has nothing to operate on and my counts run over the remaining 2,004. Drop them, exclude them from the denominator, report them as a separate unmeasurable class, or treat point-cloud output as its own track? How would you have them handled?
And the broader version: would a geometry axis be useful alongside human preference, or does it cut against what the leaderboard measures? It has no opinion on whether a model is good, edge flow, proportions, silhouette, whether the thing reads as the prompt, all invisible to it. Passing means structurally valid, not well made, and I do not think that distinction survives a leaderboard unless it is said loudly.
Happy to hand over the data if it is useful.