1,039,752 PBR-present assets scored by a vision-LM judge (99.995% parse). recommended = yes: 440,229 (42.3%) — lands remarkably close to the ~500k ablation target with zero manual thresholding.
Each asset gets 6 independent 1-10 axes — 3 quality (structural / texture / aesthetic) + 3 complexity (part / detail / color) — plus style, validity, and hard-exclude issue flags. Complexity is scored separately from quality, so "well-made but simple" and "cluttered but sloppy" are distinguishable.
Confirmed at scale: "has-PBR" is a weak signal — 307k assets flagged untextured and 374k single_color despite all having PBR latents. The VLM catches material rot the PBR-file check cannot.
1 · Verdict distribution
yes 440,229 (42.3%) — VLM would include in a high-quality set
maybe 5 — the force-spread prompt pushed the model to commit, near-zero fence-sitting
① yes ≈ 440k lands near the 500k target — use yes directly, or top-up from no by score to reach 500-550k.
2 · Interactive sample gallery — YES vs NO (drag to spin)
Drag any thumbnail left/right to spin the object — each card cycles the 16 rendered viewpoints the VLM saw (a pseudo-3D turntable). The two groups are the model's actual verdict: kept vs excluded. Score bars: blue = quality, purple = complexity.
recommended = YES · kept for the high-quality set
clean geometry + real materials, across styles & complexity
recommended = NO · excluded
untextured / single-color / low-poly / scan — note several have decent geometry (str≥6) but dead materials (tex≤2)
3 · Six scoring axes
Quality (blue): structural μ=5.74, texture μ=3.9, aesthetic μ=4.7 — texture is the harshest axis (lots of bare/greyed materials).
Complexity (purple): part μ=4.46, detail μ=3.82, color μ=3.14 — scored independently of quality; part vs detail genuinely diverge (single-part statues with rich relief; multi-part mecha with plain surfaces).
These two families let ablations filter on "make quality" and "geometric richness" as separate knobs.
② Six axes. Texture (μ=3.9) is the main quality bottleneck; complexity axes spread widely — the pool spans trivial primitives to dense assemblies.
4 · Style mix (ablation control)
Every asset tagged with one of 7 styles — control the realistic:stylized:cartoon:lowpoly ratio in an ablation instead of taking whatever the raw distribution gives.
