3D Data · Material-Import Bug Audit (renders_cond)

Detecting assets whose conditioning render lost its materials · 1.1M scanned · 2026-07-09

data-qualitymaterial-bugready_v4filtering

TL;DR

91,814 assets (8.4%) have a material-import bug: the Blender-3.0.1 conditioning render (the model INPUT) came out grey/untextured, while the object's real PBR material (the SUPERVISION) is colorful. Input and supervision disagree.

Detected by comparing two saturations per asset: cond_sat (render foreground) vs gt_sat (decoded PBR base_color). BUG = render grey AND material colorful. This separates true bugs from the 231,673 (21%) assets that are legitimately untextured.

The bug is not a Blender-version fault — re-rendering the same GLB (even with 3.0.1) restores the material; the original run had a transient material-import failure. Fix = re-render the flagged 91,814.

1 · Bug rate by subset

Concentrated in Sketchfab-sourced sets: TexVerse 11.3% (51,038) + SketchfabV1 9.7% (33,788) = 92% of all bugs.
ObjaverseXL_github looks clean (1.6%) but is mostly legitimately untextured raw meshes (51% legit-grey), not bug-free renders.
Cross-validated: SketchfabV1 gives 9.7% via PBR-decode and 8.7% via an independent raw-GLB read — the two methods agree.

① Material-bug rate per subset. Sketchfab-sourced sets (TexVerse, SketchfabV1) carry almost all of it.

2 · Detection logic

cond_sat = mean HSV saturation of the render's foreground (grey render → ~0).
gt_sat = mean saturation of the decoded PBR base_color (the true material).
BUG = cond_sat < 0.06 AND gt_sat > 0.12 (top-left quadrant). Both-grey = legit untextured; cond colorful = fine.

② Each point is an asset. The red top-left quadrant (grey input, colorful material) is the bug; the dashed lines are the thresholds.

3 · The bug, seen

Left of each pair = cond render (what the model sees — grey). Right = decoded PBR base_color (the true material — colorful).
TexVerse examples are decoded from latents alone (its raw meshes were deleted) — recovered via shape-decoder structure + tex-decoder color.
Clear cases: grey blob → purple/gold character; grey lump → brown-and-white cat; grey crystal → red crystal.

③ 16 flagged bugs. The grey render on the left is the training input; the colorful projection on the right is the material it was supposed to show.

4 · How the GT was recovered (the enabling trick)

Decoding a stored PBR latent needs the geometry octree, which is thrown away on save. TexVerse/Obj_sketchfab raw meshes are gone, so no re-voxelization.
Fix: run shape_dec(return_subs=True) to regenerate the subdivision octree, then tex_dec(guide_subs=subs) to paint base_color onto it. Shape and tex latents share the exact same sparse structure, so the octree transfers.
Uniform, GPU-only, ~0 decode errors — covers every subset including the raw-deleted ones.

5 · Next step

Flag list written to material_bug_flags.jsonl (1,097,054 rows: sha / subset / cond_sat / gt_sat / verdict).
Re-render the 91,814 bug assets. SketchfabV1 + github + small sets have raw on disk; TexVerse's 51k need a raw re-download (TexVerse-1K).
Keep the 231,673 legit-untextured as-is — they are not bugs.

Detector: cond_sat (renders_cond webp) vs gt_sat (PBR base_color via shape_dec->tex_dec). 2x H200 nodes + 1 CPU node, ~4h. Scores at /fsx/home/weikai.huang/3dgen/data/material_bug_flags.jsonl