Distribution takeaway: ready_v3 is dominated by TexVerse (55%) + github (27%) = 82%; the three curated subsets (ABO + HSSD + Toys4k) are <1.5% combined — small but high-quality / high-coverage. ready_v4 would add a ~42% character-heavy bump on top.
2 · Counts by modality
Live counts (one-level-deep ls of each latent dir; 2026-06-30). cov = PBR ÷ shape
(material coverage — only assets that carry a real Principled-BSDF material get a PBR latent).
Subset
shape@512
ss@64
pbr@512
cov
renders_cond
Objaverse-XL · github
278,627
277,646
159,757
57%
280,951
Objaverse-XL · sketchfab
168,291
166,383
126,741
75%
166,155
ABO
4,470
4,319
4,470
100%
4,488
HSSD
6,668
6,620
5,941
89%
6,673
Toys4k
3,232
3,131
2,234
69%
3,232
ready_v2 total
461,288
458,099
299,143
65%
461,499
TexVerse (new in v3)
562,034
561,810
456,561
81%
565,030
🟢 ready_v3 total
1,023,322
1,019,909
755,704
74%
1,026,529
ready_v4 funnel — 621,891 glb downloaded (7.14 TiB, deduped) → 611,680 aesthetic-scored (98.4%) → 426,521 selected @ score ≥ 4.5 (69.7% pass, mean 5.08 vs TexVerse 4.27). Phase-2 (render → voxelize → encode → caption) not yet started, so v4 latent counts are 0 for now.
9 quality-curated objects per subset (scored from a candidate pool on framing + colour/texture +
detail, best view shown), so these are representative good assets — not the thin/degenerate tails.
1024² source renders, LANCZOS-downscaled. Observations below each grid are first-hand.
Objaverse-XL · github — widest geometric variety (mechs, props, characters, abstract solids), but ~43% carry NO material → rendered as bare white/clay meshes (goblet, star, sword, angel). This is the texture-coverage drag (57% cov). Framing varies; thin objects sit small in frame.Objaverse-XL · sketchfab — real artist assets with genuine textures (Nike shoe w/ logo, woven basket, diorama shop). High texture fidelity, 75% material coverage. A few objects render small/dark.ABO — Amazon furniture / home-goods; clean photoreal PBR, ~100% material coverage. Narrow domain (chairs, sofas, appliances), small set (4.5k).HSSD — indoor furniture & appliances (bed, office chair, bin, grill). Materials present but renders skew dark/muted; a couple are low-contrast from the sampled angle. 89% cov, 6.7k set.Toys4k — stylized toy-scale objects (hammer, knight helmet, animals, drum). Clean, unambiguous single objects; smallest set (3.2k), 69% cov.TexVerse (v3) — texture-rich & very diverse (branded can with crisp label, golfer, palm, tiled panel). Strength = high-res PBR. Caveats: scan/photogrammetry assets show jagged/torn silhouettes; some thin/degenerate geometry (the @1024 SIGSEGV cases).
4 · Conditioning-render format
Each asset ships 16 views on a Hammersley sphere (full azimuth + varied elevation, incl. top/bottom),
512×512 webp. Below: all 16 views of one TexVerse asset.
Feedback: some assets render dark and small-in-frame (object fills a small fraction of the 512 canvas, low contrast on black). Candidate fix before tokenization: an exposure + tight-bounding-box framing normalization pass so the conditioning signal is consistent.
16 Hammersley conditioning views of a single asset — this one is under-lit and small-in-frame (the failure mode noted above). Multi-view coverage itself is correct (full sphere).
5 · ready_v4 preview
ready_v4 = Sketchfab-v1, deduped. 426,521 assets selected at aesthetic ≥4.5 (69.7% pass, higher than TexVerse's 62.6% — mean 5.08). Phase-2 (render → voxelize → encode → caption) not yet started.
Freshly rendered samples from the selected set (4-view aesthetic render, pre-Phase-2):
Material caveat: the sample shows a meaningful share of material-less character sculpts. v4 PBR coverage should be measured at Phase-2, not assumed from the sketchfab subset.
ready_v4 (Sketchfab-v1) selected samples. Distribution skews toward characters / figurines / anime sculpts, plus scanned chunks and structures — different mix from the product-heavy v2-sketchfab subset. Note several render as untextured white/clay sculpts: aesthetic-pass does not imply textured, so material coverage is uncertain (do NOT assume the v2-sketchfab ~75%) until Phase-2 dump_pbr runs.
6 · Cross-cutting feedback & open questions
Texture-coverage gap is structural, not under-run. PBR exists only where an asset has a real
material: v2 = 65% (299k/461k), dragged by github's untextured meshes; v3 = 74%; ABO ~100%. Decision needed:
do material-less assets train shape-only, or get dropped from the texture DiT? (Confirmed 24/24 of the
no-PBR github assets genuinely lack a Principled BSDF — they are bare .obj/.fbx geometry.)
Render normalization. A non-trivial fraction of cond renders are dark / small-in-frame. Worth an
exposure + bounding-box-framing pass before tokenization for a cleaner conditioning signal.
TexVerse mesh hygiene. Photogrammetry/scan assets carry jagged silhouettes; ~200 truly-degenerate
voxels were dropped (the @1024 encode SIGSEGV). Acceptable, but flagged for any geometry-quality filter.
v4 next step. 426,521 selected assets await Phase-2 (render → voxelize → encode → caption) — a
~42% bump over v3 shape (1.02M → ~1.45M). But the sample skews to character/anime sculpts with many
untextured ones, so its PBR yield is unknown; measure material coverage at Phase-2 rather than extrapolating
from the sketchfab subset.
Counts from live latent-dir listings; example renders sampled from renders_cond (v2/v3) + fresh aesthetic renders (v4). 2026-06-30.
