[00] · AI-3D READINESS CORPUS

64 AI-generated models. Every result published.

64 GLBs — 8 open generator families × 8 shared commerce prompts, so every family was handed the same job. Each one scored for Apple-AR readiness, run through one spatialpack autocorrect pass, then re-scored with the same gate. The per-asset table below is the whole corpus — nothing sampled, nothing dropped.

This is the corpus behind the “64/64” and “25/25” figures on the homepage and on Catalog Rescue. It is not the size leaderboard: that one runs the Khronos sample corpus against gltf-transform and gltfpack and lives at /benchmark. Different corpus, different question.

corpus assets
64
AR-ready after autocorrect
64/64
failing ones rescued
25/25
corpus bytes
-81%

Run 2026-06-11 · CLI 0.1.0 (git 89ce61f) · target apple-ar · preset web-mobile (autocorrect default) · 124.0 MB 23.5 MB · 0 of 64 outputs grew.

[01] · BEFORE AND

Before and after.

PopulationAssetsReadyWarningsNot ready
AI-generated, before autocorrect6439 (60.9%)250
AI-generated, after autocorrect6464 (100.0%)00
Khronos sample corpus, same gate (2026-06-10)10338 (36.9%)3431

The honest read: AI generators emit structurally simple PBR, so they clear a gate 31 of 103 human-authored Khronos showcase assets fail. Zero of the 64 AI models were not-ready before repair — there is no “broken AI slop rescued” story here. What the run shows is automatic warning cleanup plus a large, visually-lossless size cut.

swipe the table →
[02] · WHAT ACTUALLY

What actually failed.

IssueSeverityAssetsRemaining after
conformance:geometry.uv_island_spacingwarning250

No lossy material extensions (transmission, volume, clearcoat, iridescence) appear anywhere in this corpus — which is why nothing here was blocked outright, only warned on.

swipe the table →
[03] · PER GENERATOR

Per generator family.

FamilyGeneratorBefore R/W/NRAfter readyRescuedMedian size ΔFamily bytes
TRELLISTRELLIS (Microsoft)0/8/08/88/889.9%13.8 MB1.6 MB (-88.4%)
Hunyuan3D-2Hunyuan3D-2 (Tencent)0/8/08/88/885.8%16.5 MB2.4 MB (-85.7%)
3DTopia-XL3DTopia-XL (NTU S-Lab)0/8/08/88/885.3%23.5 MB3.5 MB (-85.2%)
Real3DReal3D (UT Austin)8/0/08/8n/a79.4%30.4 MB6.2 MB (-79.6%)
TripoSRTripoSR (Stability AI / Tripo)8/0/08/8n/a78.4%9.4 MB2.0 MB (-78.8%)
InstantMeshInstantMesh (TencentARC)8/0/08/8n/a77.3%13.2 MB3.0 MB (-77.4%)
SF3DStable Fast 3D (Stability AI)7/1/08/81/172.1%7.2 MB2.0 MB (-72.5%)
SPAR3DSPAR3D / Stable Point Aware 3D (Stability AI)8/0/08/8n/a71.6%10.0 MB2.9 MB (-71%)

R/W/NR = ready / warnings / not-ready before autocorrect. “Rescued” is n/a for families that had nothing to rescue. Median size Δ is the median per-asset output-vs-input byte change.

swipe the table →
[04] · EVERY

Every asset.

AssetTrisBeforeAfterInOutΔ bytes
TRELLIS/a_hammer2,798warningsready1146 kB49 kB−95.8%
TRELLIS/chair9,122warningsready1412 kB91 kB−93.5%
TRELLIS/an_avocado20,622warningsready1431 kB149 kB−89.6%
TRELLIS/a_coffee_table15,248warningsready1556 kB131 kB−91.6%
TRELLIS/a_coffee_maker70,575warningsready2646 kB446 kB−83.1%
TRELLIS/a_black_baseball_hat23,944warningsready1456 kB142 kB−90.2%
TRELLIS/black_hi-top_sneakers34,577warningsready1880 kB255 kB−86.5%
TRELLIS/an_antique_chest_with_three_drawers41,844warningsready2238 kB340 kB−84.8%
Hunyuan3D-2/a_hammer40,000warningsready1993 kB260 kB−87%
Hunyuan3D-2/chair40,000warningsready2095 kB296 kB−85.9%
Hunyuan3D-2/an_avocado40,000warningsready1960 kB266 kB−86.4%
Hunyuan3D-2/a_coffee_table40,000warningsready2107 kB289 kB−86.3%
Hunyuan3D-2/a_coffee_maker40,000warningsready1906 kB289 kB−84.8%
Hunyuan3D-2/a_black_baseball_hat40,000warningsready1930 kB291 kB−84.9%
Hunyuan3D-2/black_hi-top_sneakers40,000warningsready2256 kB351 kB−84.5%
Hunyuan3D-2/an_antique_chest_with_three_drawers40,000warningsready2256 kB323 kB−85.7%
TripoSR/a_hammer26,939readyready540 kB120 kB−77.7%
TripoSR/chair28,134readyready564 kB122 kB−78.4%
TripoSR/an_avocado49,979readyready1001 kB229 kB−77.1%
TripoSR/a_coffee_table42,123readyready844 kB180 kB−78.7%
TripoSR/a_coffee_maker75,895readyready1520 kB295 kB−80.6%
TripoSR/a_black_baseball_hat70,360readyready1409 kB281 kB−80%
TripoSR/black_hi-top_sneakers72,631readyready1455 kB316 kB−78.3%
TripoSR/an_antique_chest_with_three_drawers101,163readyready2026 kB442 kB−78.2%
InstantMesh/a_hammer25,320readyready508 kB123 kB−75.8%
InstantMesh/chair45,836readyready918 kB207 kB−77.4%
InstantMesh/an_avocado64,044readyready1283 kB299 kB−76.7%
InstantMesh/a_coffee_table73,556readyready1472 kB326 kB−77.8%
InstantMesh/a_coffee_maker111,852readyready2238 kB496 kB−77.8%
InstantMesh/a_black_baseball_hat91,360readyready1829 kB389 kB−78.7%
InstantMesh/black_hi-top_sneakers122,176readyready2445 kB556 kB−77.2%
InstantMesh/an_antique_chest_with_three_drawers126,844readyready2539 kB590 kB−76.8%
SF3D/a_hammer11,208readyready493 kB155 kB−68.7%
SF3D/chair14,488readyready679 kB216 kB−68.2%
SF3D/an_avocado16,760readyready658 kB191 kB−71%
SF3D/a_coffee_table26,300readyready975 kB259 kB−73.4%
SF3D/a_coffee_maker26,572warningsready920 kB231 kB−74.8%
SF3D/a_black_baseball_hat33,340readyready1135 kB276 kB−75.6%
SF3D/black_hi-top_sneakers28,828readyready1090 kB304 kB−72.1%
SF3D/an_antique_chest_with_three_drawers35,464readyready1292 kB360 kB−72.1%
SPAR3D/a_hammer12,096readyready568 kB188 kB−66.9%
SPAR3D/chair16,276readyready765 kB252 kB−67.1%
SPAR3D/an_avocado18,376readyready814 kB273 kB−66.5%
SPAR3D/a_coffee_table32,632readyready1262 kB359 kB−71.5%
SPAR3D/a_coffee_maker30,984readyready1204 kB342 kB−71.6%
SPAR3D/a_black_baseball_hat49,232readyready1814 kB477 kB−73.7%
SPAR3D/black_hi-top_sneakers39,372readyready1491 kB423 kB−71.7%
SPAR3D/an_antique_chest_with_three_drawers55,712readyready2090 kB594 kB−71.6%
Real3D/a_hammer74,553readyready1493 kB321 kB−78.5%
Real3D/chair97,211readyready1946 kB398 kB−79.5%
Real3D/an_avocado137,025readyready2743 kB584 kB−78.7%
Real3D/a_coffee_table197,524readyready3954 kB781 kB−80.3%
Real3D/a_coffee_maker209,170readyready4187 kB849 kB−79.7%
Real3D/a_black_baseball_hat245,723readyready4918 kB936 kB−81%
Real3D/black_hi-top_sneakers275,167readyready5509 kB1167 kB−78.8%
Real3D/an_antique_chest_with_three_drawers284,422readyready5694 kB1182 kB−79.2%
3DTopia-XL/a_hammer22,054warningsready1905 kB211 kB−88.9%
3DTopia-XL/chair26,368warningsready1994 kB262 kB−86.8%
3DTopia-XL/an_avocado34,594warningsready2100 kB307 kB−85.4%
3DTopia-XL/a_coffee_table44,884warningsready2880 kB391 kB−86.4%
3DTopia-XL/a_coffee_maker60,388warningsready2832 kB446 kB−84.2%
3DTopia-XL/a_black_baseball_hat72,550warningsready3740 kB585 kB−84.4%
3DTopia-XL/black_hi-top_sneakers100,000warningsready4609 kB769 kB−83.3%
3DTopia-XL/an_antique_chest_with_three_drawers65,054warningsready3404 kB502 kB−85.2%

Δ bytes is output vs input for the whole GLB. 0 assets grew; 0 size-guard trips (the guard ships the original bytes when a repair would grow a file without improving the verdict).

swipe the table →
[05] · METHODOLOGY

Methodology, precisely.

Corpus provenance

64 GLBs ingested 2026-06-10 from the 3d-arena/3d-arena Hugging Face dataset (MIT license, 3D Arena (maintainer: Dylan Ebert)). Same input prompt/image set across all generator families (paired design). Inputs sourced from dylanebert/iso3d. Every asset carries source URL, dataset license, generator repo and weights license in the corpus manifest. Commercial generators (Meshy, Tripo API, Rodin, Luma) were excluded because their redistribution rights could not be verified.

prompts: a black baseball hat · a coffee maker · a coffee table · a hammer · an antique chest with three drawers · an avocado · black hi-top sneakers · chair

Generators
  • TRELLIS TRELLIS (Microsoft); weights MIT
  • Hunyuan3D-2 Hunyuan3D-2 (Tencent); weights Tencent Hunyuan Community License
  • 3DTopia-XL 3DTopia-XL (NTU S-Lab); weights S-Lab License 1.0
  • Real3D Real3D (UT Austin); weights MIT
  • TripoSR TripoSR (Stability AI / Tripo); weights MIT
  • InstantMesh InstantMesh (TencentARC); weights Apache-2.0
  • SF3D Stable Fast 3D (Stability AI); weights Stability Community License
  • SPAR3D SPAR3D / Stable Point Aware 3D (Stability AI); weights Stability Community License
Gate

spatialpack readiness --target apple-ar before, then spatialpack autocorrect (preset web-mobile (autocorrect default)), which re-scores the repaired GLB. Same scorecard both times — conformance checks plus AR-parity self-consistency.

Artifact verification

4 repaired GLBs (TRELLIS/a_coffee_maker, Hunyuan3D-2/chair, TripoSR/chair, 3DTopia-XL/a_black_baseball_hat) were re-parsed and rendered before-vs-after across 6 camera views each — 24 views total — with the SpatialPack visual-diff harness (model-viewer + headless Chromium). Worst case across all 24 views: SSIM 0.9886, mean ΔE94 0.63, MAE 0.0026. No visible change from the repair.

What this evidence does NOT show
  • — Commercial generators (Meshy, Tripo API, Rodin, Luma). Excluded for license verifiability, not because they did worse.
  • — Apple AR Quick Look ground truth. “Ready” is our scorecard (conformance + AR-parity self-consistency), not a RealityKit render on a device.
  • — Generality beyond 8 prompts and one arena snapshot per family.
  • — A rescue-from-broken story. Zero assets were not-ready before repair; the 25/25 figure is warnings cleared, not blockers fixed.
Prior run (superseded, kept on the record)

The first run of this corpus (2026-06-10, pre-fix toolchain) produced the same readiness verdicts but only −29.8% corpus bytes, because autocorrect ran a normal-generation pass that un-welded meshes and inflated the three vertex-colour families by 13–21%. That was a real regression, it was filed against ourselves, and the fix (drop the pass; add a keep-smaller-of size guard) is what the numbers above measure.

swipe the table →
Have generated models of your own?

The free check runs the same readiness gate used on all 64 assets above — no account, no card.

Check a model — free