← Back to project
← 3D Generation & Editing

3D Data Overview — June 2026 milestone

4 datasets · ~11.4M assets · ~30 TB backed up to R2
11.4M
3D assets
2.4M
4-view rendered
1.8M
VLM scored
~30 TB
R2 backup
v1 · Sketchfab
our scrape — raw GLBs
1M CC-licensed 3D models pulled from the public Sketchfab API. GLB format with embedded textures + metadata. VLM-scored, high quality, ready to use.
1.01M
downloaded GLBs
1.07M
VLM scored
12 TB
on disk
sample sample sample sample sample sample
sourcesketchfab.com (public) format.glb (textures embedded) licenseCC (manifest carries it) statuscompleted
Main viewer VLM filter Filter compare Quality pilot Unified
v1 · GitHub (SWH pilot)
predecessor to v2
First Software Heritage scrape limited to GitHub. 225K textured assets, all rendered + VLM-scored + LAION-aesthetic-rated. Superseded by v2 SWH (full-archive scan).
225K
textured assets
207K
4-view rendered
1.9 TB
on disk
sample sample sample sample sample sample
sourceSWH archive · GitHub only format.fbx .obj .glb mix licenseper GitHub repo statuscompleted (production)
Manifest Textured browser Filter compare
v2 · SWH (full scan)
largest crawl
Full Software Heritage scan across GitHub / GitLab / PyPI / npm. 10M real meshes, 6M with textures. Render stopped at 20%; can resume.
10.0M
real_mesh
6.0M
textured
181K
filter kept
sample sample sample sample sample sample
sourceSWH full archive render progress1.22M / 6M (20%) VLM progress523K / 6M (8.7%) filter kept rate34.7%
5K sample Filter compare
TRELLIS.2 processed
downstream — encoded latents
Pre-encoded training data — sparse-structure / shape / PBR latents from the 3-stage TRELLIS.2 diffusion pipeline. Includes 16-view conditioning renders + voxels + dumps. 398K fully trainable.
398K
trainable (3-latent ∩)
577K
16-view cond renders
18.6 TB
on disk
sample sample sample sample sample sample
6 subsetsOXL_gh/sf + ABO/HSSD/Toys4k + TexVerse vs paper (T.2)54% (398K / 737K) gapTexVerse expansion ongoing statusincremental
No live viewer yet

Live viewers (embedded)

Hosted on the workstation via a Cloudflare quick tunnel — click a viewer to load it in-page. If the iframe is blank the tunnel may be down; full URL below.
Direct URL: —

R2 backup — weikai3dimdata

All ~30 TB uploaded to Cloudflare R2 (WNAM region) · source file counts == R2 object counts · $0 egress
30.9M
files
~30 TB
total size
17/17
dirs verified ✓
~$45
storage / 20 days
R2 prefixfilessize on R2
v1_sketchfab/glbs1,009,17911.57 TiB
v1_sketchfab/scores1,025,672108.8 GiB
v1_sketchfab/{manifest, vlm_results, download_human}~11K~4 GiB
v2_swh/content10,970,2028.12 TiB
v2_swh/textures1,600,0451.25 TiB
v2_swh/converted_fbx117,37562.7 GiB
v2_swh/lfs_pointers1,692,246210 MiB
v2_swh/{vlm_filter, filter_outputs}~10K~570 MiB
trellis2/ObjaverseXL_github6,923,6653.16 TiB
trellis2/ObjaverseXL_sketchfab4,364,3582.82 TiB
trellis2/TexVerse (processed, excl raw+glbs)3,827,7281.85 TiB
trellis2/{ABO, HSSD, Toys4k}418,522379 GiB