TexSet

Textured meshes with three UV parameterisations each: 20,000 objects, every one shipped as its source GLB, three xatlas UV layouts of the same mesh (A, B, C), and the object's albedo baked into each layout as a 1024² atlas. The mesh is never duplicated: the UV sets bind to the GLB's faces by position, and a fingerprint tells you whether your loader produced the mesh they were built against.

Objects	20,000 — train 19,200 · val 400 · test 400
Faces	32,983,516 (per object p50 1,264, max 5,000)
UV layouts	3 per object (xatlas default · straighter/fewer charts · normal-deviation weighted)
Atlases	3 × 1024² RGBA per object, alpha = chart coverage
Shards	34 WebDataset tars (train 32 · val 1 · test 1), 74,690,918,400 bytes (69.6 GiB)
Sources	CanoVerse (Objaverse / Objaverse-XL, canonical pose) 7,774 · TexVerse-1K 12,226

hf download Tome1212/TexSet --repo-type dataset --local-dir TexSet

What is in one sample

Members of one object are adjacent in the tar and share the uid as key:

Member	Contents
<uid>.glb	the source GLB, byte for byte as admitted (see admitted_via for the few repaired TexVerse files)
<uid>.uv.npz	A/uv_vertices, B/uv_vertices, C/uv_vertices — (K,2) uint16, uv = q / 65535; A/uv_faces, B/uv_faces, C/uv_faces — (F,3) int32 indices into that variant's uv_vertices, row i is mesh face i
<uid>.tex_A.png tex_B.png tex_C.png	1024² RGBA; RGB = the GLB's diffuse map baked into that layout, alpha = chart-coverage mask (not transparency)
<uid>.meta.json	loader, fingerprint, n_vertices, n_faces, per-variant charts, utilization, fill, L2_stretch, faces_without_texel, bake_res

A, B, C are three different xatlas layouts of the same object, not three different pictures: one surface point returns the same albedo in all three (measured MAE 1.95/255).

Loading

pip install numpy trimesh pillow
python examples/read_one_sample.py data/val/val-00000-of-00001.tar

uid 007da24c29154f99b1250541f4fb08dc: 1,870 faces, 941 vertices, loader TextureTools.load_whole_mesh+Texture.from_trimesh
  variant A: 1,702 uv vertices, 101 charts, utilization 0.638, faces without texel 0
  tex_A (1024, 1024, 4) uint8, covered texels 60.0%
  mesh from glb: 1,870 faces; fingerprint OK

webdataset and datasets consume the shards directly. manifests/samples.parquet is the queryable index (uid, split, batch, shard, source, category, licence, author, faces, richness, digests).

Four conventions you must know

1. Face binding is positional. uv_faces row i pairs with face i of the mesh as produced by the declared loader. Different loaders weld vertices differently, so meta.json carries fingerprint = sha256(V float32 ‖ F int32) of the mesh the UV sets were built against. examples/read_one_sample.py::load_mesh_like_texset reproduces that loader with plain trimesh (5.1): trimesh.load(process=False), bake the scene graph (Scene.dump), concatenate, then merge_vertices(merge_tex=False, merge_norm=True) twice and merge_vertices(merge_tex=True, merge_norm=True) once. Never load with process=True — it reorders faces and silently mis-textures everything.
2. The atlas alpha is the coverage mask. Use it as the weight when downsampling; a plain box filter mixes the empty gutter into island borders (measured p95 error 22.6/255 vs 1.50/255 for the masked average).
3. Row 0 of tex_*.png is v = 0 — no flip. Texel lookup is col = u·W, row = v·H.
4. baseColorFactor is NOT applied. The atlas is the GLB's diffuse map in its own colour. This is deliberate: 69.9 % of the CanoVerse objects carry a factor of exactly 0.40 that belongs to that supplier's pipeline, not to the object (TexVerse's independent copy of the same object carries 1.00). Multiply it in yourself if your renderer needs glTF semantics. Texture sampling outside [0,1] follows the glTF sampler (REPEAT tiles; CLAMP_TO_EDGE clamps).

How the 20,000 were chosen

Two batches, marked by batch in samples.parquet:

• s21_10k (10,000): a frozen research set, a seeded uniform sample of the 137,529-object core, shipped unchanged with its own 9,600 / 200 / 200 split.
• extra_10k (10,000): chosen for texture richness and diversity relative to s21_10k. Candidates were the core minus s21_10k, minus four quality screens (texel-starved, overlapping texels, tiling residual, planar billboards), restricted to licence-clean objects. A richness gate kept objects above the s21_10k median on a composite of colour entropy, hue entropy, gradient energy, luminance spread and dominant-colour share (plus ≥ 100 distinct colours, no colour over 50 % of the texels). From those, a k-center greedy over [RGB histogram, texture statistics, mesh statistics] seeded with s21_10k picked the object farthest from everything already in the set, 10,000 times, capped at 50 objects per uploader and 1,000 per category. Split 9,600 / 200 / 200, seed 3407.

Result: richness composite mean 0.635 vs 0.470, near-black atlases 1 vs 100, 6,831 distinct uploaders vs 5,527, category entropy 6.25 vs 5.68 bits, face-count distribution matched.

Repository layout

data/train/train-000{00..31}-of-00032.tar       600 objects each
data/val/val-00000-of-00001.tar                 400 objects
data/test/test-00000-of-00001.tar               400 objects
manifests/samples.parquet     one row per object: identity, split, batch, shard, source, category,
                              licence, author, faces, UV metrics, richness, per-member sha256
manifests/shards.json         shard -> counts, bytes, sha256
manifests/release.json        version, counts, member list
manifests/{train,val,test}.txt
manifests/SHA256SUMS
examples/read_one_sample.py   dependency-light reader that also verifies the fingerprint

Provenance and terms

Source objects come from Objaverse / Objaverse-XL via the CanoVerse canonical-pose release, and from TexVerse-1K (YiboZhang2001/TexVerse-1K). Both collections are ODC-By 1.0; each object remains under its author's licence, recorded per object in source_license / licence_class with source_author for attribution. extra_10k is 100 % licence-clean; s21_10k is shipped as frozen and contains 1,326 objects with no recorded licence, 574 NonCommercial, 22 NoDerivs and 66 under other licences. Read LICENSE_OR_TERMS.md before redistributing or using this data commercially.

Known limitations

• CanoVerse objects are canonically posed; TexVerse objects keep their original Objaverse pose (source and canoverse_pose_class tell them apart).
• 2 objects are pruned (pruned = true, fingerprint = null): faces no camera could reach were deleted from the GLB after baking, uv_faces rows were dropped to match, and the atlas was left as is, so the texels of the deleted faces belong to no face. One of them (99b2fde7…) lost its embedded texture in the re-export; its atlas still carries the albedo.
• Faces smaller than a texel get no texel: 2.02 % of all faces in the core; faces_without_texel per variant in meta.json. On regular flat-panelled geometry the three layouts can collapse to the same segmentation.
• The admitted_via column names 7 TexVerse objects shipped with a KHR_texture_transform baked into the image and 73 whose CLAMP_TO_EDGE sampler was honoured in the bake; their GLB differs from the upstream file and source_sha256 records the upstream digest.