TopoObjaverse

Texture supervision on Objaverse topology: 76,278 objects, each delivered as three UV parameterisations of the same mesh, per-texel face and barycentric queries, tri-state visibility against four rendered condition views, and the ground-truth texture those queries address.

This is the complete dataset. One set, one download.

Objects	76,278 — train 75,278 · validation 500 · test 500
Faces	121,720,350
UV query rasters	3 families × 256² per object
Condition views	4 × 512² RGBA per object, each with its camera
Shards	32 WebDataset tars (train 30 · validation 1 · test 1), 67,912,366,080 bytes
Split identity	TOPOTEX_SPLIT_V6 (frozen 2026-08-24, seed splitv6-20260824)

# the whole dataset
hf download Tome1212/TopoObjaverse --repo-type dataset --local-dir TopoObjaverse

# or one split
hf download Tome1212/TopoObjaverse --repo-type dataset \
    --include "data/validation/*" "manifests/*" "examples/*" --local-dir TopoObjaverse

What is in one sample

Every sample is a group of tar members sharing the object id as key. Tensor and metadata members are zstd-compressed inside the tar; the PNGs are stored raw.

Member	Contents
mesh.safetensors	vertices [V,3] float32 (Y up, right-handed, centred on the vertex centroid and scaled so the farthest vertex is at distance exactly 1 — a unit sphere, not a unit cube), faces [F,3] int32, the object's own uv_vertices/uv_faces, global_scale [1] = sqrt(total surface area) (the length unit graph_rel's edge lengths are divided by, not the canonicalisation factor), and a face-adjacency graph: graph_edges [E,2], graph_rel [E,3], graph_boundary [F] (fraction of a face's edges with no neighbour)
queries.safetensors	for each family in xatlas, smart_uv, partial: {fam}_face_id [256,256] int32 — int64 in some objects, so cast on read — (−1 invalid), {fam}_barycentric [3,256,256] float16, {fam}_gt_texture [256,256,3] uint8 sRGB, {fam}_gt_alpha [256,256] uint8, {fam}_valid_mask [256,256], {fam}_uv_vertices, {fam}_uv_faces
visibility.safetensors	per family {fam}_vis [4,256,256] uint8 tri-state (0 invalid, 1 hidden, 2 visible) and {fam}_count [256,256] uint8, the number of views that see each texel
view_000.png … view_003.png	512² RGBA condition renders, sRGB, transparent background
meta.json	uid; each view's camera (azimuth, elevation, focal_mm, distance, shift_x, shift_y) and its statistics; file digests; the errata trail; and a topoobjaverse block naming the schema, pipeline, render-spec and split versions plus the release and errata cutoff

Three parameterisations of the same mesh make UV layout an input variable rather than a fixed assumption: xatlas is a standard automatic atlas, smart_uv is Blender's angle-based unwrap, and partial is a deterministic partial atlas seeded from the object id. Visibility is tri-state rather than boolean because "the model could not have seen this texel" and "the model saw it and got it wrong" are different training signals.

Loading

Nothing beyond safetensors and zstandard is required: the shards are ordinary WebDataset tars, and examples/read_one_sample.py reads one with the standard library plus those two packages.

pip install safetensors zstandard
python examples/read_one_sample.py data/validation/validation-00000-of-00001.tar

uid 0004b0...  faces 1,204  vertices 812
  xatlas    texture (256, 256, 3) valid 68.2%
  visibility (4, 256, 256)  visible texels per view [17087, 17091, 12662, 17554]
  views ['view_000.png', 'view_001.png', 'view_002.png', 'view_003.png']

webdataset and datasets consume the shards directly as well. manifests/samples.parquet is the queryable index (uid, split, faces, valid texels, size, shard, digests, source provenance) and manifests/shards.json maps a uid to the shard holding it. The typed loader, the validators and the packer live in the code repository:

from topoobjaverse.loader import WebDatasetBackend
ds = WebDatasetBackend(repo_id="Tome1212/TopoObjaverse")
for s in ds.iter_samples("validation"):
    print(s.uid, s.n_faces, s.queries["xatlas_gt_texture"].shape)
    break

Repository layout

data/train/train-000{00..29}-of-00030.tar
data/validation/validation-00000-of-00001.tar
data/test/test-00000-of-00001.tar
manifests/samples.parquet     one row per sample: identity, split, size, shard, digests, provenance
manifests/shards.parquet      shard -> counts, bytes, sha256
manifests/shards.json         the index the loader reads
manifests/release.json        versions, build commit, config digests, counts
manifests/{train,validation,test}.txt
manifests/SHA256SUMS
examples/                     dependency-light readers

Provenance

Source meshes come from Objaverse and Objaverse-XL (Allen Institute for AI), canonicalised with the CanoVerse rotation annotations; environment maps are from Poly Haven; rendering is Blender/Cycles. Each sample records its own file digests and errata trail, and manifests/samples.parquet carries the per-object source fields (source_dataset, source_license, source_author, source_author_display, source_name, source_identifier) as reference provenance. Version identifiers on every sample and in manifests/release.json: schema_version topoobjaverse.sample/1, pipeline_version topoobjaverse.pipeline/1, render_spec_version canoverse.render/1, split_version TOPOTEX_SPLIT_V6, release_version, errata_cutoff 2026-09-04.

Terms

The source objects come from Objaverse and Objaverse-XL and remain the property of their authors under the licences they chose; the collections themselves are ODC-By 1.0. Per-object licence fields ship in the manifest for reference. See LICENSE_OR_TERMS.md before redistributing or using this data commercially — in particular, the ground-truth textures of some objects were produced with FLUX.1-dev, whose licence makes model outputs non-commercial. The construction code is MIT-licensed in the GitHub repository and says nothing about the data.

Known limitations

Measured properties of this frozen release, none of them worked around at load time:

• Baked lighting. Condition views are rendered under a fixed diffuse protocol; ground-truth textures are base colour. Shading already present in a source texture stays in it.
• Low-information views. Flat or plate-like objects can produce views that show very little of the atlas. Only the worst bands were re-cameraed; roughly 185,000 views scoring above 0.25 on info_frac were left as rendered.
• Flat-colour textures. 13,766 objects are painted in one or a few flat colours and carry almost no high-frequency signal.
• Anti-aliased ground-truth alpha. gt_alpha is a bake, so island borders carry intermediate values even for OPAQUE materials: of 200 sampled objects, 136 are alpha-255 everywhere and 141 strictly binary, the rest carrying intermediate values on up to 42% of their valid texels. valid_mask, not alpha, is the authority on which texels are supervised.
• An unquantified rainbow-gradient texture family. Its detector was never calibrated, so those objects are neither counted nor excluded.
• Repaired evaluation objects. 408 objects in validation and test were repaired by earlier campaigns, so metrics from older internal baselines are not comparable. Rebuild the evaluation baseline and train from scratch.
• Source-mesh quality. Objaverse meshes vary. Degenerate triangles, mirrored UV islands and interpenetrating geometry occur; they are documented, not repaired.

Citation

@misc{topoobjaverse2026,
  title  = {TopoObjaverse: texture supervision on Objaverse topology},
  author = {{TopoObjaverse maintainers}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Tome1212/TopoObjaverse}
}

Please also cite Objaverse and, where applicable, Objaverse-XL.