Splataverse: cross-dataset splits

This repo holds only split manifests. The Gaussian splats themselves live in the per-source repos, each under its own licence. Every row of a manifest names the repo (hf_repo) and file (path) it comes from.

Where the data comes from

source	repo	licence / notes
Sketchfab (Objaverse)	maxwildersmith/splataverse-sketchfab	cc-by / cc-by-sa / cc-by-nc(-sa), per object (license column); quality-filtered
Smithsonian 3D	maxwildersmith/splataverse-smithsonian	CC0; ~88% of val/test are bone specimens
TableVerse (tabletop scenes)	maxwildersmith/splataverse-tableverse	licence unknown; identical duplicate instances are ranked last
GitHub, textured (Objaverse-XL)	maxwildersmith/splataverse-github	per-repo open licence (MIT, Apache-2.0, GPL, BSD, CC0, ...); kept objects only
GitHub, untextured fallback (Objaverse-XL)	maxwildersmith/splataverse-github-untextured	same licensed repos, but geometry-only (grey) objects; training only, always the last pool to be used

Split v5

splits/v5.parquet (config v5) has one row per object: uid, hf_source, hf_repo, path, license, split, tier, is_bone, group, reasons. split is val, test, train (in a tier), train_unused, flagged (failed the splat quality check; reasons says why) or unscored.

• Training tiers are nested (1k is inside 5k is inside 10k ...): tier T is split == "train" and tier <= T. Small tiers are Sketchfab plus ~10% Smithsonian; TableVerse and GitHub ramp in from 50k up (GitHub 5% at 50k, 8% at 100k, 12% at 250k, 20% at 1M; TableVerse 5% at 100k, 10% at 250k, 30% at 1M). Sketchfab fills the remainder, and once its clean pool is used up the other sources top the tier up.
• The untextured GitHub fallback is used only after every other pool is exhausted, so it appears only in train_all. Filter on hf_source to drop it.
• val / test are fixed. Sketchfab 1,000 + 1,000 (licence-stratified), Smithsonian's frozen 150 + 150, TableVerse whole scenes (~600 each, so a scene never straddles train and test), and a seeded random 500 + 500 from textured GitHub (not yet grouped by repo, so forks can make GitHub a little optimistic). The untextured fallback is never in val/test.

list	objects	Sketchfab	Smithsonian 3D	TableVerse	GitHub	GitHub, untextured fallback
val	2,257	1,000	150	607	500	-
test	2,255	1,000	150	605	500	-
train_1k	1,000	900	100	-	-	-
train_5k	5,000	4,500	500	-	-	-
train_10k	10,000	9,000	1,000	-	-	-
train_25k	25,000	22,908	2,092	-	-	-
train_50k	50,000	45,408	2,092	-	2,500	-
train_100k	100,000	84,908	2,092	5,000	8,000	-
train_250k	250,000	192,908	2,092	25,000	30,000	-
train_1m	1,000,000	459,890	2,092	450,478	87,540	-
train_all	1,924,723	459,890	2,092	966,284	87,540	408,917

The 2M tier is not fillable from licensed data: the training pool is 1,924,723 objects, so train_all is the largest list. Unlicensed GitHub is deliberately not included.

splits/v5/<list>.txt has one <repo> <path> line per object. The previous split (v4, Sketchfab + Smithsonian only) is kept under splits/v4/. ArtVIP and HOI are not in v5 yet. TableVerse paths may be <scene>/objects.tar::<member> (bundled in a tar).

Conversion benchmark (conversion_v1)

A fixed set of source meshes for evaluating mesh -> 3D Gaussian splat conversion (not our splats: you convert the original mesh with your method and score it). splits/conversion_v1/objects.parquet (config conversion_v1) has one row per object: uid, tier, source, hf_repo, path, license, source_url, sha256, file_type, is_bone, note; source_url and sha256 are the objaverse-xl manifest entries of the original mesh, hf_repo/path point to our converted splat for reference.

• bench: the exact objects of the Splataverse paper's conversion table (Sketchfab, Smithsonian scans, the Khronos SciFiHelmet). 1 object can no longer be downloaded from the source (957028a3-c370-5c5d-9615-962c8ec5ae37: removed at source (Smithsonian API 404, 2026-10-26)), so the paper reports 20 of 21; it stays listed with a note (also where it appears in full). Cameras ship in splits/conversion_v1/cameras.json.
• full: a seeded, stratified set drawn from the v5 test lists: Sketchfab (25 per licence), Smithsonian (25 bone + 17 other: all non-bone test objects, the test list is mostly bones), textured GitHub (100), and untextured GitHub geometry (50 .stl + 50 other). The untextured pool has no val/test in v5, so these 100 are also in v5's train_all fallback pool; exclude them (hf_source == "github_untextured" and uid in this list) if you train on train_all and evaluate here.

tier	objects	github	github_untextured	khronos	sketchfab	smithsonian
bench	21	-	-	1	10	10
full	342	100	100	-	100	42

Protocol (paper, Evaluation protocol): ground truth is the textured mesh rendered unlit (emission = base colour) by Blender Cycles (16 spp, transparent film) from 16 held-out views (8 azimuths x elevations 15/45 deg, 512 px, 40 deg FoV, bounding sphere fitted with 10% margin); every method is rendered by the same gsplat rasterizer (antialiased, 4x4 supersampled, black background). Ranked on LPIPS (AlexNet), LPIPS after gain matching, structure and gradient correlation over the object mask; masked PSNR is reported but not ranked on; geometry is the Chamfer distance between Gaussian centres and 100k area samples (% of bbox diagonal).

pip install huggingface_hub objaverse trimesh pandas torch gsplat lpips plyfile pillow tyro
python eval_conversion.py --stage fetch --tier bench --out-dir conv_eval          # source meshes + cameras
blender -b -P eval_conversion.py -- gt --out-dir conv_eval                        # unlit GT renders
python eval_conversion.py --stage score --out-dir conv_eval --splats my_method    # my_method/<uid>.ply|.npz

Download a split

pip install huggingface_hub tyro
python fetch_split.py --split train_5k --out-dir splataverse_data

fetch_split.py (in this repo) skips files already downloaded and pauses whenever the Hugging Face API rate limit (1,000 requests per 5 minutes) is reached, so large tiers can be fetched unattended. Use --version v4 for the old split.

Complete split and protocol catalog

The split catalog documents scientific partitions, nested tiers, historical versions, duplicate exclusions, the matched 600-object training and160-candidate paired-evaluation lists, the 1,119-object confirmatory test, retrieval galleries, and the new frozen balanced-profile cohorts. Validation checks and artifact hashes make these metadata additions auditable. Existing v4/v5 assignments are preserved.

Loader distinction: load_dataset(..., "v5")["train"] loads the metadata table. Select rows with split == "train" and the desired tier for scientific training; its val and test rows are evaluation data. UID-only lists and repo path lists are different formats.

Additional sources: ArtVIP, HOI, cities, ortho/Berlin, SimReady, and the private Tartan release have source-specific manifests and access rules; they are not included in v5 train/validation/test. Source-level inventory or conversion completion is not a benchmark split assignment. Workshop leaderboard splits will be versioned separately once frozen.

New conversion jobs use balanced by default. Existing split manifests continue pointing to the original payloads; they must not be relabeled as balanced measurements. See the catalog for channel/provenance availability and the limitations of historical duplicate controls.