Read date for every source below: 2026-09-16. Raw page copies were kept only for the reading session; the measurement outputs and the GitHub join script are beside this file. "L<n>" = line number in the pinned file. Status: COMPLETE (all 8 candidates + measured per-object licence mixes + summary + gaps).
Verdict vocabulary: explicit-training-prohibition | non-commercial-terms | mixed-per-object | permissive | third-party-copyright-unlicensed | unknown.
1. TRELLIS.2-4B#
Sources read:
- README @ commit: https://raw.githubusercontent.com/microsoft/TRELLIS.2/75fbf0183001ed9876c8dbb35de6b68552ee08bd/README.md
- data_toolkit README @ same commit: https://raw.githubusercontent.com/microsoft/TRELLIS.2/75fbf0183001ed9876c8dbb35de6b68552ee08bd/data_toolkit/README.md
- Model card @ af44b45f2e35a493886929c6d786e563ec68364d (HF API sha of main on read date = this pin): https://huggingface.co/microsoft/TRELLIS.2-4B/raw/main/README.md
- Paper (README L5 names it): https://arxiv.org/abs/2512.14692v1 (only version; HTML https://arxiv.org/html/2512.14692v1)
A. Training datasets as stated#
- Paper §4 "Implementation details": "Our SC-VAE is trained using the Trellis-500K setup [65] after filtering out assets without PBR materials, resulting in a curated collection from Objaverse-XL [12], ABO [9], and HSSD [28]." and "The generative model is trained on an extended collection of about 800K assets, augmented with TexVerse [77] to enrich PBR diversity and realism."
- Paper Appendix C, Table 6 "Composition of the training set and evaluation set" (Source / Shape available / Material available): TexVerse 503387 / 382996; ObjaverseXL (sketchfab) 168307 / 141623; ObjaverseXL (github) 293887 / 202188; ABO 4485 / 4485; HSSD 6670 / 6670; SC-VAE training set 473349 / 354966; All training set 976736 / 737962; Toys4k (evaluation set) 3229 / 2282.
- Appendix C: "We begin by curating a collection of 3D assets but exclude the 3D-FUTURE [17] dataset due to its lack of Physically-Based Rendering (PBR) materials." → 3D-FUTURE is NOT in TRELLIS.2 training (unlike TRELLIS).
- Eval-only: Toys4k and a 90-asset "Sketchfab Featured" test set ("Both test sets are unseen during training.").
- data_toolkit/README.md L22: "
SUBSET: Target dataset subset. Options:ObjaverseXL,ABO,HSSD,TexVerse(Training sets);SketchfabPicked,Toys4k(Test sets)." L24: ObjaverseXLSOURCE"Options:sketchfab,github." - Completeness: Table 6 is an enumerated closed list (5 sources). Note
inconsistency: §4 says "about 800K", Table 6 says 976,736 total. The loader
modules the toolkit imports
(
importlib.import_module(f'datasets.{sys.argv[1]}')) are NOT present in the TRELLIS.2 tree at the pin (tree listing has nodata_toolkit/datasets/), so the exact per-object ID lists for TRELLIS.2 (esp. TexVerse subset) are not published at this commit. - Model card: no dataset list; only "the outputs reflect the distribution of the training data" (L64) and "The code and dataset are publicly released to facilitate reproduction and further research." (L139).
B. Per-object licence handling (this candidate)#
- Not measurable: TRELLIS.2 ObjaverseXL (github) = 293,887 ≠ TRELLIS-500K's
311,843, so the TRELLIS-500K GitHub list is not the same set; no TRELLIS.2
list published. For reference only, TRELLIS-500K's GitHub list is 87.15%
license=None (candidate 2 §B). - Measured: TRELLIS.2 Table 6 gives ObjaverseXL (sketchfab) = 168,307, the same count as TRELLIS-500K; TRELLIS.2 publishes no ID list, so identity is not verifiable. If identical, the TRELLIS-500K measurement applies: 19,121 NC-family objects (11.36%) (see candidate 2 §B). TexVerse subset (503,387) cannot be measured (no list published).
- Filters stated in paper: PBR-material filter, "we filter the assets based on an aesthetic score ... Objects with an estimated aesthetic score below 4.5 are excluded from the training set." No licence filter stated. Word "licen*" appears in the paper HTML only in arXiv's own "License: arXiv.org perpetual non-exclusive license" banner.
- Paper says the SC-VAE uses "the Trellis-500K setup" for Objaverse-XL/ABO/HSSD (TRELLIS-500K itself has no licence column and no licence filter — candidate 2 §B); Table 6 counts match TRELLIS-500K for sketchfab/ABO/HSSD but not for github (293,887 vs 311,843).
C. Dataset terms (full clause quotes in candidate 2 §C for Objaverse-XL, ABO, HSSD, Toys4k)#
- TexVerse —
https://huggingface.co/datasets/YiboZhang2001/TexVerse/raw/8166bb11d24c03fb2c34d0ee099d754ce492bb54/README.md
(HF sha on read date; lastModified 2026-09-12; paper arXiv 2508.10868 per
TRELLIS.2 ref [77]). Collection licence L2
license: odc-by. L40: "TexVerse dataset has 858,669 unique 3D models curated from Sketchfab". L82: "Individual objects in TexVerse are all licensed as creative commons distributable objects. The metadata will provide the license for each object." L84–90 mix: CC BY 699,075; CC BY-NC-SA 88,710; CC BY-SA 14,624; CC BY-NC 31,514; CC BY-NC-ND 11,599; CC BY-ND 7,008; CC0 6,139. (i) commercial: NC objects forbid; (ii) derivatives: ND objects forbid sharing adaptations; (iii) no training clause found in the card. - Objaverse-XL (sketchfab + github), ABO, HSSD: see candidate 2 §C.
D. Verdicts (TRELLIS.2-4B)#
- TexVerse →
mixed-per-object(131,823 NC-family [BY-NC-SA 88,710 + BY-NC 31,514 + BY-NC-ND 11,599] and 18,607 ND-family [BY-NC-ND 11,599 + BY-ND 7,008] objects in the 858,669-object collection per the card's counts; TRELLIS.2 states no licence filter). - Objaverse-XL (sketchfab) →
mixed-per-object. - Objaverse-XL (github) →
mixed-per-object(Objaverse-XL: "Individual objects in Objaverse-XL are licensed under different licenses"; no licence filter by TRELLIS.2). - ABO →
permissive(CC BY 4.0). - HSSD →
non-commercial-terms(CC BY-NC 4.0). - Toys4k (evaluation only, not training) →
mixed-per-object.
2. TRELLIS image-large / text-large / text-xlarge#
Sources read:
- README @ 442aa1e1afb9014e80681d3bf604e8d728a86ee7: https://raw.githubusercontent.com/microsoft/TRELLIS/442aa1e1afb9014e80681d3bf604e8d728a86ee7/README.md
- DATASET.md @ same: https://raw.githubusercontent.com/microsoft/TRELLIS/442aa1e1afb9014e80681d3bf604e8d728a86ee7/DATASET.md
- Loader: https://raw.githubusercontent.com/microsoft/TRELLIS/442aa1e1afb9014e80681d3bf604e8d728a86ee7/dataset_toolkits/datasets/ObjaverseXL.py (and ABO.py, HSSD.py, 3D-FUTURE.py, Toys4k.py)
- Paper: https://arxiv.org/abs/2412.01506v3 (v3 = 30 May 2025, latest; HTML read)
- Dataset card: https://huggingface.co/datasets/JeffreyXiang/TRELLIS-500K/raw/288b06111bc4d06800293ab087989f852a632af2/README.md (sha on read date; lastModified 2024-12-26)
- Model cards (sha on read date): TRELLIS-image-large 25e0d31ffbebe4b5a97464dd851910efc3002d96, TRELLIS-text-large 4aad9f4a110329a410974d7f41ce5333a9a1fc87, TRELLIS-text-xlarge e0b00432b8e3a8ecee0df806ab1df9f7281f2be4 — each 16 lines, only "introduced in the paper [Structured 3D Latents ...]"; no training-data statement in any of the three cards, and none says the image vs text models used different subsets.
A. Training datasets as stated#
- Paper §4: "For training, we carefully collect approximately 500K high-quality 3D assets from 4 public datasets: Objaverse (XL) [16], ABO [13], 3D-FUTURE [20], and HSSD [34]." "For quantitative evaluations, we use Toys4k [80], which is not part of our training set".
- Paper App. B.1: "we include only the objects from Sketchfab (also known as ObjaverseV1 [15]) and GitHub in our training dataset" (i.e. NOT Thingiverse, Polycam, Smithsonian).
- README L205 / DATASET.md L3: "TRELLIS-500K is a dataset of 500K 3D assets curated from Objaverse(XL), ABO, 3D-FUTURE, HSSD, and Toys4k, filtered based on aesthetic scores."
- DATASET.md L13–21 (= HF card L46–54): ObjaverseXL (sketchfab) threshold 5.5 → 168,307; ObjaverseXL (github) 5.5 → 311,843; ABO 4.5 → 4,485; 3D-FUTURE 4.5 → 9,472; HSSD 4.5 → 6,670; All (training set) 500,777; Toys4k (evaluation set) 4.5 → 3,229.
- Completeness: closed enumerated list with per-object IDs published (CSV per source).
B. Per-object licence handling#
- TRELLIS-500K CSV columns (header read by HTTP range on all six CSVs at sha
288b0611…):
sha256,file_identifier,aesthetic_score,captions— no licence column. file_identifier is a Sketchfab URL (e.g.https://sketchfab.com/3d-models/18e8e405...), a GitHub blob URL,3/B07YBH2SR3.glb(ABO),3D-FUTURE-model/<uuid>,objects/.../<hash>.glb(HSSD). - Selection criterion stated: aesthetic score only ("filtered based on aesthetic scores"; App. B.2 uses "a pretrained aesthetic assessment model"). No licence filter stated in paper, README, DATASET.md or card; "licen*" appears in the paper only in arXiv's banner.
- Loader ObjaverseXL.py L29–37:
annotations = oxl.get_annotations();annotations = annotations[annotations['sha256'].isin(metadata['sha256'].values)];oxl.download_objects(annotations, ...)— selection by sha256 only, no licence predicate.grep -i licenover all five loader modules: 0 hits. - Measured (this session) TRELLIS-500K
ObjaverseXL_sketchfab.csv(168,307 rows; all matched by UID and by sha256): CC-BY 146,096 (86.80%); CC-BY-NC-SA 13,382 (7.95%); CC-BY-NC 5,739 (3.41%); CC-BY-SA 2,091 (1.24%); CC0 999 (0.59%) → 19,121 NC-family objects (11.36%) in the training list. Method for measured mixes below: per-objectlicensecolumn of https://huggingface.co/datasets/allenai/objaverse-xl/resolve/93fa0ca5a3067268290c673497bf974cebc83fa6/sketchfab/sketchfab.parquet (796,031 rows; whole-file tally CC-BY 702,400 / CC-BY-NC-SA 50,677 / CC-BY-NC 24,148 / CC-BY-SA 15,350 / CC0 3,456) joined on Sketchfab UID (or sha256) to the candidate's published list. Licence = value recorded in Objaverse-XL metadata at its 2023 release, not re-checked against live Sketchfab pages. Script output:licence-mix-sketchfab.txt(beside this file). - Measured TRELLIS-500K
ObjaverseXL_github.csv(311,843 sha256; 311,813 matched, 30 unmatched): None 271,750 (87.15%); MIT 23,088 (7.40%); GPL-3.0 8,843 (2.84%); Apache-2.0 3,678 (1.18%); CC0 1,144; Unlicense 844; GPL-2.0 688; AGPL-3.0 425; BSD-3 395; 33 distinct values. GitHub method:licensecolumn of https://huggingface.co/datasets/allenai/objaverse-xl/resolve/93fa0ca5a3067268290c673497bf974cebc83fa6/github/github.parquet (5,236,361 rows; whole-file: None 3,827,628 = 73.10%; 'MIT License' 19.14%; GPL-3.0 4.09%; 38 distinct values) joined on sha256 (or fileIdentifier). Values are GitHub-style licence names (e.g. 'BSD 3-Clause "New" or "Revised" License'); what None means and whether values are repo- or file-level is not documented (objaverse/xl/github.py @ 6996a386 lists the column only). Output:licence-mix-github.txt(beside this file; scriptgh_mix.py). - TRELLIS-500K card itself:
license: mit(card L2) — applies to the CSV metadata, it does not relicense the referenced assets (card says nothing about asset licences).
C. Dataset terms#
- Objaverse 1.0 (= Objaverse-XL sketchfab source) — https://huggingface.co/datasets/allenai/objaverse/raw/21e4e142159e2153706c23a3a02e55cec5591cea/README.md L16: "The use of the dataset as a whole is licensed under the ODC-By v1.0 license. Individual objects in Objaverse are all licensed as creative commons distributable objects, and may be under the following licenses:" L18–22: CC-BY 4.0 721K; CC-BY-NC 4.0 25K; CC-BY-NC-SA 4.0 52K; CC-BY-SA 4.0 16K; CC0 1.0 3.5K. L24: "The metadata will provide the license for each object." Paper https://arxiv.org/abs/2212.08051v1 §3: "Objects selected for OBJAVERSE have a distributable Creative Commons license and were obtained using Sketchfab's public API." No training clause found in the card or paper.
- Objaverse-XL — https://huggingface.co/datasets/allenai/objaverse-xl/raw/93fa0ca5a3067268290c673497bf974cebc83fa6/README.md L67: "The use of the dataset as a whole is licensed under the ODC-By v1.0 license. Individual objects in Objaverse-XL are licensed under different licenses." L63: "Polycam data is available by Polycam to academic researchers for non-commercial use upon request". Paper https://arxiv.org/abs/2307.05663v1 (PDF): GitHub source "We index 37M public files that contain common 3D object extensions" (§3.1) — no licence gate described for GitHub; datasheet: "Users must follow the license of the content with which the original files were distributed and the terms of service for each platform." and "The individual objects are subject to the licenses that they are released under, and users need to assess license questions based on downstream use." No training clause found (card or paper). No per-source licence mix stated for GitHub.
- ODC-By 1.0 — https://opendatacommons.org/licenses/by/1-0/ §3.1: "These rights explicitly include commercial use, and do not exclude any field of endeavour." ... "b. Creation of Derivative Databases". §2.4: "this License does not cover any rights (other than Database Rights or in contract) in individual Contents contained in the Database." No training clause.
- CC BY-NC 4.0 legal code — https://creativecommons.org/licenses/by-nc/4.0/legalcode.en: "NonCommercial means not primarily intended for or directed towards commercial advantage or monetary compensation."; grant: "produce, reproduce, and Share Adapted Material for NonCommercial purposes only." No training clause.
- ABO — https://amazon-berkeley-objects.s3.amazonaws.com/index.html (L39) "A CC BY 4.0-licensed dataset of Amazon products with metadata, catalog images, and 3D models." L513–514 "This work is licensed under the Creative Commons Attribution 4.0 International Public License (CC BY 4.0)." (i) commercial allowed, (ii) derivatives allowed with attribution, (iii) no training clause found in that page.
- 3D-FUTURE — "3D-FUTURE Data Sets Use License Agreement, Version: March 28, 2020": https://terms.aliyun.com/legal-agreement/terms/suit_bu1_ali_cloud/suit_bu1_ali_cloud202004171628_60052.html (the Tianchi landing page https://tianchi.aliyun.com/specials/promotion/alibaba-3d-future is JS-rendered; not readable via curl). §3.2: "Tao Bao hereby grants you a [worldwide], revocable, non-exclusive, non-transferable and royalty-free license, with no right to sub-license, ... but for scientific research purpose only". §3.2.2 "Create derivative works based on the data sets, reproduce, distribute, and disseminate the derivative works". §3.3.1 "You shall not commercialize the data sets;" §3.3.3 "You shall not use the data sets, derivative works of the data sets, and the results obtained based on the data sets for commercialization or providing external services;". No training/model/algorithm clause found in that agreement.
- HSSD — https://huggingface.co/datasets/hssd/hssd-models (gated=auto;
README not readable without login — raw returns HTTP 401). API cardData (sha
208ea6bc22d863bbc1cbaef1829b3e646600dfb6):
license: cc-by-nc-4.0; gate prompt: "You agree to use this dataset under the CC BY-NC 4.0 license terms". Same on hssd/hssd-hab card (sha 4369cb98…) L8–10. No training clause found. - Toys4k (eval only) — https://github.com/rehg-lab/lowshot-shapebias README @6cfc6bd8267d283261229aabb02be5dd3535b1b3 L81 "Please do not distribute the data made available here in any way". Access form https://forms.gle/w7Zf82umwaKxr9L7A "Terms of Use": "As the assets' licenses are mixed (some may be CC0, while others CC BY-SA or CC-BY-NC-SA), if considering whether you can use the complete dataset, it's helpful to consider whether your use case will comply with the more restrictive Creative Commons license CC BY-NC-SA." Attribution PDF lists source URLs only (sketchfab 2,034, poly.google 409, blendswap 412 lines), no licence per row. No training clause found.
D. Verdicts (all three TRELLIS weights; cards give no per-weight split)#
- Objaverse-XL (sketchfab = Objaverse 1.0) →
mixed-per-object. - Objaverse-XL (github) →
mixed-per-object. - ABO →
permissive. - 3D-FUTURE →
non-commercial-terms(research-only; §3.3.3 also bars commercial use of "results obtained based on the data sets", but no words about training). - HSSD →
non-commercial-terms. - Toys4k (evaluation only) →
mixed-per-object.
3. TripoSG#
Sources read:
- README @ fc5c40990181e2a756c4e0b1c2f4d6b5202faf8c:
https://raw.githubusercontent.com/VAST-AI-Research/TripoSG/fc5c40990181e2a756c4e0b1c2f4d6b5202faf8c/README.md
(paper link L6 → arXiv 2502.06608). Repo tree at the pin has no data/training
code (only inference scripts,
triposg/package, LICENSE, NOTICE). - Model card VAST-AI/TripoSG @ 2c1c516d22d58db486a058d98d31bb6177344e06 (HF sha
on read date):
https://huggingface.co/VAST-AI/TripoSG/raw/2c1c516d22d58db486a058d98d31bb6177344e06/README.md
— 44 lines,
license: mit, no training-data statement. - Paper: https://arxiv.org/abs/2502.06608v3 (v3 = 27 Mar 2025, latest; HTML read incl. math alt-text).
A. Training datasets as stated#
- README L30: "High-Quality Dataset: Trained on 2 million meticulously curated Image-SDF pairs, ensuring superior output quality" (no source named in README). L36: released weights = "TripoSG 1.5B parameter rectified flow model and VAE trained on 2048 latent tokens".
- Paper §4 "Data-Building System": "TripoSG is trained on existing open-source datasets such as Objaverse (-XL) (Deitke et al., 2023; Deitke et al., 2024) and ShapeNet (Chang et al., 2015), which contains approximately 10 million 3D data."
- Paper Fig. 2 caption: "The data-building system processes the 3D models from various datasets (e.g., Objaverse and ShapeNet) through a series of data processing steps to create the training data."
- Paper §5.2.1 "Dataset": "We train our model on Objaverse (-XL) (Deitke et al., 2023; Deitke et al., 2024), the largest publicly available 3D dataset, which contains over 10 million unique 3D objects from various sources. ... After preprocessing, we get 2 million high-quality 3D objects."
- Paper §5.1: "we trained a 1.5B parameter model on a 2M dataset with a latent resolution of 512 tokens ... Next, we switched to a latent resolution of 2048 tokens and continued training for an additional 300k steps" (this matches the released 1.5B/2048 weights); the 4B MoE stage used "1M high-quality dataset" (not released per README).
- Ablations: "our data-building system reduced the 800K Objaverse dataset to 180K" (§7.3).
- Completeness: open-ended ("such as", "e.g."); §4 names ShapeNet but §5.2.1 names only Objaverse(-XL); which Objaverse-XL sub-sources (GitHub / Thingiverse / Sketchfab / Polycam / Smithsonian) were used is not stated; no object ID list released. Also "for all untextured models, we ... use ControlNet++ to generate corresponding multi-view RGB data" (§4.3) — synthetic conditioning images.
B. Per-object licence handling#
- Filtering described (§4.1–4.4): quality scoring by a regression model trained on scores from "10 professional 3D modelers", filtering, orientation fixing, watertight conversion. No licence filter stated; "licen*" appears in the paper HTML only in arXiv's banner. Card and README silent.
C. Dataset terms#
- Objaverse 1.0 / Objaverse-XL / ODC-By: see candidate 2 §C.
- ShapeNet — terms linked from https://shapenet.org/terms (now a Hugging Face org page) → https://huggingface.co/spaces/ShapeNet/README/raw/33d099e7a7cb4122718d0c742e1441b61187f7cf/TERMS.md (space sha on read date). L3: "Researcher shall use the Database only for non-commercial research and educational purposes." L5: "...including but not limited to Researcher's use of any copies of copyrighted 3D models that he or she may create from the Database." L8: "If Researcher is employed by a for-profit, commercial entity, Researcher's employer shall also be bound by these terms and conditions". (i) non-commercial only; (ii) no derivative grant; (iii) no training clause found in that TERMS.md.
D. Verdicts (TripoSG)#
- Objaverse(-XL), sub-sources unstated →
mixed-per-object. - ShapeNet (named in §4/Fig. 2 as "such as"/"e.g."; not in §5.2.1) →
non-commercial-terms. - Any further "such as" sources →
unknown(not identified).
4. Step1X-3D (Geometry-1300m + Texture)#
Sources read:
- README @ cb5ac944709c6c913109070c7b90c3447f57f3d4:
https://raw.githubusercontent.com/stepfun-ai/Step1X-3D/cb5ac944709c6c913109070c7b90c3447f57f3d4/README.md
(paper link L11 → arXiv 2505.07747); data/README.md @ same (describes only
watertight_and_sampling.py). - Model card stepfun-ai/Step1X-3D @ bf7084495b3a72222f36549b7942948aa4d9daa7
(sha on read date):
license: apache-2.0; L23/L31 repeat README L32/L54; no further data statement. - Released asset list: HF dataset stepfun-ai/Step1X-3D-obj-data @
7acf2121a1c88edfd753e2425aba72fff1553990 (lastModified 2025-05-22): files
objaverse_320k.json(16,788,875 B),objaverse_texture_30k.json(1,692,597 B),objaverse_xl_github_url_480k.json(102,691,992 B), README (card:license: apache-2.0; body "Thhis is the training subdataset for [Step1X-3D ...]"). ModelScope mirror https://www.modelscope.cn/api/v1/datasets/stepfun-ai/Step1X-3D-obj-data lists the same three JSONs with identical byte sizes (ModelScopeLicense: Apache License 2.0). - Paper: https://arxiv.org/abs/2505.07747v1 (only version, 12 May 2025; HTML read).
A. Training datasets as stated#
- Paper §3.1: "Through our comprehensive data processing pipeline, we successfully curated roughly 2 million high-quality 3D assets from multiple sources: extracting 320k valid samples from the original Objaverse dataset, obtaining an additional 480k from Objaverse-XL, and combining these with carefully selected data from ABO, 3D-FUTURE, and some internal created data."
- Paper §1: "The Step1X-3D data pipeline begins with over 5M 3D assets sourced from public datasets (e.g., Objaverse [18], Objaverse-XL [17], and etc) and proprietary collections. ... This process yields a curated dataset of 2M high-quality assets, a subset of which (alomost 800K assets derived from public data) will be openly released."
- Paper Fig. 3 caption: "(a) we filter bad quality data from web captured 3D assets (e.g., Objaverse, ABO, Thingiverse, and self-collected assets)."
- Paper §4.2 (texture model): "From the 320K Objaverse dataset cleaned in Sec. 3.1, we further curated 30K 3D assets for multi-view generative model training."
- README L32: "We release the 800K uids of high quality 3D assets (excluding
self-collected assets) ... for both training 3D geometry and synthesis."
README L72–75 table: Objaverse
objaverse_320k.json320K; Objaverse-XLobjaverse_xl_github_url_480k.json480K; "Assets for texture synthesis"objaverse_texture_30k.json30K; "Assets in ModelScope" 830K. - JSON formats (first bytes read by HTTP range): objaverse_320k / texture_30k =
Objaverse 1.0 paths
"000-130/f7654b3f336249d8bb6ae12503e6b543.glb"(Sketchfab UIDs); objaverse_xl_github_url_480k ={sha256: "https://github.com/<owner>/<repo>/blob/<commit>/<path>"}→ the 480K Objaverse-XL part is the GitHub source. - Completeness: open-ended and partly unreleased. Named: Objaverse 1.0 (320K), Objaverse-XL GitHub (480K), ABO, 3D-FUTURE, Thingiverse (figure caption only), "internal created data"/"proprietary collections"/"self-collected assets". No ABO/3D-FUTURE/Thingiverse/internal ID lists released. Arithmetic from the authors' own numbers (not a stated figure): 2M − 320K − 480K ≈ 1.2M assets are outside the released lists.
- Texture weights: 30K subset of the 320K Objaverse 1.0 list (stated). Texture module is "SD-XL-fine-tuned" (base encoder out of scope).
B. Per-object licence handling#
- Measured:
objaverse_320k.jsonactually lists 322,883 paths → 321,460 matched, 1,423 unmatched: CC-BY 290,425 (90.35%); CC-BY-NC 12,182 (3.79%); CC-BY-NC-SA 9,374 (2.92%); CC-BY-SA 6,973 (2.17%); CC0 2,506 (0.78%) → 21,556 NC-family (6.71%).objaverse_texture_30k.jsonlists 32,550 → 32,270 matched, 280 unmatched: CC-BY 29,486 (91.37%); CC-BY-NC 1,540 (4.77%); CC0 450; CC-BY-SA 444; CC-BY-NC-SA 350 → 1,890 NC-family (5.86%) in the texture-model list. (Method: see candidate 2 §B.) - Measured
objaverse_xl_github_url_480k.json(480,654 sha256 keys; 480,634 matched, 20 unmatched): None 417,853 (86.94%); MIT 35,224 (7.33%); GPL-3.0 13,806 (2.87%); Apache-2.0 7,090 (1.48%); CC0 1,418; Unlicense 1,259; GPL-2.0 883; BSD-3 672; AGPL-3.0 655; CC-BY 417; 34 distinct values. (GitHub method: see candidate 2 §B.) - Filters stated (§3.1 "Data Filter"): texture quality, single-surface, small-object coverage, transparent materials, wrong normals, "Name and Mesh Type Filtering". No licence filter stated; "licen*" appears in the paper HTML only in arXiv's banner. The released JSONs carry no licence field (plain path list / sha256→URL map). README/model card/dataset card silent on per-object licences; the uid lists themselves are tagged Apache-2.0, which cannot relicense the referenced objects.
C. Dataset terms#
- Objaverse 1.0, Objaverse-XL (GitHub), ODC-By, ABO, 3D-FUTURE: see candidate 2 §C.
- Thingiverse — terms pages https://www.thingiverse.com/legal, /legal/terms, /terms return a Next.js shell (HTTP 200, no text; legal page chunk loads content client-side); WebFetch saw only the page title. Not readable. Objaverse-XL paper (2307.05663v1 §3.1) describes Thingiverse objects as "predominantly released under Creative Commons licenses" (i.e. per-object).
- "internal created data" / "proprietary collections" / "self-collected assets" — no terms published.
D. Verdicts (Step1X-3D geometry; texture = Objaverse 1.0 row only)#
- Objaverse 1.0 (320K; texture 30K subset) →
mixed-per-object. - Objaverse-XL GitHub (480K) →
mixed-per-object. - ABO →
permissive. - 3D-FUTURE →
non-commercial-terms. - Thingiverse (figure caption "e.g.") →
unknown(terms not readable; objects per-object CC per Objaverse-XL paper). - Internal/self-collected/proprietary (~1.2M by arithmetic) →
unknown.
5. SkinTokens / TokenRig#
Sources read:
- README @ 273b691d35989d71cd17ff2895fdc735097b92d1:
https://raw.githubusercontent.com/VAST-AI-Research/SkinTokens/273b691d35989d71cd17ff2895fdc735097b92d1/README.md
(paper link L4 → arXiv 2602.04805). Repo has no data lists (src/data/*.py
loaders only;
grepfor VRoid/ModelsResource/Articulation/licence in src/data/datapath.py: 0 hits). - Model card VAST-AI/SkinTokens @ 79736cad0fd84de384d5eede659b4ebd24effe33 (sha
on read date, lastModified 2026-04-20):
https://huggingface.co/VAST-AI/SkinTokens/raw/79736cad0fd84de384d5eede659b4ebd24effe33/README.md
(
license: mit). Files:experiments/articulation_xl_quantization_256_token_4/grpo_1400.ckpt,experiments/skin_vae_2_10_32768/last.ckpt. - Paper: https://arxiv.org/abs/2602.04805v1 (only version, 4 Feb 2026; HTML read).
A. Training datasets as stated#
- README L89: "
articulation_xl_quantization_256_token_4| TokenRig autoregressive rigging model, trained on ArticulationXL 2.0 + VRoid Hub + ModelsResource and refined with GRPO". - Card L63: "Qwen3-0.6B-based Transformer trained on a composite of
ArticulationXL 2.0 (70%), VRoid Hub (20%), and ModelsResource
(10%) ... then refined with GRPO for 1,400 steps." Card L146: "The
checkpoints were trained on a composite of ArticulationXL 2.0 (70%), VRoid Hub
(20%), and ModelsResource (10%); the processed data splits will be released as
a separate dataset repository later." Card L49: "The training data
(
ArticulationXLsplits and processed meshes) used to train these checkpoints will be released separately in a future update." → per-object lists not released on read date. - Paper §4.1.1 "Dataset Configuration": "we construct a composite dataset sourcing from Articulation2.0 (Song et al., 2025b) (70%), VRoid Hub (Isozaki et al., 2021) (20%), and ModelsResource (Models-Resource, 2019; Xu et al., 2020) (10%)." §4.1.4: "For the post-training stage, we curate a smaller, high-complexity dataset of AI-generated meshes." (generator not named). Refs: Song et al. 2025b = MagicArticulate (CVPR 2025); Isozaki et al. 2021 = "VRoid studio: a tool for making anime-like 3d characters" (a tool paper, not a dataset release); Models-Resource 2019 = "The models-resource"; Xu et al. 2020 = RigNet.
- Completeness: closed list of three sources + unnamed AI-generated meshes for GRPO. Sizes of the VRoid Hub / ModelsResource subsets not stated. (Note: paper says GRPO "800 steps"; card/checkpoint say 1,400.) Context only, not claimed by SkinTokens: predecessor UniRig (arXiv 2504.12451v1 §4.1) "curated a dataset of 2,061 anime-style 3D models from VRoidHub ... compiled by first filtering the available models on VRoidHub based on the number of bones" — no licence/condition filter stated there either.
B. Per-object licence handling#
- No licence handling stated in README, card, or paper (paper "licen*" hits: only arXiv banner "License: CC BY 4.0"). No statement that VRoid Hub models were selected by their per-model Conditions of Use.
- Measured: the 6,622 Articulation-XL2.0
sketchfabrows → CC-BY 6,276 (94.77%); CC-BY-NC 190; CC-BY-NC-SA 94; CC-BY-SA 61; CC0 1 → 284 NC-family (4.29%). Which Articulation-XL2.0 objects SkinTokens used is not published. (Method: see candidate 2 §B.) - Measured: the 42,015 Articulation-XL2.0
githubrows (all matched by fileIdentifier): None 37,220 (88.59%); MIT 2,699 (6.42%); GPL-3.0 824 (1.96%); Apache-2.0 534; GPL-2.0 262; CC0 149; Unlicense 137; AGPL-3.0 54; 24 distinct values. (GitHub method: see candidate 2 §B.) - Articulation-XL2.0 metadata (read in full, 48,637 rows):
https://huggingface.co/datasets/Seed3D/Articulation-XL2.0/resolve/11f777d83a1e7ee7c5a22201da4e140cafac1837/meta_Articulation_XL_2.0.csv
— header
uuid,source,vertex_count,face_count,joint_count,bone_count,category_label,fileType,fileIdentifier(no licence column);sourcetally: github 42,015; sketchfab 6,622.
C. Dataset terms#
- Articulation-XL2.0 — card
https://huggingface.co/datasets/Seed3D/Articulation-XL2.0/raw/11f777d83a1e7ee7c5a22201da4e140cafac1837/README.md
(
chaoyue7/Articulation-XL2.0resolves to the same repo/sha): L2license: cc-by-4.0; L55 "a large-scale dataset featuring over 48K 3D models with high-quality articulation annotations, filtered from Objaverse-XL." MagicArticulate paper https://arxiv.org/abs/2502.12135v2: "carefully curated from Objaverse-XL" — no licence filter stated ("licen*" only in arXiv banner). The CC-BY-4.0 tag covers the released annotations/processed data; underlying objects are Objaverse-XL GitHub/Sketchfab objects → Objaverse-XL terms (candidate 2 §C). No training clause found. - VRoid Hub — https://hub.vroid.com/en/terms (serves pixiv's policy bundle;
"English translation ... the original Japanese version shall prevail").
- pixiv Service Master Terms of Use, Article 14 "Prohibited Conduct"
(applies to VRoid Hub via VRoid Hub Terms Art. 4: "User must not engage in
prohibited conduct specified in Article 14 of the pixiv Master Terms of
Use."):
- 3rd listed item: "Engaging in activities that have a commercial, business, or profit-making objective, or using the Service or any part of the Service (including content, information, functions, systems, and programs) outside the Service for such purposes or in preparation for such purposes;"
- 4th listed item (AI/training): "Conduct that the Company determines to unreasonably prejudice the interests of the User who Posted etc. the Posted Information to the Services or a related services by using the results of data analysis (including the conduct of learning to develop artificial intelligence.The same shall apply hereafter) of Posted Information Posted etc. to the Services or related services. However, this does not apply if the User has obtained the permission of the User who Posted etc. the Posted Information for this data analysis;"
- Revision log: "May 31, 2023: ... The current Item 4 and Item 5 were newly added to 'Article 14: Prohibited Conduct'." and "June 22, 2026: ... Article 14. Prohibited Conduct ... ha[s] been revised." (items 4–5 = the two data-analysis/AI items by list order in the extracted text).
- "VRoid Hub" Terms of Use (in force "December 21, 2018", revised "July 5, 2021"): Art. 3 per-model "Conditions of Use" set by the uploader: "Whether to allow downloading by User ... Whether to allow commercial use by corporations / Whether to allow commercial use by individuals ... Whether to allow modifications / Whether to allow redistribution / Whether to require credit attribution"; "such Registered Works can be used only within the scope of the Conditions of Use". Art. 4: "User must not reverse engineer, decompile, disassemble, or otherwise analyze Software, systems, contents, Registered Works, or other data or programs on the Individual Service". Art. 5: rights "will belong to the creator of the relevant Registered Work(s)". No AI/ML toggle appears among the Art. 3 Conditions of Use items.
- pixiv Service Master Terms of Use, Article 14 "Prohibited Conduct"
(applies to VRoid Hub via VRoid Hub Terms Art. 4: "User must not engage in
prohibited conduct specified in Article 14 of the pixiv Master Terms of
Use."):
- ModelsResource — site https://www.models-resource.com/ (and /faq/, /tos/, /legal/, robots.txt, sister sites spriters-resource.com, vg-resource.com): not readable — HTTP 403 Cloudflare "Just a moment..." challenge to curl and WebFetch; Internet Archive returned 429 / "Temporarily Offline"; Claude-in-Chrome extension not connected. Site's own page titles as indexed by search (not page bodies): "Bowser's Car - Super Mario 3D World - Wii U - The Models Resource", "GameCube - The Models Resource" (site organised by game console/title). Dataset provenance: RigNet README (zhan-xu/RigNet @ 589a928b9731b2693165e80d190e343bd6d6b762) L63 "Our dataset ModelsResource-RigNetv1 has 2,703 models." (L111–112 "This project is under LICENSE-GPLv3." = code); RigNet paper https://arxiv.org/abs/2005.00559v2: "2703 rigged characters mined from an online repository [Models-Resource 2019]"; Xu et al. 2019 https://arxiv.org/abs/1908.08506v1: "We first collected a dataset of 3277 rigged characters from an online repository, called Models Resource". No licence or rights statement for the models in any of these.
D. Verdicts (SkinTokens FSQ-CVAE + TokenRig)#
- Articulation-XL2.0 (Objaverse-XL GitHub 42,015 / Sketchfab 6,622) →
mixed-per-object. - VRoid Hub →
explicit-training-prohibition— pixiv Master Terms Art. 14 names "learning to develop artificial intelligence" as prohibited when the Company determines it unreasonably prejudices the poster, unless the poster's permission is obtained (conditional wording), on top of per-model Conditions of Use (commercial/modification/redistribution toggles). SkinTokens states no permission or per-model filter. - ModelsResource →
unknown(site terms not readable; no licence stated by RigNet/Xu 2019/SkinTokens; site titles indicate video-game models, but "unlicensed third-party copyright" is not established from a readable source). - AI-generated meshes (GRPO stage, generator unnamed) →
unknown.
6. Paint3D (incl. its UV-position ControlNet GeorgeQi/Paint3d_UVPos_Control)#
Sources read:
- README @ d5545d38db6aaf78efc563269f70177688a4218e: https://raw.githubusercontent.com/OpenTexture/Paint3D/d5545d38db6aaf78efc563269f70177688a4218e/README.md (L8 arXiv 2312.13913; L52 "For UV-position controlnet, you can find it here").
- Split files @ same commit: tools/idx_split/idx_all.txt (105,301 lines),
idx_train.txt (105,000), idx_test.txt (301) — one 32-hex UID per line (e.g.
000074a334c541878360457c672b6c2e), i.e. Objaverse-1.0-style UIDs; no other columns. - Model card GeorgeQi/Paint3d_UVPos_Control @
1d904225f9b4ff1cd2ff414082b9b211e4df6be5 (sha on read date, lastModified
2024-04-26):
https://huggingface.co/GeorgeQi/Paint3d_UVPos_Control/raw/1d904225f9b4ff1cd2ff414082b9b211e4df6be5/README.md
—
license: apache-2.0; 46 lines; no training-data statement. - Paper: https://arxiv.org/abs/2312.13913v2 (latest, 22 Dec 2023; HTML read).
A. Training datasets as stated#
- Paper §4 "Datasets.": "We conduct experiments on a subset of textured meshes from the Objaverse [13] dataset. We exclude meshes devoid of textures, those with monochromatic texture, and 3D scene objects composed of multiple meshes. The filtered subset contains 105,301 texture meshes, with 105,000 meshes utilized for training the position encoder and 301 meshes employed for evaluating our model. Additionally, we gather 30 meshes in the wild to assess our model." Ref [13] = Objaverse (Deitke et al., CVPR 2023) = Objaverse 1.0.
- Paper §3: "Our texture diffusion model is trained using a dataset consisting of paired position maps and texture maps" ... "we freeze εθ as suggested by [71] and only optimize the position encoder τp" → the released UVPos ControlNet is trained on that Objaverse subset.
- The split files' line counts (105,301 / 105,000 / 301) match the paper's numbers exactly.
- Completeness: closed (Objaverse 1.0 only) with a published per-object UID list. ShapeNet appears only in an appendix comparison (chair/table, "text and images as provided in [69]"), not as training data.
- README L172: "uses datasets which each have their own respective licenses that must also be followed."
B. Per-object licence handling#
- Measured: idx_train.txt (105,000 UIDs) → 104,361 matched, 639 unmatched: CC-BY 87,463 (83.81%); CC-BY-NC-SA 9,276 (8.89%); CC-BY-NC 4,575 (4.38%); CC-BY-SA 2,207 (2.11%); CC0 840 (0.80%) → 13,851 NC-family (13.27%). (Method: see candidate 2 §B.)
- Filters stated: textured / non-monochromatic / single-object. No licence filter stated; paper "licen*" hits only the arXiv banner ("License: CC BY-NC-SA 4.0" — the paper's own licence). Split files carry no licence field.
C. Dataset terms#
- Objaverse 1.0 + ODC-By: see candidate 2 §C (card mix: CC-BY 721K; CC-BY-NC 25K; CC-BY-NC-SA 52K; CC-BY-SA 16K; CC0 3.5K).
D. Verdicts (Paint3D UVPos ControlNet)#
- Objaverse 1.0 (105,000 train UIDs) →
mixed-per-object.
7. TEXGen#
Sources read:
- README @ bbed677afd290dee103014a134609df01b3d6422:
https://raw.githubusercontent.com/CVMI-Lab/TEXGen/bbed677afd290dee103014a134609df01b3d6422/README.md
(L28 arXiv 2411.14740). Repo tree at pin has no training-data list (only
assets/input_list/test_input.jsonl, 32-hex ids + captions for example meshes). - Model card Andyx/TEXGen @ b960de1d47cb8172214254177f74392d37347107 (sha on
read date, lastModified 2024-12-24):
license: apache-2.0, body = one sentence linking the paper; no training-data statement. Files:texgen_v1.ckpt,texgen_v2.ckpt. - Paper: https://arxiv.org/abs/2411.14740v1 (only version, 22 Nov 2024; HTML read). Published as ACM TOG 43(6) art. 213 (README citation).
A. Training datasets as stated#
- Paper §5: "We use Objaverse (Deitke et al., 2023) as our raw data source, which comprises over 800,000 3D meshes. After processing and cleaning this dataset, we extracted a total of 120,400 data pairs. Of these, 120,000 pairs are designated for training and the remaining 400 pairs are set aside for evaluation."
- Paper App. A.1: "we first filter out meshes with poor texture quality ... we use Gemini (Team et al., 2023) to acquire each mesh's caption based on their rendered images. Ultimately, we processed and obtained 120,400 meshes with their corresponding texture images, using 120,000 for training and 400 for evaluation."
- README L95: "Note that the released model is trained with Flow Matching which is different from the paper version, since we find it more stable." — the training data of the released checkpoints is not separately stated (paper describes the paper version).
- Completeness: closed (Objaverse 1.0 only) but no per-object UID list released; exact 120K subset not reproducible from public artefacts.
B. Per-object licence handling#
- Filter stated: texture quality only. No licence filter stated; paper "licen*" hits only the arXiv banner ("License: CC BY 4.0"). README/card silent.
C. Dataset terms#
- Objaverse 1.0 + ODC-By: see candidate 2 §C.
D. Verdicts (TEXGen)#
- Objaverse 1.0 (120K, list unreleased) →
mixed-per-object. - Released flow-matching checkpoints' data (not stated separately) →
unknownbeyond the paper's Objaverse statement.
8. SAM 3D Objects#
Sources read (public only):
- README @ f91db411c50efee93d8db7aeb323885650f6f722: https://raw.githubusercontent.com/facebookresearch/sam-3d-objects/f91db411c50efee93d8db7aeb323885650f6f722/README.md — L81 "The SAM 3D Objects model checkpoints and code are licensed under SAM License."; L152 arXiv 2511.16624; no training-data statement. LICENSE @ same commit read (grant/derivative clauses only; it governs Meta's materials, not training data).
- HF model card facebook/sam-3d-objects (sha
2e73555018d2741ccd486e56c24fac41155a1dc6, gated=manual): README raw returns
HTTP 401 → card body not readable; API cardData only (
license: other, gate form fields). - Paper: https://arxiv.org/abs/2511.16624v2 (v2 = 2 Jun 2026, latest; HTML read).
A. Training datasets as stated#
- §3.1.1 Pretraining: "we gather a set of image I, shape S, and texture T triplets, using 2.7 million object meshes from Objaverse-XL (Deitke et al., 2023) and licensed datasets, and render them from 24 viewpoints ... We call this dataset Iso-3DO". Table (training stages): Pre-training Iso-3DO — 2.7M meshes, 64.8M images. Texture & Refinement pretraining: "Iso-3DO-500K, a partition of Iso-3DO data with high aesthetics" (§C.5).
- Mid-training: RP-3DO render-paste — "we replace the object in the image with a synthetic 3D object drawn from the same synthetic sources used in Iso-3DO" (App. B.2); FO 55.1M samples / 2.87M unique meshes / 11.17M images; OS-R + "ProcThor" (Table 1; ref Deitke et al. 2022 ProcTHOR).
- Post-training (SFT/alignment): MITL-3DO (annotator-verified model/retrieval outputs; "Retrieval: The nearest 3D object is retrieved from a shape object library (pretraining data)") and Art-3DO ("route a small percentage of these hardest cases to professional 3D artists"). App. A.1 image sources: "SA-1B, MetaCLIP ... SA-V ... Ego4D, Ego-Exo4D, AEA, AEO, Nymeria ... Food Recognition (Bossard et al., 2014) ... BDD100k"; texture MITL: "The images are sourced from SA-1B".
- Completeness: open — Objaverse-XL sub-sources not stated; "licensed datasets" not named; per-object lists not released. Eval-only: SA-3DAO, ISO3D, Aria Digital Twin, held-out Objaverse split.
B. Per-object licence handling#
- Filters stated (App. B.1): aesthetic score (texture model), rule-based geometry filters. No per-object licence filter stated for the Objaverse-XL portion; paper "licen*" hits: the phrase "licensed datasets" (§3.1.1) and arXiv's banner only.
C. Dataset terms#
- Objaverse-XL + ODC-By: see candidate 2 §C.
- "licensed datasets" (unnamed): not identified → no terms.
- ProcTHOR — https://raw.githubusercontent.com/allenai/procthor-10k/439193522244720b86d8c81cde2e51e3a4d150cf/LICENSE (and allenai/procthor @ 53d5bd4c…/LICENSE): "Apache License Version 2.0". (Paper does not say which ProcTHOR release.) No training clause.
- SA-V — https://raw.githubusercontent.com/facebookresearch/sam2/2b90b9f5ceec907a1c18123530e92e794ad901a4/sav_dataset/README.md L162: "The videos and annotations in SA-V Dataset are released under CC BY 4.0." No training clause.
- SA-1B — segment-anything README @ dca509fe793f601edb92606367a655c15ac00fdf L118: "By downloading the datasets you agree that you have read and accepted the terms of the SA-1B Dataset Research License." Licence text at https://ai.facebook.com/datasets/segment-anything-downloads/ → ai.meta.com returns HTTP 400 to curl: not readable.
- Ego4D — https://ego4d-data.org/pdfs/Ego4D-Licenses-Draft.pdf (first licence in the PDF, University of Bristol): "the “Purpose” is limited to: ... (2) using the Database to research, develop, train, evaluate, or improve any software, algorithm, machine learning model, technique, or technology designed to: ... (b) model or reconstruct at least one three-dimensional object or environment; ... wherein all permitted uses in subsections 2(a)-2(c) of this section may be used for academic research, commercial or noncommercial product development". "Such software, algorithms, machine learning models ... may be used for academic, commercial, or noncommercial purposes."
- Ego-Exo4D ("EgoExo License" linked from https://ego4ddataset.com/ego4d-license/), AEA / AEO / Nymeria (https://www.projectaria.com/datasets/... → HTTP 500): not read / not readable.
- BDD100K — https://raw.githubusercontent.com/bdd100k/bdd100k/9ac17c6c7c51d2fc83065fccd707cd5b1882a293/doc/source/license.rst L16: "Permission to use, copy, modify, and distribute this software and its documentation for educational, research, and not-for-profit purposes ... and permission to use, copy, modify and distribute this software for commercial purposes (such rights not subject to transfer) to BDD and BAIR Commons members and their affiliates". No training clause.
- Food-101 — https://data.vision.ee.ethz.ch/cvl/datasets_extra/food-101/ : page text has no licence statement (archive-internal terms not read; 5 GB).
- MetaCLIP — https://raw.githubusercontent.com/facebookresearch/MetaCLIP/f47f7841f6a91cc5676729a3d125519393d87d1e/README.md L206 "The majority of Meta CLIP is licensed under CC-BY-NC" (that is the code/project; the image data behind "MetaCLIP (Xu et al., 2024a)" in SAM 3D is not identified).
D. Verdicts (SAM 3D Objects geometry + texture)#
- Objaverse-XL (2.7M meshes incl. unnamed sources) →
mixed-per-object. - "licensed datasets" (unnamed) →
unknown. - ProcTHOR →
permissive. - SA-V (images/video for data engine) →
permissive. - Ego4D →
permissive(licence expressly allows training 3D-reconstruction models for commercial product development). - BDD100K →
non-commercial-terms(commercial rights only for BDD/BAIR Commons members). - SA-1B, Ego-Exo4D, AEA, AEO, Nymeria, Food-101, MetaCLIP image data →
unknown. - Art-3DO / MITL annotations (Meta-commissioned) →
unknown(no terms published).
Summary table#
Terms URLs are the ones quoted in the candidate sections (read 2026-09-16). "eval" = evaluation-only, not training.
| candidate | weight repo | dataset | verdict | terms URL | key quote (≤25 words) |
|---|---|---|---|---|---|
| TRELLIS.2-4B | microsoft/TRELLIS.2-4B | TexVerse (503,387 used) | mixed-per-object | https://huggingface.co/datasets/YiboZhang2001/TexVerse | "Individual objects in TexVerse are all licensed as creative commons distributable objects. The metadata will provide the license for each object." |
| TRELLIS.2-4B | microsoft/TRELLIS.2-4B | Objaverse-XL sketchfab (=Objaverse 1.0) | mixed-per-object | https://huggingface.co/datasets/allenai/objaverse | "Individual objects in Objaverse are all licensed as creative commons distributable objects" (CC-BY-NC / CC-BY-NC-SA among them) |
| TRELLIS.2-4B | microsoft/TRELLIS.2-4B | Objaverse-XL github | mixed-per-object | https://huggingface.co/datasets/allenai/objaverse-xl | "Individual objects in Objaverse-XL are licensed under different licenses." |
| TRELLIS.2-4B | microsoft/TRELLIS.2-4B | ABO | permissive | https://amazon-berkeley-objects.s3.amazonaws.com/index.html | "This work is licensed under the Creative Commons Attribution 4.0 International Public License (CC BY 4.0)." |
| TRELLIS.2-4B | microsoft/TRELLIS.2-4B | HSSD | non-commercial-terms | https://huggingface.co/datasets/hssd/hssd-models | "You agree to use this dataset under the CC BY-NC 4.0 license terms" |
| TRELLIS.2-4B | microsoft/TRELLIS.2-4B | Toys4k (eval) | mixed-per-object | https://forms.gle/w7Zf82umwaKxr9L7A | "consider whether your use case will comply with the more restrictive Creative Commons license CC BY-NC-SA." |
| TRELLIS image-large / text-large / text-xlarge | microsoft/TRELLIS-image-large, -text-large, -text-xlarge | Objaverse-XL sketchfab (168,307) | mixed-per-object | https://huggingface.co/datasets/allenai/objaverse | "Individual objects in Objaverse are all licensed as creative commons distributable objects" — measured 11.36% NC-family in TRELLIS-500K list |
| TRELLIS (3 weights) | same | Objaverse-XL github (311,843) | mixed-per-object | https://huggingface.co/datasets/allenai/objaverse-xl | "Individual objects in Objaverse-XL are licensed under different licenses." — measured 87.15% license=None |
| TRELLIS (3 weights) | same | ABO | permissive | https://amazon-berkeley-objects.s3.amazonaws.com/index.html | "A CC BY 4.0-licensed dataset of Amazon products with metadata, catalog images, and 3D models." |
| TRELLIS (3 weights) | same | 3D-FUTURE (9,472) | non-commercial-terms | https://terms.aliyun.com/legal-agreement/terms/suit_bu1_ali_cloud/suit_bu1_ali_cloud202004171628_60052.html | "You shall not use the data sets, derivative works of the data sets, and the results obtained based on the data sets for commercialization" |
| TRELLIS (3 weights) | same | HSSD (6,670) | non-commercial-terms | https://huggingface.co/datasets/hssd/hssd-models | "You agree to use this dataset under the CC BY-NC 4.0 license terms" |
| TRELLIS (3 weights) | same | Toys4k (eval) | mixed-per-object | https://forms.gle/w7Zf82umwaKxr9L7A | "As the assets' licenses are mixed (some may be CC0, while others CC BY-SA or CC-BY-NC-SA)" |
| TripoSG | VAST-AI/TripoSG | Objaverse(-XL), sub-sources unstated | mixed-per-object | https://huggingface.co/datasets/allenai/objaverse-xl | "Individual objects in Objaverse-XL are licensed under different licenses." |
| TripoSG | VAST-AI/TripoSG | ShapeNet ("such as") | non-commercial-terms | https://huggingface.co/spaces/ShapeNet/README/blob/main/TERMS.md | "Researcher shall use the Database only for non-commercial research and educational purposes." |
| TripoSG | VAST-AI/TripoSG | further "such as" sources | unknown | — | "trained on existing open-source datasets such as Objaverse (-XL) ... and ShapeNet" |
| Step1X-3D (geometry) | stepfun-ai/Step1X-3D | Objaverse 1.0 (322,883 listed) | mixed-per-object | https://huggingface.co/datasets/allenai/objaverse | measured 6.71% NC-family; card: "may be under the following licenses: ... CC-BY-NC 4.0 ... CC-BY-NC-SA 4.0" |
| Step1X-3D (texture) | stepfun-ai/Step1X-3D | Objaverse 1.0 texture list (32,550 listed) | mixed-per-object | https://huggingface.co/datasets/allenai/objaverse | measured 5.86% NC-family |
| Step1X-3D (geometry) | stepfun-ai/Step1X-3D | Objaverse-XL github (480,654 listed; measured 86.94% license=None) |
mixed-per-object | https://huggingface.co/datasets/allenai/objaverse-xl | "The individual objects are subject to the licenses that they are released under" (Objaverse-XL datasheet) |
| Step1X-3D (geometry) | stepfun-ai/Step1X-3D | ABO | permissive | https://amazon-berkeley-objects.s3.amazonaws.com/index.html | "licensed under the Creative Commons Attribution 4.0 International Public License (CC BY 4.0)" |
| Step1X-3D (geometry) | stepfun-ai/Step1X-3D | 3D-FUTURE | non-commercial-terms | https://terms.aliyun.com/legal-agreement/terms/suit_bu1_ali_cloud/suit_bu1_ali_cloud202004171628_60052.html | "for scientific research purpose only" |
| Step1X-3D (geometry) | stepfun-ai/Step1X-3D | Thingiverse (fig. caption) | unknown | https://www.thingiverse.com/legal (not readable) | — |
| Step1X-3D (geometry) | stepfun-ai/Step1X-3D | internal / self-collected / proprietary (~1.2M by arithmetic) | unknown | — | "combining these with carefully selected data from ABO, 3D-FUTURE, and some internal created data" |
| SkinTokens / TokenRig | VAST-AI/SkinTokens | Articulation-XL2.0 (70%; its github rows 88.59% license=None, sketchfab rows 4.29% NC) |
mixed-per-object | https://huggingface.co/datasets/Seed3D/Articulation-XL2.0 | "over 48K 3D models with high-quality articulation annotations, filtered from Objaverse-XL" |
| SkinTokens / TokenRig | VAST-AI/SkinTokens | VRoid Hub (20%) | explicit-training-prohibition | https://hub.vroid.com/en/terms | "Conduct that the Company determines to unreasonably prejudice the interests of the User who Posted ... (including the conduct of learning to develop artificial intelligence" (conditional; poster permission exempts) |
| SkinTokens / TokenRig | VAST-AI/SkinTokens | ModelsResource (10%) | unknown | https://www.models-resource.com/ (403, not readable) | RigNet: "2703 rigged characters mined from an online repository [Models-Resource 2019]" |
| SkinTokens / TokenRig | VAST-AI/SkinTokens | AI-generated meshes (GRPO) | unknown | — | "we curate a smaller, high-complexity dataset of AI-generated meshes" |
| Paint3D | GeorgeQi/Paint3d_UVPos_Control | Objaverse 1.0 (105,000 UIDs) | mixed-per-object | https://huggingface.co/datasets/allenai/objaverse | measured 13.27% NC-family; "subset of textured meshes from the Objaverse dataset" |
| TEXGen | Andyx/TEXGen | Objaverse 1.0 (120,000; list unreleased) | mixed-per-object | https://huggingface.co/datasets/allenai/objaverse | "We use Objaverse ... as our raw data source" |
| SAM 3D Objects | facebook/sam-3d-objects | Objaverse-XL (part of 2.7M meshes) | mixed-per-object | https://huggingface.co/datasets/allenai/objaverse-xl | "Individual objects in Objaverse-XL are licensed under different licenses." |
| SAM 3D Objects | facebook/sam-3d-objects | "licensed datasets" (unnamed) | unknown | — | "2.7 million object meshes from Objaverse-XL ... and licensed datasets" |
| SAM 3D Objects | facebook/sam-3d-objects | ProcTHOR | permissive | https://github.com/allenai/procthor-10k/blob/439193522244720b86d8c81cde2e51e3a4d150cf/LICENSE | "Apache License Version 2.0" |
| SAM 3D Objects | facebook/sam-3d-objects | SA-V | permissive | https://github.com/facebookresearch/sam2/blob/2b90b9f5ceec907a1c18123530e92e794ad901a4/sav_dataset/README.md | "The videos and annotations in SA-V Dataset are released under CC BY 4.0." |
| SAM 3D Objects | facebook/sam-3d-objects | Ego4D | permissive | https://ego4d-data.org/pdfs/Ego4D-Licenses-Draft.pdf | "train ... any ... machine learning model ... designed to ... model or reconstruct at least one three-dimensional object ... commercial or noncommercial product development" |
| SAM 3D Objects | facebook/sam-3d-objects | BDD100K | non-commercial-terms | https://github.com/bdd100k/bdd100k/blob/9ac17c6c7c51d2fc83065fccd707cd5b1882a293/doc/source/license.rst | "for educational, research, and not-for-profit purposes"; commercial only "to BDD and BAIR Commons members and their affiliates" |
| SAM 3D Objects | facebook/sam-3d-objects | SA-1B, Ego-Exo4D, AEA, AEO, Nymeria, Food-101, MetaCLIP data, MITL/Art-3DO | unknown | — | SA-1B: "accepted the terms of the SA-1B Dataset Research License" (text not readable) |
Could not establish#
- TRELLIS.2 per-object lists: no ID lists at commit 75fbf018
(
data_toolkit/datasets/loader modules imported bybuild_metadata.pyare absent from the tree); whether its ObjaverseXL-sketchfab 168,307 equals TRELLIS-500K's cannot be verified; its TexVerse 503,387 subset cannot be measured. Paper text says "about 800K" while Table 6 says 976,736. - Which TRELLIS weight used what: the three TRELLIS model cards carry no training-data statement; paper describes one 500K set for all.
- TripoSG: sources are "such as"/"e.g."; Objaverse-XL sub-sources, whether ShapeNet was actually used, and the object list are not published.
- Step1X-3D: ~1.2M of the stated 2M assets (arithmetic: 2M − 320K − 480K) are outside released lists — ABO/3D-FUTURE/Thingiverse/internal subsets unnamed in size; "internal created data"/"proprietary collections" have no terms. 1,423 of the 322,883 Objaverse paths and 280 of the 32,550 texture paths did not match Objaverse-XL Sketchfab metadata (licence unknown).
- Thingiverse terms: https://www.thingiverse.com/legal, /legal/terms, /terms are client-rendered (Next.js shell); WebFetch saw only the title. Not readable.
- SkinTokens: per-object splits unreleased ("will be released separately in a future update"); sizes of the VRoid Hub and ModelsResource portions not stated; whether VRoid models were chosen by per-model Conditions of Use or with poster permission not stated; generator of the GRPO "AI-generated meshes" not named.
- The Models Resource site terms/ownership statement: every
models-resource.com URL (and sister sites) returns a Cloudflare 403 challenge
to curl/WebFetch; Internet Archive returned 429 then "Temporarily Offline";
Claude-in-Chrome extension not connected. Neither RigNet (2005.00559v2), Xu
et al. 2019 (1908.08506v1) nor SkinTokens states any licence for these
models. Search-index page titles show game-console/game-title organisation,
but ownership/no-licence is not established from a readable primary source →
verdict left
unknown. - VRoid Hub: the English text is a reference translation ("the original Japanese version shall prevail"; Japanese not read). Whether any per-model "Other Conditions" address AI training was not checked (would require reading individual model pages). The AI clause is conditional on a Company determination.
- HSSD card body (hssd/hssd-models) is gated (HTTP 401); only API cardData and gate prompt read.
- 3D-FUTURE Tianchi landing page is JS-rendered; the separately hosted "Data Sets Use License Agreement" (version March 28, 2020) was read instead.
- TEXGen: no UID list; released flow-matching checkpoints' training data not separately stated.
- Paint3D: 639 of 105,000 train UIDs did not match Objaverse-XL Sketchfab metadata (licence unknown).
- SAM 3D Objects: gated HF card body not readable (401); "licensed datasets" unnamed; Objaverse-XL sub-sources unstated; SA-1B licence text (ai.meta.com → HTTP 400), Ego-Exo4D licence (not fetched), Aria AEA/AEO/Nymeria licences (projectaria.com → HTTP 500), Food-101 terms (not on landing page), MetaCLIP image data (not identified) — all unread.
- Licence values measured come from Objaverse-XL's 2023 metadata snapshot, not from live Sketchfab pages; creators may have changed licences since.
- Objaverse-XL GitHub
license=None meaning: measured None shares are 87.15% (TRELLIS-500K github), 86.94% (Step1X-3D 480K), 88.59% (Articulation-XL2.0 github), 73.10% (whole github.parquet). Whether None means "no licence file in the repo" (i.e. all rights reserved by default) or merely "not detected", and whether values are repository- or file-level, is not documented in objaverse/xl/github.py @ 6996a3861b75abb0e0bb036a0766cf3263c708cd or the Objaverse-XL card/paper — so these rows staymixed-per-object, notthird-party-copyright-unlicensed. Small unmatched remainders: 30 (TRELLIS github), 20 (Step1X github).