fffiloni/idol-community-assets

IDOL Commons

A community collection of 3D assets: every contribution gives someone else a starting point for a scene, a game or an experiment. Explore the objects, credit their contributors and share what you make.

Development status: contributions are enabled in the protected development Space. The catalogue is the authoritative list of published assets; illustrative schema examples are not contributions. The application is being developed separately in IDOL Commons.

What is in the dataset?

The original, validated 3D Gaussian splat PLY from IDOL, transparent previews rendered from that result, contributor-authored descriptions and labels, verified HF account attribution, generation settings and server-recorded provenance. These PLY files are not polygon meshes or GLB files. Source images are not distributed. Text-to-image prompts are private by default; from schema 1.3.0, an author may explicitly opt in to publish the original prompt with the asset.

V1 accepts results produced with the application's pinned IDOL pipeline, including the maintainer-requested GPU batch described below. A valid PLY is not, by itself, proof of IDOL provenance, originality, ownership or lawful use. HF sign-in verifies the contributing account at publication; it does not verify that account's rights in the depicted subject. Content is not represented as legally cleared or individually reviewed by the maintainer.

IDOL demo batch

The maintainer requested generation of all 94 examples in the application's pinned IDOL catalogue, using their original presets. The inputs come from SperidLabs' IDOL demo at revision 2fd68f6aaa69df7c178503a788ad8deb4c7d3117. Generation runs through the same pinned application pipeline in an authenticated Hugging Face GPU job. The batch checkpoint records progress and any failures; this paragraph does not imply that the entire batch has finished.

New records are contributed by @fffiloni with Apache-2.0 as their preferred licence, source credits, modification notices and the full licence text. The upstream repository declares Apache-2.0; third-party rights in depicted subjects remain applicable. Existing identified examples are reused, with earlier valid licence grants retained. Source images are not redistributed.

Each example is generated, validated, rendered and committed atomically. Provenance records the exact model, pipeline, encoder and application revisions, runtime packages and generation parameters. The reproducible operator script is tools/batch_examples.py in the development Space. This owner-requested batch authenticates the contributing account directly through the Hub; it does not impersonate a browser OAuth session.

The contributor can edit these records in My contributions. Verified moderators can use Moderation → All contributions to edit details and preview angles, hide or restore objects, and remove contributions from the current dataset. Moderator edits preserve original authorship and are recorded in the private moderation audit.

Licences are per asset

Each record identifies the preferred licence and, for schema 1.2.0, any earlier licence grants. The whole collection is not blanket-licensed under a single asset licence. Read the asset metadata, LICENSES.txt and NOTICE.txt before reuse. The maintainer offers only their own rights in the compilation and original documentation under CC BY 4.0. Model weights, software, upstream works and third-party rights have their own terms.

Licence	Main condition
CC BY 4.0	Attribution and change notices.
CC BY-SA 4.0	Attribution and share-alike for adaptations.
CC0 1.0	Dedication/fallback licence for rights the contributor can waive.
Apache 2.0	Licence copy, copyright/change notices and applicable NOTICE obligations.
MIT	Copyright and permission notices retained.

Contributors offer only rights they are entitled to grant in the object, rendered previews and editorial metadata. A licence does not create copyright in unprotectable AI output, clear third-party copyright, trademarks, privacy or personality rights, or imply endorsement. A verified HF account and a technical file check are not legal clearance. Source permissions and intended use must be reviewed independently.

Existing records keep their original grants. Adding a preferred licence does not revoke permissions already validly granted; old source notices remain in the grant history. Later withdrawal cannot recall lawful copies. For each release, use its revision-pinned metadata and notices. Public-domain material remains public domain.

A useful attribution includes the title, verified contributor, source record URL, licence link, modifications and required upstream notices. Basket packages include the legal documents for 1.2.0 records; downloading the PLY alone does not embed those documents. See LICENSE.md.

Contributing responsibly

Read CONTRIBUTING.md and the application's publication terms. Publication is a separate, affirmative action after inspecting the result. The form records a rights basis, public credits where required, and confirmation that the result was reviewed for third-party rights and inappropriate content. These are contributor declarations, not an automated rights audit.

Only submit content you can share under your selected licence and applicable upstream terms. Owning a photograph does not necessarily give you rights in a logo, character, sculpture, design or identifiable person it depicts. Generating an image with AI does not resolve those issues either. Do not upload private information, non-consensual likenesses, infringing material or content prohibited by HF's Content Policy. Source material with incompatible restrictions needs separate permission; naming its licence is not enough.

Provenance and schema

New web contributions use 1.3.0; strict 1.0.0 / 1.1.0 / 1.2.0 records remain readable without inventing missing rights or image-origin information. The collection manifest envelope remains 1.0.0. See schema/record.schema.json, schema/publish-request.schema.json and the complete, explicitly fictional example record.

The server derives the stable ID, account, date, paths, byte counts, SHA-256 hashes, IDOL model/pipeline revisions, encoder revisions, actual parameters and preview camera. Records from 1.1 distinguish an uploaded input from an image generated by the app's pinned Z-Image-Turbo model. For an upload, an external model or prompt is not guessed. For app-generated images, the actual image model, revision and numeric settings are recorded; prompt text remains private by default. From schema and publication terms 1.3.0, contributors can separately opt in to publish the actual original Z-Image-Turbo prompt in provenance.source_image.prompt, with prompt_public: true. The app displays this text in the object information panel. Existing private prompts are not retroactively published; an expired or missing prompt cannot be reconstructed.

All levels forbid unknown fields, use strict types and bound lengths/ranges. Categories are controlled. Additional server checks enforce identity, ownership, cross-field consistency and real file content. Neither JSON Schema nor binary validation is a content, provenance or legal review on its own.

Files and publication consistency

collection.json                     versioned catalogue with stable sequence positions
data/records.jsonl                  Hub export with derived splat URI
assets/ab/<uuid>/metadata.json      per-asset canonical record
assets/ab/<uuid>/LICENSES.txt       full applicable licence texts (1.2+)
assets/ab/<uuid>/NOTICE.txt         source/copyright/change notices (1.2+)
assets/ab/<uuid>/model.ply          original validated IDOL result
assets/ab/<uuid>/preview.png        transparent 1024 × 1024 render
assets/ab/<uuid>/preview-small.png  512 × 512 derivative of that render
runtimes/<sha256>.txt               installed generation-runtime package versions

Files, metadata and catalogue become visible together in one Hub commit, with parent-revision conflict checks. The viewer refreshes the catalogue and uses revision-pinned file URLs. Contributions are account-limited and identical 3D bytes are deduplicated. The current pilot caps the active catalogue at 10,000 entries; a larger collection requires sharded storage and paginated discovery.

Community signals

Optional aggregate metrics are in social/stats.json (versioned envelope), social/stats.jsonl (one row per asset) and social/stats.schema.json. Join them to canonical records using asset_id. They update asynchronously, usually within a minute when the Space is running, without modifying asset records.

• likes: current positive votes, at most one per verified HF account.
• basket_adds: cumulative app additions, deduplicated per account or anonymous session, asset and UTC day. This is not current basket occupancy.
• download_requests: app download/ZIP/script/command requests, with the same daily deduplication. This does not prove completion and excludes direct Hub downloads. Failed or cancelled requests may still count.

Only totals are public, never voter identities or visitor event logs. Anonymous sessions can reset, and client requests can be missed, so activity totals are estimates rather than unique-person counts. These metrics start at feature activation; past activity is not inferred. Withdrawn assets leave current metric snapshots; repository history can retain earlier totals.

Reports, corrections and withdrawal

Use Report on a selected object for rights, privacy, inappropriate content or other content concerns. To withdraw your own asset, open My contributions → Withdraw contribution. App reports are private to the maintainer and require HF sign-in to limit abuse. An accepted report immediately hides the object from the public app viewer and baskets pending review. Its author can still inspect it in My contributions, and the maintainer can review it in Moderation. Hiding does not delete the public dataset files; removal is a separate action. Include the asset ID and a concise explanation; avoid uploading sensitive evidence into public discussions or the report text. See RIGHTS-AND-REMOVAL.md.

You can also use HF's reporting channels, including its safety and copyright-notice channels, independently of app sign-in. HF repository reports are public; use the appropriate private channel for sensitive material. The dataset maintainer is @fffiloni.

Current catalogue entries and their files can be withdrawn atomically. Earlier Git commits, forks, caches and downloaded copies may remain. Validly granted licence permissions are not revoked by later withdrawal; a contributor cannot grant rights they never held. Sensitive cases may require repository-history remediation and HF intervention. Complete erasure, automatic legal clearance and a guaranteed response deadline are not promised.

Authors can also use Edit thumbnail from their object or My contributions to choose a new public preview angle. The app re-renders the actual PLY and updates the previews and framing atomically; the 3D file, authorship, provenance and licence remain unchanged. Thumbnail edits are paused during moderation review.

Model and software credits

Component	Source / licence scope
IDOL	speridlabs/idol, model revision 2731c2291fc65a797c5cb627a5e5526fc1490902; publisher declares Apache-2.0
Optional input image generation	Tongyi-MAI/Z-Image-Turbo, revision f332072aa78be7aecdf3ee76d5c247082da564a6; publisher declares Apache-2.0
IDOL image encoders	DINOv3 under Meta's DINOv3 licence; FLUX.2 autoencoder identified separately as Apache-2.0 by Black Forest Labs
Collection interface	Adapted from Scenario explorer by fffiloni, Apache-2.0; native splats use Spark, MIT

The upstream IDOL demo and model declare Apache-2.0; no individual Creative Commons grant or complete rights clearance has been verified for the demo inputs. The app does not treat an example as the contributor’s own source creation or automatically prove permission from the repository licence. Do not assume an example or depicted subject is cleared for redistribution.

No model weights are included in this dataset. Model licences do not automatically license every depicted subject or generated output. The development source contains a separate component inventory and outstanding release checks; it does not claim that unresolved upstream provenance has been cleared.

Authors may edit titles, descriptions, categories, tags, public credits and the preferred licence in My contributions → Edit details. Rights changes require confirmation. Additional licence grants preserve earlier grants and notices; editing does not restore an object pending moderation.

Bilingual catalogue metadata

All 94 IDOL demo examples have English and French visual descriptions, descriptive tags and reviewed categories. Their geometry, authorship, provenance and licence grants are unchanged. See the editorial update report. Existing custom titles are retained.

For new contributions, the app proposes editable metadata from the rendered 3D preview with Qwen/Qwen3-VL-4B-Instruct, pinned revision ebb281ec70b05090aa6165b016eac8ec08e71b17 (publisher declares Apache-2.0). Suggestions can be inaccurate and must be reviewed before publication. They do not infer permissions or replace licence consent. Known IDOL demo inputs reuse the reviewed metadata table. The original image and prompt are not sent to an external inference provider.

The viewer uses bilingual catalogue text plus a bounded French/English synonym dictionary, accent normalization and prefix/fuzzy matching. This improves French discovery but is not an unrestricted semantic translation service.

Hub 3D gallery export

The top-level splat column is a derived Mesh feature for the Hugging Face dataset viewer. It references the existing public PLY using hf://datasets/OWNER/REPO@main/assets/ab/UUID/model.ply. The asset remains a 3D Gaussian splat, not a polygon mesh. No source image is added to this export.

collection.json and each metadata.json retain their strict canonical schema; only data/records.jsonl contains the extra column. Every publication, metadata edit, preview edit and withdrawal must regenerate this export through dataset_export.records_jsonl. Operator batches must ship the exporter from the same reviewed checkout as the batch script; historical scripts that write raw canonical rows must be updated before use. Keep dataset_info.features aligned with future canonical fields and test the viewer after schema changes.

Gallery links follow main. For reproducible downloads, use the revision-pinned canonical metadata and its SHA-256 checksums. The gallery shows public dataset records; private moderation hiding in the Space does not hide dataset files.

The gallery integration was proposed by Caleb Fahlgren (@cfahlgren1) in dataset PR #2.

Python load_dataset consumers need a datasets build that supports the new Mesh feature. The PyPI release 4.8.4 does not support it; the integration was tested with upstream revision a1269ca9dbb6e0858a9c375e8982329a3fcc84af. Use Mesh(decode=False) when you want original file paths/bytes instead of automatic trimesh decoding. Raw JSONL readers and the Space do not depend on this optional Python decoder. See the official mesh loading guide.