Datasetpublic

fernandotonon/QtMeshEditor-motion-corpus

QtMeshEditor Motion Corpus A permissively-licensed animated-humanoid corpus: rigged 3D characters with skeletal animation clips, harvested for QtMeshEditor's text-to-motion v2 work (epic #837) — the template clip library and the training set for a from-scratch flow-matching motion model. Every asset is CC0 or CC-BY — nothing here derives from Mixamo, LAFAN1, Bandai-Namco, AMASS/HumanML3D, or game rips (all license-poisoned for commercial redistribution). That makes this corpus —… See the full description on the dataset page: https://huggingface.co/datasets/fernandotonon/QtMeshEditor-motion-corpus.

sourceHugging Faceotherupdated 23d agoView on Hugging Face

1likes1.2kdownloads

Dataset Card

QtMeshEditor Motion Corpus

A permissively-licensed animated-humanoid corpus: rigged 3D characters with skeletal animation clips, harvested for QtMeshEditor's text-to-motion v2 work (epic #837) — the template clip library and the training set for a from-scratch flow-matching motion model.

Every asset is CC0 or CC-BY — nothing here derives from Mixamo, LAFAN1, Bandai-Namco, AMASS/HumanML3D, or game rips (all license-poisoned for commercial redistribution). That makes this corpus — and anything trained on it — redistributable, including commercially, with the attributions in `ATTRIBUTION.md`.

Layout

manifest.json          per-asset provenance: source, source_url, title,
                       author, license, tags, files, and a validation record
                       (bone count, animation names, vertex count) produced
                       by the editor's own loader (`qtmesh info --json`)
ATTRIBUTION.md         CC-BY credits — ship this with derived works
raw/packs/...          Quaternius CC0 animated packs (FBX)
raw/opengameart/...    OpenGameArt CC0/CC-BY rigs (mixed formats)
raw/sketchfab/...      Sketchfab CC0/CC-BY downloads (glTF), when present
derived/
  motion-library-v2.json   the 150-clip CURATED LIBRARY built from raw/ —
                       per-clip canonical 22-joint world-frame quats. This is
                       the exact training source for the v8.0 text-to-motion
                       model, and the app also ships it as the template-clip
                       fallback for prompts outside the model vocabulary.

The derived library

derived/motion-library-v2.json is what a consumer most likely wants: the raw assets already parsed onto one canonical skeleton, gated for quality, and labelled by action. 150 clips from 31 source assets — 27 CC-BY-4.0 (credited in ATTRIBUTION.md), CC0 material (largely Quaternius packs), and 17 clips owned by the author.

Rebuild it from raw/ with `scripts/build-motion-library-v5.py`, or train on it directly — see fernandotonon/QtMeshEditor-t2m for the model, its metrics, and the exact prep/train commands.

manifest.json also records discovery-only rows (license-cleared Sketchfab models not yet downloaded) — their source_url can be fetched with the scraper below.

Reproducing / extending

The corpus is produced by `scripts/scrape-motion-corpus.py` (license-filtered discovery + download + validation; incremental manifest):

python3 scripts/scrape-motion-corpus.py --out motion_corpus --packs \
    --opengameart "animated character" \
    --sketchfab "walk animation,run animation"   # SKETCHFAB_API_TOKEN to download

Re-running against this dataset's directory merges new finds into the manifest instead of overwriting it.

Provenance & licensing

Per-asset license and author live in manifest.json; the license filter accepts only CC0-1.0, CC-BY-3.0 and CC-BY-4.0 (exact-label matching — no NonCommercial/NoDerivs/ShareAlike variants). If you redistribute the corpus or ship anything derived from it (clip libraries, model weights), include ATTRIBUTION.md.