Every motion source, its licence, and which mode it is allowed in.
The motion she plays at home is cut from datasets released for research and non-commercial use, and from footage that may not be redistributed. None of it is displayed on this site, distributed, or reachable, and models trained on it never leave the machine she runs on. That is what "private forever" means: not a setting, a property of where the data can be. In public mode the borrowed corpus is closed and only authored and heuristic layers move her, so nothing on a stream is derived from a source that forbids it.
Non-commercial data may play in private. It is never training input for anything public. A gate can refuse a clip; nothing can un-train a weight, so the line is drawn at the training set, not at the output.
All 35 sources in the registry, about 5220.6 hours of motion, with each licence as the source states it and the mode it is allowed in. "Private (by choice)" marks sources whose licence would allow more; they stay private because everything trained alongside them is.
| Source | Licence | Format | Size | Mode | Why it is here |
|---|---|---|---|---|---|
| Talking With Hands 16.2M (Meta) | CC BY-NC 4.0 | bvh+audio | 20 h | private | two-person conversation with fingers |
| ZeroEGGS (Ubisoft) | research | bvh+audio | 2 h | private | 19 styles, one actor; already partly in the corpus |
| BEAT2 / EMAGE (MPI) | CC BY-NC 4.0 | smplx+audio | 60 h | private | the big one, 20 GB on HF; partly in the corpus already |
| BEAT v1 (BVH) | CC BY-NC 4.0 | bvh+audio | 76 h | private | same recordings as BEAT2 in BVH with facial blendshapes |
| TalkSHOW / SHOW | research | smplx+audio | 27 h | private | talk-show hosts fitted from video: real conversational gesture at a desk |
| Trinity Speech-Gesture | research | bvh+audio | 4 h | private | GENEA 2020/2022 base data |
| GENEA Challenge 2023 data (TWH-based, dyadic) | CC BY-NC 4.0 | bvh+audio+tsv | 18 h | private | cleaned TWH with transcripts and the interlocutor |
| PATS (CMU, 2D poses of 25 speakers) | research | 2d-keypoints+audio | 251 h | private | huge but 2D; needs lifting before it is any use |
| TED Expressive (HA2G) | research | 3d-keypoints+audio | 27 h | private | lifted TED talks; upper body only |
| Seamless Interaction (Meta) | CC BY-NC 4.0 | video+smplx codes | 4000 h | private | 27 TB of tar shards; seamless_stream.py keeps everything but the video |
| DnD Group Gesture (Aalto) | research | bvh+audio | 6 h | private | five-person tabletop conversation |
| Motion-X (IDEA) | research | smplx | 144 h | private | 81k sequences incl. in-the-wild video |
| AMASS (MPI) | research | smplx/smplh npz | 40 h | private | the mocap library; HumanML3D is its labelled subset |
| HumanML3D | research | amass subset + text | 29 h | private | text labels for AMASS motions (needs AMASS) |
| LAFAN1 (Ubisoft) | research | bvh | 4.6 h | private | locomotion, clean, five subjects |
| Bandai Namco Research Motiondataset | CC BY-NC-ND 4.0 | bvh | 3 h | private | styled everyday motion incl. gestures |
| 100STYLE | research | bvh | 4 h | private | 100 walking styles; idle/character (CC BY 4.0) |
| CMU Motion Capture (BVH conversion) | free | bvh | 9 h | private (by choice) | 2,500 clips, the classic (Hahne BVH port) |
| SFU Motion Capture | research | bvh | 2 h | private | clean everyday motion; per-clip .bvh links under nusmocap/ |
| KIT Motion-Language | research | bvh/mmm + text | 11 h | private | text-described motions |
| Motorica Dance | research | bvh+audio | 6 h | private | dance; rhythm and weight for the reflex layer |
| AIST++ | research | smpl+audio | 5 h | private | dance; SMPL motions 306 MB + 3D keypoints (terms of use accepted by pax for private use) |
| Learning2Listen (Ng et al. 2022) | research | DECA face+head coeffs + mel audio | 72 h | private | LISTENER head and expression aligned to the speaker: her nods while pax talks |
| ViCo listening-head | research | 3DMM coeffs + video | 2 h | private | speaker and listener 3DMM in conversation |
| IEMOCAP (USC) | research | face+head mocap markers + audio + emotion | 12 h | private | dyadic face and head markers with emotion labels |
| VOCASET (MPI) | research | 4D face scans + audio | 1 h | private | audio-driven mouth and jaw, the lipsync reference |
| MEAD emotional talking face | research | video, 60 actors, 8 emotions | 40 h | private | expression under emotion at three intensities |
| HDTF talking heads | research | YouTube video list + script | 16 h | private | high-res talking heads; video via yt-dlp, not automated |
| CelebV-HQ | research | YouTube video list + script | 68 h | private | 35k face clips with expression/action labels; video via yt-dlp, not automated |
| Gaze360 (MIT) | research | images + 3D gaze | 0 h | private | only if we ever train a gaze estimator; not animation data |
| InterHuman (InterGen) | research | smpl | 7 h | private | two-person interaction |
| Inter-X | research | smplx | 13 h | private | two-person with hands |
| InterAct (2025) | research | smplx | 241 h | private | daily two-person activities, large |
| Goliath-SC self-contact poses (ICCV 2025) | research | smplx poses | 0 h | private | 383K self-contact poses: exactly the fence's problem |
| TUCH / MTP self-contact (MPI) | research | smplx poses | 0 h | private | mimic-the-pose self-contact data |