Py-Feat v2.0 licensing¶

This document covers the multitask Detectorv2 family. The detailed dataset inventory refers to face_multitask_v28.safetensors in py-feat/face_multitask_v2, not every older checkpoint in that repository. Reviewed October 8, 2026.

Software¶

Py-Feat's original source code and documentation remain under the MIT License, including commercial use. Third-party code retains its own notices. The noncommercial restrictions described below concern particular pretrained assets and data permissions, not a blanket restriction on the Py-Feat software.

Pretrained weights and upstream components¶

The multitask weights have been designated for noncommercial research. They cannot be represented as MIT-only:

The backbone is initialized from convnextv2_tiny.fcmae_ft_in22k_in1k, whose pretrained weights are CC BY-NC 4.0. The ConvNeXt-V2 code license is MIT and explicitly distinguishes the ImageNet weights. The multitask checkpoint modifies the pretrained network through joint fine-tuning and adds task heads. Preserve the upstream attribution and license notice.
Gaze360's research license expressly excludes commercial applications of models trained on its data. Its limits on use and redistribution also require review for public sharing. ETH-XGaze's access conditions expressly prohibit training models for commercial products. These are additional restrictions beyond the backbone license.
Geometry targets are distilled from MediaPipe and img2pose. img2pose is CC BY-NC 4.0. Teacher licensing and permission to distribute student weights require separate consideration; distillation is not a blanket permission to relicense.
The AU architecture adapts ME-GraphAU. Its upstream source terms must be retained for any copied code; architecture attribution does not establish a license for checkpoints or training data.
RetinaFace and optional identity weights are separate downloads with their own component notices. ArcFace is not the source of the multitask network's licensing restrictions. Disabling identity extraction does not remove the network's own upstream restrictions.

A research-only label is not, by itself, authorization for unrestricted public distribution. Aff-Wild2's newer agreement explicitly regulates trained-model sharing; BP4D and BP4D+ contain provisions concerning parameter sets and algorithms. Their applicability and any necessary permission must be resolved for the particular checkpoint. This notice does not grant rights beyond those available from the relevant holders.

Commercial-use inquiries¶

The current multitask checkpoints are not offered as commercially cleared weights. Py-Feat does not broker third-party permissions or offer an additional commercial sublicense that overrides upstream restrictions. The MIT software grant and existing valid component licenses are unchanged.

Users seeking commercial use must assess the exact assets and applicable terms, and obtain any necessary permissions directly from the relevant rights holders. The dataset register links provider sources; the backbone and teacher sources are listed above. A dataset download approval is not necessarily permission for commercial use of an already trained checkpoint. Py-Feat cannot certify that independently obtained permissions resolve every applicable right.

Alternatively, use independently cleared weights with the MIT software. The commercial retraining table identifies the remaining work for a future release. Downstream permissions do not by themselves resolve the distributor's own agreement obligations or weight sharing authority.

Current checkpoint's dataset roles¶

The current manuscript's training tables distinguish these sources:

Use	Corpora
Stage 1 geometry distillation	CelebV-HQ, BP4D, BP4D+, DISFA, EmotioNet, CK+, UNBC-McMaster Shoulder Pain
Stages 2-3 AU supervision	BP4D, BP4D+, DISFA, EmotioNet, CK+, UNBC-McMaster Shoulder Pain, AM-FED, Aff-Wild2 AU
Stages 2-3 emotion supervision	AffectNet, RAF-DB, FER+, ExpW, MELD
Stages 2-3 valence/arousal supervision	AffectNet, Aff-Wild2 V/A, AFEW-VA
Stages 2-3 gaze supervision	ETH-XGaze, Gaze360, MPIIGaze, Columbia Gaze
No gradient training for this checkpoint	DISFA+, FacePlace, 300W, WFLW, AFLW, EYEDIAP
Excluded duplicate corpus	MPIIFaceGaze (overlaps MPIIGaze)

Aff-Wild2's AU and V/A tracks are separate training pools from one dataset. Each dataset's train/validation/test partitions must also be distinguished: a held-out split does not mean the whole corpus was excluded from training. The manuscript reports selection among six candidate checkpoints using benchmark scores, including DISFA+ and EYEDIAP. These corpora are excluded from gradient training but are not untouched by model selection.

The stage-1 and stage-2/3 training counts are 944,248 and 1,568,745 rows, respectively. These counts and current-checkpoint roles come from the manuscript's Tables 3 and S18. The inspected local training repository confirms relevant source exclusions but does not contain a complete frozen v2.8 run manifest. Earlier _v2, _v26, and _v27 files require their own inventory; this notice does not retroactively assign them the v2.8 split.

See the dataset register for individual source terms. Validation-only use does not automatically relicense an otherwise unmodified model. Data access, evaluation, model selection, and publication still must satisfy the applicable agreement.

Visualization models¶

The later au_to_mesh, emotion_to_mesh, and bs_to_mesh PLS artifacts use CelebV-HQ frames and detector-generated targets; the v6 cards identify Detectorv2 v2.8 as the teacher. Their fitted parameters have their own provenance. Neither the software MIT license nor an assumption that all teacher restrictions automatically transfer establishes their complete distribution permissions. Existing valid grants are not revoked here.