View on GitHub

Py-Feat v1.0 licensing

This document covers the classic model family described by Cheong et al. (2023), now exposed as Detectorv1. The family name is not a claim that a PyPI package version 1.0 was released. Reviewed October 8, 2026.

Software

Py-Feat's original source code and documentation are licensed under the MIT License, including commercial use. Third-party source files retain their own copyright notices and licenses. This notice does not add a research-only restriction to the software.

Models

There is no single license covering every Detectorv1 configuration. Downloaded weights, fitted PCA/scaler/PLS parameters, datasets, and software must be considered separately. The repository's MIT license does not grant rights that belong to model or dataset providers. Existing valid license grants are not revoked by this clarification.

Component / Hugging Face repository	Software / upstream notice	Weight-specific position
retinaface, retinaface_r34	RetinaFace MIT	Existing model cards declare MIT; WIDER FACE data has separate noncommercial terms. A code license alone does not establish all weight rights.
img2pose	CC BY-NC 4.0	Preserve the upstream noncommercial license and attribution; not MIT or BSD-3.
mobilefacenet, mobilenet, pfld	Upstream implementation and checkpoints	A sufficiently explicit upstream weight redistribution grant has not been established in this review. Do not describe these as relicensed under Py-Feat's MIT license.
xgb_au, svm_au	Py-Feat integration: MIT	Previously labeled MIT; training agreements require separate review of trained-parameter redistribution. The MIT label is not a grant of dataset-provider rights.
svm_emo	Py-Feat integration: MIT	Previously labeled MIT; ExpW, CK+, and JAFFE training-data permissions must be considered separately.
resmasknet	ResidualMaskingNetwork MIT	Preserve the upstream MIT notice; FER2013 provenance is separate from the implementation license.
facenet	facenet-pytorch MIT	The shipped checkpoint uses VGGFace2. Its data terms must be assessed separately; selecting FaceNet does not by itself establish commercial clearance.
arcface_r50	InsightFace code MIT; pretrained-model terms	Pretrained weights are for noncommercial research. Conversion to safetensors does not change that.
l2cs	L2CS-Net MIT	The shipped Gaze360 checkpoint has an additional source license expressly excluding commercial applications of models trained on Gaze360; sharing scope needs review. MPIIFaceGaze is a separate checkpoint configuration.
pose_mlp_v2	Py-Feat implementation: MIT	Distilled from img2pose using CelebV-HQ. The previous card's BSD-3 teacher claim was incorrect. Student-weight permissions require their own assessment.

The paper-era FaceBoxes and MTCNN options have their own 3DDFA_V2 and MTCNN source notices. They are historical components, not additional current Detectorv1 options.

Training, validation, and visualization

The 2023 paper and supplementary materials describe these distinct uses:

Artifact	Training	Validation / evaluation
Classic SVM and XGBoost AU detectors	The supplement explicitly names BP4D, BP4D+, DISFA, CK+, and UNBC-McMaster Shoulder Pain; current model documentation additionally names Aff-Wild2.	Three-fold cross-validation for tuning; DISFA+ for held-out AU evaluation.
Emotion SVM	ExpW, CK+, JAFFE	Three-fold cross-validation for tuning; separate paper benchmarks. JAFFE is training data for this artifact.
Original AU-to-68-landmark visualization PLS	EmotioNet, BP4D, DISFA+	Three-fold cross-validation. DISFA+ is not globally evaluation-only across Py-Feat artifacts.
Later au_to_landmarks, bs_to_au, and landmarks68_to_mesh478 mappings	CelebV-HQ with detector/teacher-generated targets	See the exact checkpoint's model card; these are not the original paper's PLS weights.

The supplement's AU-training sentence ends with references 99-101 without a dataset name, and its training-dataset section also describes EmotioNet. The exact AU classifier mixture therefore requires a checkpoint-specific manifest; the list above must not be read as exhaustive. Separate training runs and file-format conversions must retain their own provenance. For dataset conditions and unresolved rights, see the dataset register.

Distribution status

The reviewed agreements do not establish unrestricted public redistribution or commercial use for every trained artifact above. In particular, the Aff-Wild2 agreements and the BP4D/BP4D+ parameter-set clauses require review against the relevant training date and checkpoint. Provider permissions, amendments, and renewals must be checked alongside the original agreements.

This document records scope and provenance; it is not a new sublicense for third-party datasets or weights and does not assert that dataset terms automatically attach to every trained model or inference output.