yakhyo/uniface-weights

UniFace model weights

Weight mirror for UniFace. 62 files, primarily ONNX.

This repository is a mirror, not the primary source. UniFace downloads from GitHub Releases first and falls back here when GitHub is unreachable. Files are byte-identical to the release assets and are SHA256-verified by the library on download.

from uniface import SCRFD

# Downloads and verifies automatically; no token required.
detector = SCRFD()

Nothing to configure — the fallback is automatic, and no Hugging Face token is needed because this repository is public.

Filenames

Files are named by UniFace's internal model key, not by the original release asset name. Five asset basenames (mobilenetv2.onnx, mobilenetv3_small.onnx, mobilenetv3_large.onnx, resnet18.onnx, resnet34.onnx) appear in two different source repositories with different weights, so a mirror keyed on the original names could not hold all 62 models.

Licenses

The UniFace library is MIT. These weights are not. They come from many upstream projects under different terms, including copyleft and non-commercial-research-only licenses. Read this table before using any of them in a product.

This table records the license of the cited upstream source, verified against that project's own LICENSE file or README statement (see Verification). It is provenance information, not legal advice. Where a license could not be established from an authoritative statement, the row says so rather than guessing.

Permissive

Files	Family	Upstream source	License
retinaface_mnet025, retinaface_mnet050, retinaface_mnet_v1, retinaface_mnet_v2, retinaface_r18, retinaface_r34, retinaface_r50	RetinaFace	yakhyo/retinaface-pytorch	MIT
centerface	CenterFace	Star-Clouds/CenterFace	MIT
blazeface, face_mesh, face_landmarker	BlazeFace / Face Mesh	google-ai-edge/mediapipe via yakhyo/mediapipe-face-mesh-onnx	Apache-2.0
adaface_ir_18, adaface_ir_101	AdaFace	mk-minchul/AdaFace	MIT
edgeface_xxs, edgeface_xs_gamma_06, edgeface_s_gamma_05, edgeface_base	EdgeFace	otroshi/edgeface (Idiap Research Institute)	BSD-3-Clause
pipnet_r18_wflw_98, pipnet_r18_300w_celeba_68	PIPNet	jhb86253817/PIPNet	MIT
face_attrib_net	FaceAttribNet	qualcomm/ai-hub-models	BSD-3-Clause
parsing_resnet18, parsing_resnet34	BiSeNet face parsing	yakhyo/face-parsing	MIT
modnet_photographic, modnet_webcam	MODNet	ZHKKKe/MODNet	Apache-2.0 (README states code and models)
minifasnet_v1se, minifasnet_v2	MiniFASNet	minivision-ai/Silent-Face-Anti-Spoofing	Apache-2.0
ediffiqa_t, ediffiqa_s, ediffiqa_m, ediffiqa_l	eDifFIQA	LSIbabnikz/eDifFIQA	MIT
gaze_resnet18, gaze_resnet34, gaze_resnet50, gaze_mobilenetv2, gaze_mobileone_s0	MobileGaze	yakhyo/gaze-estimation	MIT
headpose_resnet18, headpose_resnet34, headpose_resnet50, headpose_mobilenetv2, headpose_mobilenetv3_small, headpose_mobilenetv3_large	6DRepNet-style head pose	yakhyo/head-pose-estimation	MIT

Non-commercial research only

InsightFace releases its code under MIT but states, twice in its README and again in its model zoo, that all pretrained models are for non-commercial research purposes only, and that this applies to auto-downloaded models. Do not use these six in a commercial product.

Files	Family	Upstream source	License
arcface_mnet, arcface_resnet	ArcFace	deepinsight/insightface	Non-commercial research only
scrfd_10g, scrfd_500m	SCRFD	deepinsight/insightface	Non-commercial research only
age_gender	Age/Gender	deepinsight/insightface	Non-commercial research only
2d_106	106-point landmarks	deepinsight/insightface	Non-commercial research only

Copyleft (GPL-3.0)

These derive from GPL-3.0 projects. GPL-3.0 imposes obligations on distribution of derived works that MIT does not, and its applicability to model weights is unsettled. Take advice before shipping these in a closed-source product.

Files	Family	Upstream source	License
yolov5n, yolov5s, yolov5m	YOLOv5-Face	deepcam-cn/yolov5-face	GPL-3.0
xseg	XSeg	iperov/DeepFaceLab	GPL-3.0

Unresolved — no license located

No authoritative license statement was found for these. Absent a license grant, assume no redistribution or use rights. They are listed here rather than assigned a guessed license.

Files	Family	Cited source	Status
yolov8_lite_s, yolov8n_face	YOLOv8-Face	yakhyo/yolov8-face-onnx-inference	No LICENSE file in the cited repository, and the original training source is not recorded. YOLOv8 derivatives are commonly AGPL-3.0 via Ultralytics, but that was not verified for these weights.
affecnet7, affecnet8	DDAMFN emotion	SainingZhang/DDAMFN	No LICENSE file and no license statement in the README.
fairface	FairFace	yakhyo/fairface-onnx	No LICENSE file. Upstream dchen236/FairFace has no LICENSE file; its README states CC BY 4.0 under the Data heading, which covers the dataset, not the model weights.
sphere20, sphere36	SphereFace	yakhyo/face-recognition	Repository is MIT, but the weights were trained on MS1M-V2. See Training-data restrictions.
mobilenetv1_025, mobilenetv2, mobilenetv3_small, mobilenetv3_large	MobileFace	yakhyo/face-recognition	Same as above: MIT repository, MS1M-V2 training data.

Training-data restrictions

A permissive repository license does not always carry over to weights, because the training data may be restricted independently. Two cases affect models above:

• MS1M-V2 is distributed by InsightFace, whose README states that the training data "and the models trained with these data are available for non-commercial research purposes only." On a plain reading this reaches the SphereFace and MobileFace weights.
• WIDER FACE, used by RetinaFace, SCRFD, CenterFace, YOLOv5-Face and YOLOv8-Face, is published for non-commercial research use.

These constraints come from the dataset publishers, not from the repositories mirrored here.

Verification

Licenses above were read from the upstream projects rather than inferred:

• GitHub's license API (GET /repos/{owner}/{repo}) for the SPDX identifier of each cited repository.
• deepinsight/insightface README ("The training data ... and the models trained with these data are available for non-commercial research purposes only") and model_zoo/README.md ("ALL models are available for non-commercial research purposes only.").
• ZHKKKe/MODNet README License section, which explicitly covers models as well as code.
• otroshi/edgeface LICENSE file, which is BSD-3-Clause text (GitHub reports NOASSERTION because of the copyright block format).
• qualcomm/ai-hub-models face_attrib_net/README.md, which points at the repository LICENSE.

If you own one of these upstreams and this table is wrong, please open an issue on UniFace and it will be corrected.

Citation

Cite the original authors of each model, not this mirror. Upstream repositories are linked in the tables above.