face-detect-gguf

GGUF model packs for the face-detect backend of LocalAI.

Each .gguf here is a self-contained, metadata-driven pack (detector + recognizer, plus genderage / anti-spoof heads when the source provides them) produced by face-detect.cpp, a standalone C++/ggml port of the insightface and OpenCV-Zoo face pipelines. No Python or ONNX runtime is needed at inference time: the GGUF carries the weights verbatim plus the forward-graph topology in its KV metadata, and the C++ engine replays it.

• Source code commit: 636a19631a400694a08edb7e707288003b7093aa (face-detect.cpp)
• Format: GGUF, general.architecture = "facedetect"
• Dtype: all packs published as f16 (near-lossless canonical; see parity below)
• Consumed by: LocalAI face-detect backend

License - read before use

The packs in this repo carry two different licenses depending on their source weights. Pick the pack that matches your use case.

Pack	Source	License	Commercial use
buffalo_l.gguf	insightface buffalo_l	Non-commercial, research-only	No
buffalo_m.gguf	insightface buffalo_m	Non-commercial, research-only	No
buffalo_s.gguf	insightface buffalo_s	Non-commercial, research-only	No
yunet-sface.gguf	OpenCV-Zoo YuNet + SFace	Apache-2.0	Yes

The insightface buffalo packs (SCRFD + ArcFace) are released by insightface for NON-COMMERCIAL research purposes only. They are redistributed here as derived GGUF artifacts under those same upstream terms. If you need a commercial-friendly option, use yunet-sface.gguf (Apache-2.0).

Models

buffalo_l.gguf - SCRFD det_10g + ArcFace ResNet50 (512-d)

The primary, highest-accuracy insightface pack. SCRFD det_10g detector + ArcFace w600k_r50 (IResNet50) recognizer producing a 512-d embedding, plus the genderage head and the MiniFASNet anti-spoof ensemble (V2@2.7 + V1SE@4.0, 80x80) bundled in. License: non-commercial / research-only.

buffalo_m.gguf - SCRFD det_2.5g + ArcFace ResNet50 (512-d)

Mid-size insightface pack. SCRFD det_2.5g detector + ArcFace w600k_r50 512-d recognizer (+ genderage + anti-spoof when present). The det_2.5g topology is replayed through the metadata-driven graph interpreter. License: non-commercial / research-only.

buffalo_s.gguf - SCRFD det_500m + ArcFace MobileFaceNet (512-d)

Smallest insightface pack. SCRFD det_500m detector + ArcFace MobileFaceNet (w600k_mbf) 512-d recognizer (+ genderage + anti-spoof when present). Both the det_500m detector and the MobileFaceNet recognizer are replayed metadata-driven. License: non-commercial / research-only.

yunet-sface.gguf - YuNet detector + SFace recognizer (128-d), Apache-2.0

The commercial-friendly alternative. OpenCV-Zoo YuNet (face_detection_yunet_2023mar) anchor-free detector + SFace (face_recognition_sface_2021dec) recognizer producing a 128-d embedding. SFace carries its (x-127.5)/128 normalization in-graph. License: Apache-2.0 (commercial use OK).

Parity

Each pack was validated against its reference pipeline (insightface for buffalo, cv2 FaceDetectorYN/FaceRecognizerSF for yunet+sface) before upload. The decode-isolated embedding gate is cosine >= 0.9999 and max|d| <= 1e-3.

Pack	Dtype	Embedding cosine (gate)	Result
buffalo_l.gguf	f16	1.000000	PASS
buffalo_m.gguf	f16	1.000000	PASS
buffalo_s.gguf	f16	1.000000	PASS
yunet-sface.gguf	f16	1.000000	PASS

f16 quantization is applied only to the large 2-D Gemm weights (the ArcFace / SFace embedding head); every conv kernel, BN stat, bias and projection head stays F32, so f16 is near-lossless. All four packs meet the strict near-lossless bound at f16.

Usage

These packs are intended to be installed through the LocalAI model gallery (face-detect-buffalo-l / -m / -s, and face-detect-yunet-sface) and run by the face-detect backend. See the LocalAI and face-detect.cpp documentation.