globalnebula/insightface-buffalo-l-onnx

InsightFace buffalo_l (SCRFD + ArcFace), ONNX mirror

A mirror of the two ONNX models from the InsightFace buffalo_l pack used by the Piccy AI worker: SCRFD-10GF for detection (with 5-point landmarks) and ArcFace w600k_r50 for recognition (512-d embeddings). Both run on onnxruntime with no PyTorch. The weights are unchanged from upstream; this mirror just gives the build a stable, self-controlled download source.

Files

File	What it is	Notes
scrfd_10g.onnx	SCRFD-10GF detector	upstream det_10g.onnx, renamed
arcface_w600k_r50.onnx	ArcFace ResNet50 recognizer	upstream w600k_r50.onnx, renamed
buffalo_l.zip	zip of det_10g.onnx + w600k_r50.onnx	for the Docker build arg below

Use with the Piccy Docker build

Point the build at the zip in this repo (works with the existing INSIGHTFACE_BUFFALO_URL build arg, which extracts det_10g.onnx and w600k_r50.onnx):

INSIGHTFACE_BUFFALO_URL="https://huggingface.co/globalnebula/insightface-buffalo-l-onnx/resolve/main/buffalo_l.zip" \
  docker compose build ai-worker

Use directly / manual dev

Drop the two standalone files into the worker's models/ directory:

ai-worker/models/scrfd_10g.onnx
ai-worker/models/arcface_w600k_r50.onnx

Signatures

• SCRFD input: a BGR image; the detector handles resize + (x-127.5)/128 internally (see ai-worker/insight_face.py). Outputs boxes, per-face scores, and 5 landmarks.
• ArcFace input: (batch, 3, 112, 112), an ArcFace-aligned 112x112 crop normalized (x-127.5)/127.5. Output: (batch, 512) embedding; L2-normalize before cosine comparison.

Provenance & license

• Source: InsightFace buffalo_l model pack (https://github.com/deepinsight/insightface).
• InsightFace code is MIT. The pretrained weights derive from the WebFace600K / MS1M training data and carry their own dataset terms; review before production or commercial use.