AuraFace-v1 face embedding (INT8 quantized ONNX)

File: glintr100_int8.onnx (104.6 MB) — INT8-quantized face-recognition backbone outputting a 512-dimensional embedding from a 112×112 aligned face.

Source & provenance

Base model: AuraFace-v1 (fal/AuraFace-v1), ResNet-100 + Additive Angular Margin Loss (ArcFace, arXiv:1801.07698), licensed Apache-2.0. Copyright remains with the AuraFace authors.
Only the quantized file in this repo is distributed here. The original FP32 file glintr100.onnx (~260 MB) is not included — get it from fal/AuraFace-v1 if needed.
Reported AuraFace-v1 benchmarks: LFW 0.99650, CFP-FP 0.95186, AGEDB 0.96100, CALFW 0.94700, CPLFW 0.90933.

What changed (Apache-2.0 §4 modification notice)

The FP32 glintr100.onnx was quantized to INT8 with onnxruntime quantization (per-channel weight quantization; compute graph otherwise unchanged): same input data [N, 3, 112, 112] float32, same output 1333 [1, 512]. File size: ~260 MB → ~105 MB. Expect a small accuracy drop versus FP32, typical for INT8 quantization.

License

Licensed under the Apache License 2.0 (same as the base model). You may use, modify, and distribute this file provided you preserve attribution to the AuraFace authors, keep this modification notice, and comply with the full license text.

Model spec

Item	Value
Task	Face feature extraction → 512-d embedding (cosine similarity for recognition)
Input `data`	`[N, 3, 112, 112]`, float32, aligned RGB face, normalized per InsightFace convention
Output `1333`	`[N, 512]`, float32 (L2-normalize before comparing)

Usage

import onnxruntime as ort
import numpy as np

sess = ort.InferenceSession("glintr100_int8.onnx",
                            providers=["CPUExecutionProvider"])
face = np.random.rand(1, 3, 112, 112).astype(np.float32)  # aligned + normalized face
emb = sess.run(None, {"data": face})[0]
emb = emb / np.linalg.norm(emb, axis=1, keepdims=True)     # cosine-sim ready