AuraFace-v1 (glintr100) — TFLite, int8 weights
On-device TFLite conversion of the glintr100.onnx face recognition model from
fal/AuraFace-v1 (Apache-2.0).
Only the recognition model is included. The other files bundled in the AuraFace repo (scrfd, 2d106det, 1k3d68, genderage) are InsightFace models under a non-commercial license and are not part of this repository.
Modifications
- Converted ONNX (NCHW) → TFLite (NHWC) with onnx2tf, batch fixed to 1.
- Dynamic-range quantization: weights int8, activations float32.
- Cosine similarity to the original ONNX embedding on aligned real faces: mean 0.998.
Usage
- Input:
float32[1, 112, 112, 3], RGB, ArcFace 5-point aligned crop,(pixel - 127.5) / 127.5. - Output:
float32[1, 512]embedding. Compare L2-normalized embeddings with cosine similarity.
Attribution
AuraFace-v1 by fal.ai, licensed under the Apache License 2.0. See LICENSE.
- Downloads last month
- 37
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for eungsang/auraface-v1-tflite
Base model
fal/AuraFace-v1