AI powered face verification for apps that need fast identity checks.
TinyFaceMatch compares two face images and returns a clean match decision with similarity scores. Try the hosted model for free, then use the open-source API in your own product.
Built for practical face API workflows
Measured on validation pairs with a threshold tuned for low false accepts.
- ROC AUC
- 0.9983
- Accuracy
- 0.9972
- Balanced accuracy
- 0.9902
- True accept rate
- 0.9830
- False accept rate
- 0.0025
- False reject rate
- 0.0170
- Threshold
- 0.2856
- Model size
- 13.238 MB
Simple response, useful signal
Use TinyFaceMatch for onboarding checks, account recovery, duplicate profile detection, or internal verification flows where a compact open-source model is easier to audit and adapt.
- Free hosted model test
- Two-image verification flow
- Similarity and distance scores
- Open-source Python package
Smaller than popular face recognition models, with standout accuracy
TinyFaceMatch is designed for teams that want strong verification quality without shipping a huge model. Always review upstream model and dataset licenses before using any third-party weights commercially.
| Model | License situation | Public benchmark | Size | TinyFaceMatch advantage |
|---|---|---|---|---|
| TinyFaceMatch | Open-source project; release terms controlled here when training data is clean | 99.72% accuracy, AUC 0.9983 | 13.238 MB | Excellent accuracy-to-size balance |
| OpenCV SFace | Apache 2.0 for the model directory | LFW 99.60% | 36.9 MB | +0.12 percentage points and about 64% smaller |
| dlib face recognition ResNet | Face-recognition model files are effectively public-domain style; avoid the 68-point landmark model for commercial use due to dataset restrictions | LFW around 99.38% | 21.4 MB | +0.34 percentage points and about 38% smaller |
| FaceNet PyTorch VGGFace2 | MIT code, but pretrained weights depend on VGGFace2 / CASIA data, so commercial use needs review | LFW 99.65% | 107 MB | Slightly higher accuracy and about 88% smaller |