Accuracy & Model

Built on the
best face model

Dollganger uses ArcFace — a state-of-the-art facial recognition model that consistently ranks at the top of global benchmarks.

By the numbers
512
Embedding dimensions
99.8%
LFW benchmark accuracy
< 2s
End-to-end search time
cosine
Similarity metric
The Model

ArcFace

ArcFace (Additive Angular Margin Loss) was developed by researchers at Imperial College London. It achieves state-of-the-art performance on face verification benchmarks by maximising the angular margin between face classes in embedding space.

Each face is encoded as a 512-dimensional unit vector. The similarity between two faces is measured as the cosine of the angle between their vectors — the closer to 1.0, the more similar.

Why ArcFace?

  • Top-ranked on LFW, CFP-FP, and AgeDB benchmarks
  • Robust to lighting, angle, and age variation
  • 512-dim embeddings capture fine facial detail
  • Open research — independently audited
  • No GPU required for inference
Understanding Confidence Scores

Confidence scores represent cosine similarity, converted to a 0–100% scale. Here's what typical scores mean in practice:

Note: uploading the same photo twice typically yields 95–98%, not 100%, due to minor differences in image compression and resizing during processing.

Processing Pipeline
1

Face detection

RetinaFace locates and crops the face region from your photo with sub-pixel precision, normalising for rotation and scale.

2

Embedding generation

ArcFace converts the cropped face into a 512-dimensional vector, then normalises it to unit length for consistent comparison.

3

Vector search

pgvector computes cosine similarity between your embedding and every person in our database using an IVFFlat index for speed.

4

Multi-embedding averaging

Each person in our database has multiple embeddings from different photos. The best match across all embeddings is returned, improving robustness.

5

Cleanup

Your photo and computed embedding are immediately discarded from memory. Nothing is written to disk.

Known Limitations

Accuracy decreases with side profiles, heavy makeup, masks, or extreme lighting conditions. Front-facing, well-lit photos perform best.

The database covers celebrities, historical figures, and fictional characters. If your lookalike isn't in the database, results will be the closest available match.

Confidence scores reflect geometric facial similarity only — not ethnicity, identity, or biological relationship. A 90% score between two people does not mean they are related.

Group photos or images with multiple faces may produce unpredictable results. For best accuracy, upload a single-person portrait.

See it in action

Upload a portrait and experience the accuracy yourself.

Try it free →