AI Headshot Identity Checker

The "does it still look like me?" problem

AI headshot tools have a known failure mode: the generated face is technically polished but the identity has drifted. Skin smoothing, jaw normalisation, eye-shape standardisation — small changes to many features add up to a photo that looks great but no longer looks like you. The first time someone meets you in person after seeing your AI LinkedIn photo, the disconnect is jarring.

The Identity Checker measures this drift quantitatively. Upload your real selfie + the AI render and you get a single number that tells you whether the AI preserved your identity well enough to use, or whether it slid too far.

What ArcFace actually measures

ArcFace (the algorithm we use) is the industry standard for face recognition since 2019. It converts each face into a 512-dimensional vector — an "embedding" — where each dimension captures a specific facial feature: eye spacing, nose bridge angle, jaw width, cheekbone position, the distance between mouth corners, and dozens more. Two photos of the same person produce nearby vectors; two photos of different people produce distant vectors.

The cosine similarity between two embeddings ranges from -1 to +1. We map that to 0-100% for readability:

• 83%+ score (0.66+ raw): Above the strict same-person threshold used by Apple Face ID, Microsoft Hello, and most border-control systems.
• 68-82% score (0.36-0.65 raw): Standard same-person threshold from the ArcFace paper.
• 50-67% score (0-0.35 raw): Ambiguous — some shared features but enough deviation that automated systems would flag it.
• Below 50% (negative raw): Different people.

How AI headshots typically score

On FreeHeadshot's own Studio output, the median AI headshot vs the source selfie scores 0.72-0.78 raw cosine (87-89% on our 0-100 scale). That puts it solidly in "likely same person" territory — the AI preserved the identity. Outliers below 0.65 are usually cases where the source selfie had unusual lighting (very warm indoor or very cool blue tone) that the AI normalised away.

Competitor AI headshot tools that use 10-30 reference photos and per-user model training (Aragon, HeadshotPro) typically score 0.78-0.85 — slightly higher because they have more data to anchor on. Our single-selfie pipeline trades a small amount of identity precision for the speed and convenience advantage.

The privacy posture

Both uploaded photos hit our private VPS worker over HTTPS. The worker decodes each photo, detects the largest face via InsightFace, generates the 512-dim ArcFace embedding, computes the dot product, returns the score, and immediately destroys the in-memory photo buffers. Nothing is written to disk, nothing is logged beyond the request line + timing, nothing is used to train AI models.

We chose InsightFace specifically because it's an open-source, locally-running model — no third-party API is involved. Apple Face ID and Microsoft Hello use closed, cloud-API-backed pipelines; this tool deliberately doesn't.

What this tool isn't

• Not a legal identity verification. Government identity systems combine face recognition with liveness checks (blink, head turn, smile-on-cue) and physical document verification. We do none of that.
• Not a deepfake detector. If both photos are of the same person but one is heavily edited, this tool will say "same person" — which is correct. To detect "is this an AI generation at all", that's a different research problem.
• Not perfect on edge cases. Identical twins, heavy makeup transformations, age progressions beyond ~10 years, and severe pose differences all reduce accuracy.

Related tools

Use the Quality Checker if you want to score a single photo on lighting / sharpness / framing rather than compare two. If you've generated an AI headshot and the identity score came back low, regenerate with a clearer source selfie via the Studio.