How do I find every photo of one person in a large photo library?
Search by face. Detect every face in every photo, turn each face into a face embedding (a list of numbers that is close for two photos of the same person and far apart for different people), and compare those embeddings with a few reference photos of the person you want. Every face within a set distance of the references is a match. For a personal library, Apple Photos, Google Photos and the open-source Immich already do this. For an archive of hundreds of thousands of photos, you run the same three steps yourself or with a service, then check the borderline matches by hand.
How do the photo apps find a person?
These work well on one person's camera roll. They stop being enough when the photos sit in a shared drive, an agency archive or a cloud bucket, when several teams need the same search, or when you need the results in your own software.
How does face search work across millions of photos?
| Step | What happens | Typical tools |
|---|---|---|
| 1. Detect | Find each face and its box in the photo, with landmarks for the eyes, nose and mouth | RetinaFace, SCRFD |
| 2. Align | Rotate and crop each face to a standard pose so the model sees faces the same way | Landmark-based alignment |
| 3. Embed | Turn each aligned face into a face embedding, often 512 numbers | ArcFace-family models |
| 4. Index | Store the embeddings in a vector index next to the photo's ID and the face's box | Any vector database |
| 5. Search | Embed 3 to 5 reference photos of the person and return every face within a distance threshold | Cosine similarity |
| 6. Group (optional) | Cluster all faces so each unknown person becomes a group you can name later | DBSCAN, HDBSCAN |
Search and grouping answer different questions. Search starts from someone you can show a photo of. Grouping starts from nothing and sorts the whole archive into people, which is what the phone apps do before you ever type a name. A face embedding differs from face recognition the way a fingerprint scan differs from a match against a database: the embedding is the measurement, and recognition is what you do with it.
Why are some photos of the person missing from the results?
Benchmarks of frontal, well-lit faces overstate how a model does on a real archive, so measure on your own photos before you trust a number.
What rules apply to searching faces?
A face embedding used to identify someone is biometric data. Under the GDPR it falls under the special categories in Article 9, which needs an explicit legal basis such as consent, and Illinois' Biometric Information Privacy Act requires written consent before collecting it. Searching your own family photos is a different situation from searching photos of employees, customers or the public, so check with counsel before you index other people's faces at scale. Some widely used pretrained face models are also licensed for non-commercial research only, so check the licence of the weights you deploy.
How do I do this with Mixpeek?
Point Mixpeek at the bucket where the photos already live. The face identity extractor detects each face with SCRFD, aligns it and embeds it with ArcFace into a 512-number vector, for images, video frames and PDF pages. A retriever takes your reference photos and returns every matching face with its photo and box, and clustering groups the unknown faces across the archive. The photos stay in your storage. On the pricing page, face detection and identity is $1.00 per thousand images, and clustering is included.
Related: face recognition and identity clustering in video, the best face recognition APIs, how to find duplicate photos, including edited copies and the InsightFace buffalo_l model.
Frequently Asked Questions
Can I search my photos with one photo of a person's face?
Yes. Upload the photo, the face is turned into an embedding, and every face in the library close to it comes back. Results are better with three to five reference photos taken at different angles and ages.
Why does my photo app not show some photos of a person?
The app did not detect the face (turned away, small, covered) or the face fell outside its match threshold. Most apps let you confirm suggested matches or add a photo to a person by hand, which also helps later matches.
How do I find photos of a person in thousands of videos?
Sample frames from each video, run the same detect, align and embed steps on each frame, and store the timestamp with each face. A search then returns the video and the moment the person appears.
Is it legal to search photos for a specific person's face?
It depends on whose faces they are and where you operate. Face embeddings used for identification are biometric data under the GDPR and under Illinois' biometric privacy law, which require a legal basis or written consent. Get advice before indexing faces of people who have not agreed to it.