Find the faces. Choose your next step.

A group photo. A separate next step for each face.

Choose Find faces and upload a portrait or group photo. The result marks detected faces with numbered boxes and gives each detection its own person crop and detection-confidence value. Select a box to move to its card, or enlarge the photo and crops for a closer look.

Detection does not run a database search or establish who anyone is. You decide which person, if any, to explore next. Repeated photographs of the same person can produce several detections.

● ● ●CS Match / Face AnalysisDemo example

Two photos. One chosen reference.

Use Compare faces when you already have a reference portrait. Upload Photo 1 with one or more faces and Photo 2 as the reference. If Photo 2 contains several people, choose the reference person explicitly and press Compare faces again.

The report compares every detected face in Photo 1 with that selected reference. It shows a result for each face and a summary of strong, inconclusive, low and unavailable comparisons. More than one face can score highly; this is not a one-to-one identity assignment.

A face that cannot be aligned stays in the report with its comparison unavailable. An unclear reference needs a clearer replacement. You can still send a detected person crop to Face Search.

● ● ●CS Match / Face AnalysisDemo example

Read the score alongside the photos.

Switch between Cards and Compared in the comparison report. Cards keeps the individual face cards; Compared places a source face, percentage and selected reference inside each card, with three cards per row on wide screens. Both views use the same completed result. Find faces also provides a table of the first five detections beside the source photo, with More leading to the full list.

Detection confidence describes the detector’s confidence that it found a face. Vector similarity measures how close two numerical face representations are. The normalized 0–100% similarity score uses the same configured scale as Face Search.

A high percentage is not a probability that the photos show the same person. Different raw vector values can reach 100%. Pose, lighting, blur, face size and occlusion can change the result. Enlarge the images and review their context before drawing a conclusion.

Understand face similarity scores →Improve your source photo →

From one person crop to the public image index.

Every detected face has a Search this face button, including faces in a comparison report. It opens Face Search with the selected person’s context crop. Face Search detects the face again, builds its search representation and searches the indexed public images.

This handoff uses the person crop rather than the entire group photo. Face Search expects exactly one detected face in its input. If a close or overlapping group leaves another face in the crop, use a clearer image or a more isolated portrait.

Review ranked matches, the separate Similar faces gallery, domains and recorded source links in Face Search. Coverage is limited to the index; an absent result does not establish that a photograph is absent from the internet. Returning to Face Analysis in the same signed-in session keeps the current form and report.

Explore Face Search →Explore Similar faces →Check source domains and variants →

Start with a photo the detector can read.

Use a still JPEG, PNG or WebP photo up to 8 MiB and 20 megapixels. These limits apply separately to each photo in a comparison. A free account is required; no desktop application or reference photo is needed for Find faces.

Use visible, sufficiently large human faces. Low light, strong angles, motion blur and partial coverage can cause missed detections or unreliable comparisons. Finding every person in every image is not guaranteed.

Large images are reduced to at most 4096 pixels on the longest side for analysis. If more than 200 faces are detected, the tool asks you to split the image into smaller groups instead of silently omitting results.

Troubleshoot missed faces →How image quality affects accuracy →

The technology behind each comparison.

The current pipeline uses a SCRFD detector to find face boxes and landmarks, then a pretrained ArcFace recognition checkpoint to represent aligned faces. Compare aligns each face to a 112 × 112 RGB patch and calculates a normalized 512-value embedding. Find faces requests detections without running that embedding stage.

Each source embedding is compared with the selected reference by cosine similarity. The inspected display endpoints are 0.45 and 0.72; a clipped linear mapping turns the raw value into the displayed percentage. This scale is not an identity probability. The technical guides explain the exact formula, training versus inference, error modes and practical limits.

● ● ●CS Match / Face AnalysisTechnical schematic

Detection and alignment →Vectors and the pretrained model →The exact reference comparison →Errors and accuracy limits →

Your upload does not become a public index entry.

Face Analysis does not publish or add your uploaded photos, detected crops or comparison vectors to the public face index. Detection and comparison results remain in the current browser session; keep the page open while analysis runs and save anything you need before reloading or signing out.

Use photos you are entitled to process. Starting a database search is a separate action. In Face Search, a saved result becomes accessible by public link only when you choose to share it.

Privacy Policy →Acceptable use →Data deletion requests →