How Celebrity Look-Alike matching works

The candidate pool, the six compared features and the match score behind each Celebrity Look-Alike result

Updated
On this page

Overchat AI Celebrity Look-Alike analyzes users' facial features to identify their four closest celebrity look-alikes. It provides a similarity score (in percentages) for each match, along with a breakdown of the facial features they have in common.

Matching pipeline

Celebrity Look-Alike matching pipeline has three components:

  • AI model. A multimodal LLM (large language model) is used to analyse an input image, compare the face with celebrities based on the selected sex and era, and return the closest match with similarity score and a description of shared facial features.
  • Ranking. The matches are sorted by similarity score, from highest to lowest.
  • Reference portraits. We use publicly available celebrity photos from Wikipedia to show each match alongside the uploaded photo for easy comparison.
  1. 01PhotographOne frontal photo
  2. 02Sex and eraSet the pool of candidates
  3. 03AI modelCompares six features and the whole face
  4. 04Four matchesEach with a score from 60 to 98
  5. 05RankingHighest score first
  6. 06Reference portraitsPublic photos for comparison

The sections below follow the pipeline from left to right.

Input photograph

Celebrity Look-Alike uses one image as an input. This can be a selfie or a photograph, ideally taken in good light without perspective distortion.

Frontal view is preferred as it shows all six compared features at once and in their true proportions:

  • eyes
  • nose
  • lips
  • jawline
  • face shape
  • hair

All six features are taken into account individually and influence the overall score. However, the overall score isn't calculated as the sum of the feature similarities, because people recognise a face by the arrangement of its features, the spacing and proportions between them, as much as by the features themselves (Maurer et al., 2002). Two faces with a similar nose but different proportions don't look alike, so the AI model compares the face as a whole after comparing each feature.

Matching parameters

Before starting the scan, users choose two settings that determine which celebrities their photo will be compared against:

ParameterOptionsEffect on matching
SexMale, femaleLimits the results to celebrities of the selected sex, since facial features such as the jawline and brow tend to differ between male and female faces.
EraAny, modern, classicDetermines which celebrities are included. Modern focuses on celebrities active since 2010, classic on earlier stars, and any includes both.

These settings help narrow down the selection of celebrities and make the comparisons more relevant to the user's choices.

Matching procedure

The AI model handles the entire matching process. It takes the uploaded photo and the user's selected sex and era, then identifies the closest celebrity look-alikes. It also provides a similarity score and explains which facial features the matches have in common.

1. Selecting candidates

The comparison is limited to well-known people with an English Wikipedia page. This keeps the results focused on recognisable celebrities and provides a public portrait for each match. Familiar faces are recognised reliably across very different photographs, while unfamiliar ones are not (Jenkins et al., 2011), so a match with a face the user already knows is easy to check.

2. Comparing facial features

The model looks at six facial features when comparing the uploaded photo with each candidate:

FeatureWhat is compared
EyesShape, size and spacing
NoseWidth, length and bridge
LipsFullness and proportion
JawlineWidth and definition
Face shapeOverall shape and proportions
HairColour, texture and hairline

The model considers both individual features and how they work together. A similar nose or jawline alone does not necessarily make two people look alike. The overall arrangement and proportions of facial features also matter (Maurer et al., 2002).

3. Match score

Each match receives a score between 60 and 98. A higher score means a stronger resemblance.

  • The minimum score is 60, as the results are intended to include only celebrities with a noticeable resemblance.
  • Scores are spread across the range to show the difference in similarity between the four matches.
  • Before displaying the results, scores are rounded to whole numbers and kept within the 0–100 range. The matches are then sorted from highest to lowest score.

Reference portraits

Once the matches are selected, we load a portrait for each celebrity from their Wikipedia page. The portraits are only there to help users compare the faces side by side; they have no effect on the matching results or scores. If no portrait is available, we display the celebrity's initials instead.

Celebrity Look-Alike compares the user's face with those of well-known celebrities. It does not attempt to identify the person in the uploaded photo. Instead, it looks at facial similarities to find celebrities who share similar features.

References

  1. Maurer, D., Le Grand, R., & Mondloch, C. J. (2002). The many faces of configural processing. Trends in Cognitive Sciences, 6(6), 255–260
  2. Jenkins, R., White, D., Van Montfort, X., & Burton, A. M. (2011). Variability in photos of the same face. Cognition, 121(3), 313–323

All methodology topics