The Numbers
| Ollie's celebrity index | |
|---|---|
| Celebrities | Over 5,000 |
| Photos | Over 40,000 |
| Photos per person | 2 to 12 |
| Split | About three in five men, two in five women |
| Photo source | Wikimedia Commons only |
| Licences | Public domain, CC0, CC BY, CC BY-SA and similar |
This index is separate from the data used to train the model. The model learned what faces look like from a public research dataset. The index is the set of people it can match you with.
Choosing the Celebrities
The list starts from Wikidata, the structured database behind Wikipedia. It takes every living person covered by at least ten language editions of Wikipedia, a simple way of saying "known in many countries".
Each person is ranked by how many people read their English Wikipedia article over the previous six months, with a boost for being covered in many languages. That keeps the list current: someone who was huge in 2005 but rarely looked up now ranks lower than a rising star.
English Wikipedia under-counts stars who are famous elsewhere, so a second list ranks people from East and South Asia, Southeast Asia, Latin America, Africa and the Middle East by page views on their own languages' Wikipedias. About half of that list is women, which helps balance the overall index.
Some people are left out on purpose:
- Minors. A known birth date is required and everyone must be 18 or over.
- The deceased. Living people only.
- People known mainly for adult entertainment or serious crimes.
Read more: How Ollie's Celebrity Lookalike AI Works: A Technical Overview
Only Freely Licensed Photos
Every photo comes from Wikimedia Commons and must carry a licence that allows reuse with credit. Photos under non-commercial or no-derivatives licences are never used, and files flagged for deletion are skipped.
For each photo, Ollie stores the photographer's name, the licence and a link to the original, and shows that credit under the photo in your results. That's more work than scraping a search engine, but it means every celebrity photo on the site is used with permission granted by its licence.
Checking Every Photo
Candidate photos come from four places: the person's main Wikidata photo, photos tagged as depicting them, their Commons category and name searches. Name searches in particular bring in a lot of wrong photos, so every candidate goes through automatic checks before it's kept:
- Exactly one clear, large face, so no group shots.
- Facing roughly toward the camera.
- Sharp enough to read.
- In colour, so no black-and-white or sepia prints.
- No sunglasses.
- An adult face.
- A real photograph, not a statue, wax figure, drawing, poster or meme. This is checked using both the file's title and an image model.
Read more: How Does AI Facial Recognition Actually Work? A Plain-English Guide
Get new guides by email
Checking It's the Right Person
Passing the quality checks isn't enough: the photo also has to show the right person. A face recognition model compares each candidate with the person's own Wikidata photo and with their other candidate photos, and keeps only clear matches.
If the reference photos disagree with each other about who the person is, which happens when a Wikidata photo is wrong or shows someone else, the person is skipped rather than guessed. Near-duplicates and too many photos from the same event are dropped, so one red carpet appearance doesn't dominate someone's set.
An audit step then produces contact sheets, grids of every kept photo per person, for spot checks by eye.
Building the Search Index
Each kept photo is run through Ollie's network once, producing 512 numbers. These are stored side by side, one per photo, rather than averaged into one set of numbers per person.
That choice matters. Averaging would blur a person's look across years, hairstyles and angles. Keeping each photo separate lets a search score each celebrity by their single closest photo, and show you that exact photo. If you look like an actor did a decade ago, you'll see that older photo.
Searching is then a single matrix calculation: your 512 numbers against all 40,000+ stored sets at once, which takes milliseconds. For the rest of the pipeline, see how Ollie works.
Takedowns and Corrections
If a celebrity or photographer wants a photo or person removed, the Ollie team deletes them from the source data, rebuilds the index and keeps a record so they don't come back on the next rebuild. You can get in touch through the contact page.