How the cat breed scanner works
The scanner is a vision model we trained ourselves to tell 31 cat breeds and the ordinary house cat apart. On 6,409 test photos it had never seen, its first answer was right 81% of the time and the right breed was in its top three 98.5% of the time. Here is how it works, what it was trained on and where it still gets things wrong.
What happens to your photo
Your browser shrinks the photo to at most 512 pixels on the long side and sends it to our server over an encrypted connection. The server first checks that the photo shows a cat. If it does, the model scores the cat against 32 classes, 31 breeds plus the house cat, and sends back the top three. The photo exists only in memory while this runs, usually a second or two, and is then discarded. Nothing is written to disk or kept in a database.
The model
We started from DINOv2, an open vision model from Meta that has learned general visual features from a very large set of images. We fine-tuned its last layers and a new classification head on our cat photos, so it learned which details separate a Ragdoll from a Birman or a British Shorthair from a Russian Blue. The model has about 86 million parameters and runs on an ordinary server processor.
The training photos
We collected about 213,000 candidate photos from public cat image collections, including the Oxford-IIIT Pet dataset and adoption photos labeled by shelters and owners. Then we cleaned them:
- We removed photos that were too small, did not show a cat or showed several cats.
- We removed near duplicates, keeping the sharpest copy, which left about 124,500 photos.
- Owner and shelter labels are often wrong, because many house cats are listed as a breed they only resemble. A first model flagged 12,282 photos whose label looked doubtful. Each was re-checked twice, independently, by AI vision reviewers who did not see the original label, and a third review settled disagreements. About 9,000 of them turned out to be house cats.
- We kept the 31 breeds with at least 200 clean photos. Rarer breeds, such as the Burmese, Devon Rex and Chartreux, are left out until we have enough.
The final set has 50,850 training photos, 7,531 for tuning and 6,409 for the test. Copies and near copies of the same cat never appear on both sides of the split, so the test measures how the model does on cats it has truly never seen.
How accurate is it?
- Our test set (6,409 photos): 81.0% right on the first answer, 98.5% in the top three.
- Breed photos only: 85% right on the first answer.
- House cat or breed: it decides correctly in 85.5% of cases.
- Oxford-IIIT Pet test set (1,125 cat photos, 12 breeds): 91.8% right on the first answer across all 32 classes, 95.8% when limited to those 12 breeds.
The table shows the result for each class on our test set. Small breeds have only 30 to 60 test photos, so their numbers can move by about 10 points either way.
| Breed | First answer right | In top three |
|---|---|---|
| Cornish Rex | 100% | 100% |
| Lykoi | 100% | 100% |
| Abyssinian | 99% | 100% |
| Sphynx | 99% | 100% |
| Russian Blue | 97% | 100% |
| Savannah | 97% | 100% |
| Persian | 96% | 100% |
| Scottish Fold | 95% | 98% |
| American Curl | 94% | 98% |
| British Shorthair | 94% | 99% |
| Bengal | 93% | 98% |
| Khao Manee | 93% | 100% |
| Egyptian Mau | 92% | 98% |
| Exotic Shorthair | 92% | 98% |
| Neva Masquerade | 92% | 100% |
| Siamese | 90% | 99% |
| Siberian | 90% | 97% |
| Munchkin | 89% | 99% |
| Turkish Angora | 89% | 98% |
| Ragdoll | 86% | 98% |
| Snowshoe | 82% | 98% |
| Maine Coon | 81% | 99% |
| Bombay | 76% | 100% |
| Birman | 75% | 98% |
| House cat | 74% | 98% |
| Norwegian Forest Cat | 74% | 98% |
| Oriental Shorthair | 73% | 90% |
| Manx | 67% | 90% |
| Himalayan | 60% | 97% |
| Turkish Van | 53% | 97% |
| American Bobtail | 50% | 90% |
| American Shorthair | 44% | 98% |
Where it still makes mistakes
Almost all its mistakes sit on one boundary: house cat or look-alike breed. A black house cat can be read as a Bombay, a gray one as a Russian Blue and a silver tabby as an American Shorthair, and the reverse happens too. That is why the American Shorthair, American Bobtail, Turkish Van and Manx have the lowest first-answer scores, even though they are nearly always in the top three. When the house cat answer scores close to the top breed, the result says so.
Other known limits:
- Similar breeds. Himalayans get confused with Ragdolls and Persians, Maine Coons with Norwegian Forest Cats and Siberians, and Birmans with Ragdolls.
- Kittens. Many breed traits only appear with age, so photos of young kittens are harder.
- Missing breeds. A breed that is not in the model, such as a Burmese, will be matched to the closest-looking breed it knows.
- Mixed cats. A cat with two breeds in its family may show both in the top three, but the scanner cannot measure ancestry.
Looks are not ancestry
Every number on this page measures how well the model matches a cat's looks to the label a person gave it. It does not measure genetics. If you need to know your cat's real breed, use pedigree papers or a cat DNA test.
Curious which breed your own cat looks like? Upload one photo to the free cat breed identifier and see the three closest matches in seconds.
Identify my cat