Image Comparison — what the percent score means and how to use it

Drop two photos into the box at the top and the tool returns a single percent — how much of what is in one photo is also in the other. The page below is for people who actually need to act on that number: photographers culling near-duplicates, QA testers checking a screenshot against last week's, fraud teams hunting recycled damage photos. We cover what the number means at each level, where it lies to you, and the workflows we use ourselves.

What the percent really means

The tool runs both photos through MobileNet, a small vision model that turns each image into 1,024 numbers standing for what is in it — "dog", "wooden floor", "outdoor light". The two lists of numbers are compared with a measure called cosine similarity. The result is rescaled to a 0–100 percent. It is a measure of "how much the same stuff is in both photos", not "how identical the pixels are".

One worked example: two identical PNG screenshots of the same web page, taken on the same computer 10 seconds apart, will sometimes score around 96–98% rather than a clean 100%. That gap is anti-aliasing — the way the operating system renders text changes by a fraction of a pixel between captures. So even "the exact same thing" almost never reads 100% unless the file is byte-identical.

Three workflows worth knowing

1. Culling burst shots in a photo library

You have ten near-identical photos from a single burst, and you want to keep one. Pair them up and compare. Anything 95% or higher is safe to delete the second copy. 85–95% is two slightly different moments — keep both if you want the option to pick the better one later. Lightroom's "Stacks" and Photos.app's "Hide Duplicates" use a similar score under the hood.

2. Spotting a Photoshopped repost

Someone reposts your photo with a different filter applied, or with a logo cropped out, or with the background swapped. A side-by-side score in the 70–90% range is the giveaway — the subject is the same but something has changed. If you suspect the change is local (a face has been swapped, an object has been removed), zoom into the suspicious region and compare again with just the cropped piece — the score will fall further than you expect if the local edit is real.

3. Visual regression testing during a website redesign

Designers and QA testers compare a screenshot of today's build with last week's to catch accidental layout drift. The comparison here is rougher than a tool like Percy or Chromatic — it gives you a single number, not a pixel-diff overlay. But it is good as a smoke check before you do the full diff: anything below 90% deserves a look.

What the percent will not catch

• ⚠Small but important changes. A 10-pixel watermark added to the corner of a 4000×3000 photo will move the score by less than 1%. If you are looking for small added elements, do a side-by-side visual scan instead.
• ⚠Colour-only edits. A photo and the same photo with a sepia filter often score 85%+. The model is mostly looking at what is in the photo, not its colour palette.
• ⚠Two photos of the same kind of thing. Two different golden retrievers in a park will score around 75–88% because the model sees "golden retriever in grass" in both. The score is similarity of content, not identity.

Privacy note — where the comparison runs

Both photos are first uploaded to our hosting (Vercel Blob) so the browser can fetch them as buffers. The MobileNet model and the actual similarity comparison then run inside your browser tab — no comparison is done on our server. The first time you use the tool, the model file (about 17MB) downloads and is cached, so later comparisons are instant. You can confirm this in your browser's developer tools: after the model loads, the comparison itself produces no further network calls. Use the Purge button to delete the two uploads from our storage when you are done.

Frequently asked questions