AI Image Similarity

What it compares

Load two images and the page reports how visually similar they are, as a percentage, with a ring, a progress bar and a plain-language label. Both images are turned into numerical fingerprints by a small neural network, and the two fingerprints are compared mathematically.

All of it happens in the browser. Neither picture is uploaded, and there is no search across a collection — only the pair you provide.

The model and the comparison

An embedder does not compare pixels. It runs each image through a network trained for image classification and keeps the internal feature vector — a compact description of shapes, textures and structures — before the final label is chosen. Two images with similar content end up with similar vectors even when their lighting, size or exact framing differ.

Quantisation makes the vectors smaller and the comparison faster, at the cost of a little precision. For a two-image check that trade is invisible.

Why the percentage starts at 50

This is the single most important thing to understand about the score, and the interface does not explain it.

Cosine similarity ranges from −1 to +1. The page maps it to 0–100 with a straight formula: (similarity + 1) / 2 × 100. That means a raw similarity of zero — two vectors with no particular relationship — appears as 50 percent, not 0 percent. The scale is shifted so that negative values can be displayed, not calibrated so that "half similar" means anything.

In practice:

• Around 50 percent means the embeddings are unrelated.
• Around 70 percent corresponds to a cosine of about 0.4, which the page labels "Very similar".
• Above 85 percent is a strong match, usually a near-duplicate or the same subject from a similar angle.
• Below 40 percent is genuinely dissimilar, and uncommon between two ordinary photos.

The labels are generous by design. "Somewhat similar" begins just above 40 percent, which is still a low bar. Use the number for ranking and comparison between pairs, not as an absolute judgement about a single pair.

What it is good at

• Finding near-duplicates: the same photo resized, recompressed, or saved in another format.
• Checking a compression result: comparing an original against a heavily compressed version before publishing.
• Comparing design variants: two versions of a poster or thumbnail, to see whether the change altered the overall look.
• Spotting the same subject in two shots: a product photographed from two angles will usually score well above the unrelated baseline.
• Sanity-checking a dataset: confirming that two folders really contain different material.

What it is not good at

It is not reverse image search. There is no index, no database and no web lookup; it compares exactly two pictures you supply.

It is also not a pixel diff. Two images with identical layout but different text can score high, because the network responds to overall structure rather than the words. Conversely, the same object photographed against a very different background can score lower than expected.

Rotations, crops and flips usually reduce the score, sometimes sharply. A 90-degree rotation can drop a pair that a human would call identical. Heavy filters, colour grading and extreme exposure changes also move the number, though usually less than geometric changes.

Finally, the model was trained on object categories, so it is better at "these are both dogs" than at "these are the same document layout". For document comparison, a perceptual hash or a text diff is the right tool.

Using the page

1. Load the first image in the left panel and the second in the right. Each accepts any image type; drag and drop is not supported.
2. The first comparison downloads the runtime and the model. The panel reports each stage.
3. As soon as both images have loaded, the comparison runs automatically — there is no Compare button.
4. Read the ring: green above 70, amber above 40, red below. The label below it uses the same thresholds.
5. Replace either image to compare again; the result updates on load.

The Load AI Model button preloads the embedder so the first pair is instant. Once loaded it shows "Model Ready" and disables itself. Swapping one image reuses the loaded model, so repeated comparisons are fast.

Reading a score responsibly

Treat the output as a relative signal. If you are comparing several candidates against one reference, the ranking is more trustworthy than the absolute numbers: the pair that scores highest is probably the closest match, even if all the scores sit in a narrow band.

Do not use a fixed threshold as a decision rule without testing it on your own material. The same 75 percent can mean "clearly the same product photo" in one collection and "both are outdoor scenes" in another. Collect a few known pairs, note their scores, and choose a cut-off from that evidence rather than from the label text.

And do not treat a low score as proof of difference. A crop can lower the number while the content is identical; a human look is still the final check.

Privacy

Both images are decoded and embedded locally; nothing is uploaded, and no comparison is stored. The only network requests are for the MediaPipe bundle, the WebAssembly runtime and the model weights on first use. That makes the tool usable for private photographs, provided you are entitled to compare them.

Limits

Two images at a time, one comparison per pair, no batch ranking, no folder scanning, no database, no threshold control, no export of scores or vectors and no image editing. Animated images are treated as a single frame. Very large images take longer to decode, and the page needs a modern browser with WebAssembly plus network access on first load.

Common mistakes

• Reading 50 percent as "half the same". It means unrelated; the scale is shifted, not calibrated.
• Expecting rotation or cropping to score well. Geometric changes cost more than colour or quality changes.
• Using it as reverse image search. It only compares two files you provide.
• Setting a universal threshold. Scores depend on the kind of images being compared.
• Comparing documents. Layout and text differences are not what this model measures.
• Assuming a high score proves origin. Similar-looking images can be independent.

Frequently Asked Questions