Are these two images the same?
The point of this site: computer similarity and legal similarity are two very different things.
The tests here measure how alike two images are — their bytes, pixels and learned features. Copyright law asks different questions: was protected expression copied, was it a substantial part, and was the use fair? A pair the scanner flags as a possible copy can be lawful, and a pair it calls merely similar can infringe. The real cases below show both.
Drop two images and run 101 similarity tests side by side — from byte-level file hashes to Meta’s SSCD copy detector. Everything runs in your browser; your images are never uploaded.
Or compare the works from a real copyright case: all cases ↓
Running tests
Copyright case examples
Real cases put the site’s point to the test. Image A is the plaintiff’s work and image B the defendant’s. Scan a pair, or run them all, and set what the computer measures beside what the court held — they often disagree. Koons had Rogers’s photograph re-made as a painted wooden sculpture, so the copy detectors see only a similar subject, yet the court found infringement. Prince built his paintings from Cariou’s actual photographs, which the detectors flag, yet the court held most of the series fair use.
How it works
Computer similarity is not legal similarity
This is the point of the site. The numbers measure resemblance between files, pixels or learned features. The law asks whether protectable expression was copied (not ideas, facts or stock elements), whether the copying was substantial, and whether the use was fair. No metric answers those questions: treat the scores as evidence to interpret, not conclusions.
Many definitions of “the same”
A file hash asks whether two files are bit-for-bit identical. A perceptual hash asks whether they would look identical at thumbnail size. SSIM asks whether the pixels line up. SSCD asks whether one is an edited copy of the other. CLIP asks whether they depict the same kind of thing. Each test answers a different question — the disagreements are the interesting part.
Reading the score
Like a malware scanner, the headline counts how many tests flag the pair: Identical or Match count as flags, Partial is shown separately. Thresholds come from the original papers where they exist (SSCD, PDQ, Somepalli et al.) and are labelled as heuristics elsewhere.
Private by design
Images are decoded and analysed on your device with WebAssembly (ONNX Runtime, OpenCV). The neural models are downloaded once (≈ 410 MB in total, configurable in Settings) and cached by your browser.