ComparePicture

How similar are these two images?

No single number answers that. This page reports several that answer different parts of it, and combines them into a near-duplicate judgement that lists its evidence.

Private · Browser-based · No uploads

Two pictures in, every difference out.

  1. 1. Drop, paste or choose the earlier picture as A.
  2. 2. Add the later one as B — or drop both at once.
  3. 3. The comparison starts by itself; switch views any time.

PNG, JPEG, WebP, GIF (first frame), AVIF, BMP and HEIC where this browser can decode it. Different sizes are fine — they are aligned for you.

Both files are read by this page on your device. There is no upload step, because there is no server to upload to.

Or try an example pair

AFirstChoose image Adrag & drop, or paste with Ctrl/⌘ VPress Enter to choose image A, or drop a file here.
BSecondChoose image Bdrag & drop, or paste with Ctrl/⌘ VPress Enter to choose image B, or drop a file here.

Five measures, five questions

Each measure is computed in your browser on the two pictures you add.

  • Structural similarity (SSIM) — do the shapes, edges and textures match where they are? Sensitive to blur and to local edits; tolerant of uniform brightness changes.
  • Perceptual hashes — do the two pictures reduce to the same 64-bit fingerprint? pHash survives resizing and recompression; dHash follows gradients; aHash follows overall light and dark.
  • Histogram similarity — do they use the same colours in the same proportions, regardless of where?
  • Edge similarity — do the outlines match even if the colours do not?
  • Visual content similarity — does a compact image model (MobileNetV2) describe them the same way? About 17 MB with its runtime, downloaded once on request and run on your device.

Near-duplicate, edited copy, or a different picture

The Smart view combines the measures into one of four readings — the same picture pixel for pixel, a likely near-duplicate, likely the same picture with edits, or likely different pictures — and always lists the evidence behind it, including dimension differences and signs of recompression.

It is careful about what that means. Similarity is not provenance: it cannot tell you which file came first, who made either, or whether one was “stolen” from the other. Those are questions about people, not pixels.

Questions people ask

›What is the difference between pixel similarity and perceptual similarity?

Pixel similarity asks whether the same positions hold the same values. Perceptual similarity asks whether a person would see the same picture. A JPEG re-saved at a lower quality changes most pixels slightly but is perceptually almost identical; a one-pixel shift changes every pixel and is perceptually the same too.

›Does ComparePicture use AI?

Only where it says so: the Visual content similarity measure uses a compact image model that runs on your device after you choose to download it. Everything else — the diff, alignment, SSIM, hashes and quality measures — is ordinary computation.

›Does the AI run locally?

Yes. The model file is fetched from this site once and kept in your browser's cache; the pictures are given to it inside the page and are never sent anywhere.