Duplicate Image Finder

Phone backups, downloads folders and shared drives fill up with the same photo saved several times: resized for messaging, recompressed by an app, or copied twice. File names and sizes rarely give them away. This finder looks at the pictures themselves. It gives every image a 64-bit perceptual fingerprint, groups the ones whose fingerprints are close, and flags byte-identical files with SHA-256. It then suggests which copy to keep. Everything runs in your browser, so private photos are never uploaded.

How to use this tool

1. Select all the images you want to check (you can pick hundreds at once).
2. Keep the threshold at 5 to catch resized and recompressed copies; use 0 for exact matches only or 10+ for similar shots.
3. Optionally include rotated or mirrored copies, and choose which file to keep in each group.
4. Review the groups overview and the table: green is the suggested keep, red the duplicates.
5. Delete the listed files yourself, or download a ZIP of the files to keep.

Formula

distance = number of the 64 neighbouring-cell comparisons (difference hash) that clearly disagree between two pictures (0–64)

Each image is shrunk to 64 × 64 grey pixels and averaged into a 9 × 8 grid; each of the 64 comparisons records whether a cell is darker or brighter than its right neighbour (dHash). A comparison counts as a difference only when it is clear: the two pictures step in opposite directions, or one has a step of at least 3 grey levels where the other is flat. Plain dHash bits on flat areas such as sky, walls or screenshot backgrounds depend on fractions of a grey level and flip when a copy is resized or recompressed, so they are not counted. Nearly flat pictures must also have the same average colour and matching thumbnails. The optional average hash compares each 8 × 8 cell with the mean in the same way. Files with the same SHA-256 digest are identical byte for byte. With "rotated or mirrored" on, the eight orientations of each image are compared and the closest is used.

Example

A 4000 × 3000 photo and the 1280 × 960 copy a messaging app made of it usually score 0, so at threshold 5 they form one group; with "keep the largest" the 4000 × 3000 original is kept. A copy with 3% trimmed from each edge scores about 7–8, so light crops need a threshold of about 8–10; unrelated photos usually score 25 or more.