Loading page
About this tool
Scan, group, and remove visually similar or exact duplicate photos locally using perceptual hashing. 100% private and browser-based.
Duplicate Photo Finder compares pictures with a 64-bit difference hash (dHash). Each image is drawn onto a 9x8 canvas, converted to grayscale with the standard luma weights (0.299R + 0.587G + 0.114B), and every pixel is compared with its right-hand neighbour to produce 64 bits. Any two photos are scored by Hamming distance — similarity = (64 − differing bits) ÷ 64 × 100 — and everything at or above your threshold (60-100%, default 90%) is collected into a duplicate group. Decoding and hashing both run on a canvas element in your own browser, so the photos you add are never uploaded and no account is needed.
Open Duplicate Photo Finder on AltFTool — it loads instantly in your browser.
Click Add Images and select the photos you want to check — you can pick many files at once, and the progress bar shows a hash being computed for each one.
Set the similarity threshold (90% by default) and read the duplicate groups, each labelled with its average match percentage.
Use Visual Compare to confirm a pair, apply Smart Selection to mark the lower-resolution or smaller copies, then delete those files yourself in your own photo folder.
The 60-100% slider decides how close two hashes must be. At the default 90%, up to six of the 64 bits may differ — enough to catch a re-saved copy without pairing unrelated photos.
Every image is reduced to the same 9x8 grid before hashing, so a 4000 px original and an 800 px thumbnail of the same frame produce the same fingerprint.
One click marks everything except the highest-resolution copy in each group; another marks everything except the largest file, so you never lose the best version.
Files you add are read through local object URLs and hashed on a canvas in JavaScript. There is no upload step, no sign-up and no stored history.
By fingerprinting the picture rather than the file. This tool builds a 64-bit difference hash: the image is scaled to a 9x8 grid, converted to grayscale, and each pixel is compared with the one to its right — brighter sets a 1, darker a 0. Two photos whose hashes differ by only a few bits look identical to the eye even when the files differ byte for byte.
No. "Remove Marked" only clears images from the on-screen list — a web page has no permission to delete files on your disk. Use it to decide which copies are redundant, then delete those files yourself in Finder, File Explorer or your photo app.
No. Each file you add becomes a local object URL, is drawn onto a 9x8 canvas and hashed in JavaScript on your own device — no image data is sent anywhere and no login is required. The only exception is the four sample photos that load as a demo, which are fetched from Unsplash; press Reset Demo to clear them.
Start at the default 90%. Similarity is (64 − differing bits) ÷ 64 × 100, so 90% allows up to six of the 64 hash bits to differ — enough to catch a resized or re-compressed copy but not two different scenes. Raise it toward 100% for near-exact matches only, or drop toward 60% to catch heavier edits at the cost of some false pairs.
Usually not. A difference hash reads the whole frame, so cropping, rotating or flipping shifts every pixel comparison and produces a very different hash. It is built for resized, re-saved, re-compressed or lightly retouched versions of the same frame. Lowering the threshold helps with small crops, but not with rotation.
Anything your browser can decode in an image tag — JPEG, PNG, WebP, GIF and AVIF all work, and the file picker accepts image/* with multiple files at once. iPhone HEIC files only decode in Safari, so convert them to JPEG first if you are on Chrome, Edge or Firefox.
The Space Analytics tab estimates it: inside each group the largest file is kept and the rest are counted as wasted space, charted in KB against the total size of everything you loaded, alongside a breakdown of file formats. It covers only the files in the current session — nothing is stored once you close the tab.
There is no fixed limit, but the work grows quickly: each image is decoded and hashed one at a time, then compared pairwise, so n photos mean up to n(n−1)/2 comparisons. A few hundred images scan comfortably; several thousand will make both the scan and the grouping noticeably slow.