Image Uniquifier

What happens to your photo

The image is redrawn from scratch with small edits: a thin border grows around the frame, the frame itself shifts slightly inside its area, rotates by fractions of a degree and changes size, random noise runs across the pixels, and brightness, contrast, saturation and hue drift a little. Every copy gets its own random values within the strength you choose, so two copies of the same photo never match each other either.

Metadata does not carry over. The image is drawn on a canvas, so EXIF, GPS coordinates, the ICC profile and C2PA tags never reach the new file. You do not need to clean them separately.

Frame shift and the micro border

The shift moves the image inside the same area by a percent or two. The strip that opens up is filled by stretching the outermost row of pixels, and the frame size stays the same. For perceptual hashing this is the strongest trick after mirroring, because the content moves relative to the grid the hash is computed on.

The micro border is a narrow frame around the image, colored like its own edges: the outermost pixel row is stretched outward, so the border is invisible. Nothing is cropped out of the photo, which matters for product cards where the edges sometimes carry something you need.

Measured on a detailed photo at medium strength, ten copies per case: cropping shifts the hash by 4.2 of 64 bits, the micro border by 8.1, the shift by 11.2. Shift and border add up to 12.7, because they edit the frame in different ways and do not interfere with each other.

Cropping and the border do not combine. One shrinks the frame, the other grows it, and together they give 5.7 bits instead of eight. That is why cropping is off by default. If you can afford to lose the edges, turn it on and remove the border.

Color does not fight the hash

Shifts in brightness, contrast, saturation and hue do not move the perceptual hash at all, not by a single bit. The hash compares neighboring pixels by brightness, and this kind of edit changes brightness everywhere at once, so the order of neighbors stays the same. The same holds for gamma, levels and channel balance: tested on eight variants, zero everywhere.

These edits have their own job. With ten copies per case, geometry and noise barely touch the color histogram, while a tone correction moves it by 36%. Color comparison is a separate way to find duplicates, unrelated to the hash, so each copy shows both numbers: how many bits the hash drifted and how far the color moved as a percentage.

The two edits sit on two separate checkboxes for a reason. The tone correction changes brightness, contrast, gamma and an S-curve familiar from Curves in an editor. It produces almost the entire histogram shift and leaves hues alone. The color correction changes saturation and hue: it adds less than a percent but recolors the photo, and for a product card the exact color matters more. So the tone checkbox is on and the color one is off.

Uniqueness is shown as a number

Similar services promise uniqueness, but there is no way to verify their claim. Here each copy gets a dHash perceptual hash, and the page shows how many of its 64 bits differ from the original. This is the same hash that duplicate search runs on: it recognizes an image after recompression and light retouching.

How far the hash drifts depends on the photo itself. On a frame full of small detail a medium edit moves it by fifteen to twenty bits; on a flat background only a few. The perceptual hash is designed on purpose to tolerate small changes. Mirroring rewrites almost all of it, around 45 bits of 64, but a flipped photo is visible to the eye.

The file bytes are always different, even with a weak edit: each copy has its own MD5 sum and its own size. Matching by hash and matching by file are different things, and platforms check both.

What the tool does not do

It does not hide a photo from a person or from reverse image search. The picture stays recognizable, otherwise it would be useless. Do not count on it to get around a platform's rules either: moderators look at more than hashes.

Never use it on someone else's photos. The edits do not make a photo yours; it still has its author. The tool is for your own shots, when one listing goes to several platforms or one shoot feeds several product cards.

How to use it

• drag your photos in;
• choose how many copies to make of each and how strong the edits should be;
• untick the boxes for tricks you do not need;
• download the copies one by one or all at once as an archive. Each name gets a random tail like "divan-1-k3f7.jpg", so separate uploads do not match by file name either.