How grouping works

Finding duplicate files is a question with a right answer: either the bytes match or they do not. Finding photos that look alike is a judgement, and this page is about how DuoBolt makes it — and where it refuses to.

The screen is called Similars, and its subtitle says what it is for:

Match images by visual content. Best for bursts and rescaled copies.

What it actually compares

Every image is decoded, reduced to a 32×32 grayscale thumbnail, and run through a discrete cosine transform. The result is a 64-bit perceptual hash — a fingerprint of the picture’s structure, not of its bytes.

Two fingerprints are compared by counting the bits that differ. That count is the distance: 0 means the two images produce an identical fingerprint, and larger numbers mean less alike.

File names, folders, and dates play no part in producing it.

Why a thumbnail

Reducing to 32×32 is what makes a re-saved JPEG, a resized export and the original land on the same fingerprint. It is also why two genuinely different photos of the same flat wall can look similar to it — see the gates below.

The threshold

Scan parameters → Similarity threshold sets how many bits may differ before two images stop being candidates. It is measured in bits, from 0 to 20, and defaults to 14.

Maximum Hamming distance between perceptual hashes. Lower is stricter; higher catches more matches at the risk of false positives.

How a group is built

Not by chaining.

DuoBolt grows a group only while every pair inside it stays within the threshold — not merely each new member against one existing one. Without that rule, a chain of slightly-similar photos would slide into one enormous group where the first and last images have nothing to do with each other.

The consequence is worth stating plainly: any two photos in a group are within the threshold of each other, not just of some representative.

Max group size caps how large a group may get. Larger mutually-similar sets are split into buckets, tightest matches first. It defaults to Unlimited.

The gates

A distance below the threshold makes two photos candidates. Several checks can still reject them:

• Low-detail images — a photo with very little contrast (a wall, a sky, a scanned blank page) produces an unreliable fingerprint, so DuoBolt requires a much tighter match before pairing them.
• Distance with no corroboration — past a certain distance, DuoBolt wants a second reason to believe: capture times close together, or agreement from a second, differently-computed hash.
• Capture time — photos taken more than a day apart are held to a stricter standard.
• A learned check — when the AI runtime is available, an embedding model provides a second opinion on whether two images really depict the same thing.

This is a filter, not a promise of completeness

These gates exist to keep false pairs out, and they cost recall. A low-contrast pair with no EXIF taken years apart may not be grouped even though you would call them the same photo. DuoBolt finds similar images; it does not claim to find all of them.

Variants and Look-alikes

Every group is labeled with one of two kinds, and the distinction is about what you would do with it.

Variants

One shot, stored more than once. A RAW and the JPEG exported from it, an iCloud HEIC/JPG pair, a downscaled export beside its original.

The useful action is usually: keep the best copy, drop the rest.

Look-alikes

Different frames that resemble each other. Burst frames, retakes of the same setup, a photo re-saved after an edit.

The useful action is usually: choose which frame you like, which is a judgement no tool should make for you.

A group is Variants when all its files are image formats, they agree on capture time or share a filename stem in one folder or produce identical fingerprints, and they differ in format or dimensions — that last part is what makes it “the same shot stored differently” rather than “the same shot twice”. Everything else is Look-alikes.

Disagreeing capture times are authoritative: if the EXIF says these were taken at different moments, the group is Look-alikes whatever else matches.

Tightness

Every group carries how tightly it matched, based on the largest distance between any two of its members. There are five bands:

Band	What it means
Identical	Same fingerprint, and byte-identical files
Virtually identical	Same fingerprint, different bytes — a re-encode
Very similar	1 to 4 bits apart
Similar	5 to 8 bits apart
Loosely similar	More than 8 bits apart

The Tightness range control in the results filters by these bands. It is a range, not a floor: you can look at only the loose groups, which is a good way to audit whether your threshold is set too high.

Tip

The band uses the group’s widest pair, so a group reading Similar may contain two frames that are nearly identical. It describes the group’s spread, not its closest match.

Next

Reading the results — the grid, the filters, and how to select across groups.

Then

AI Curation ranks a group and names what it measured.