SweeperX
Get SweeperX

Similar isn’t duplicate: why iOS only finds half your redundant photos

The Photos app found eleven duplicates in a library of thirty thousand. It is not broken. It is answering a narrower question than the one you asked.

Two different problems, one word

When someone says they have thousands of duplicate photos, they almost never mean duplicates. They mean their library is full of redundancy, which comes in two quite different shapes.

Exact duplicates are the same image stored twice. You saved a photo that was already there, a restore brought one back, an app re-downloaded something. These files are often byte-for-byte identical. Removing all but one loses nothing at all, which is why iOS is willing to do it for you.

Similar photos are different photographs of nearly the same thing. Nine frames of your friend by the harbour, taken over four seconds, because one of them would have everyone with their eyes open. Every frame is a genuinely distinct photograph. Eight of them are also worthless to you, and together they are eight times the size of the one you wanted.

On most phones the second category is far larger than the first. It is also the one no built-in tool will clear.

Why Apple draws the line where it does

Apple's Duplicates album is conservative on purpose. It ships to a billion devices, and it is allowed to be wrong approximately never.

Grouping exact duplicates is a decidable problem: the images match, so one of them is unambiguously redundant. Deciding that one of nine similar frames is the good one is a judgement about a photograph, and judgements are wrong sometimes. A system that made that call at Apple's scale would delete somebody's only good photo of a person who has since died, and it would do it thousands of times a day.

So iOS finds the safe category and stops. That is a correct engineering decision, and it also leaves the bulk of the problem unaddressed — which is the gap third-party tools exist to fill, honestly or otherwise.

How software recognises “similar”

Comparing every photo to every other photo pixel by pixel is impossible: forty thousand photos is eight hundred million comparisons, each over several million pixels. The technique that makes it tractable is perceptual hashing, and the idea is simpler than the name.

For each photo:

  1. Shrink it drastically — to something like a 32×32 grid. Almost all the detail is thrown away; the broad structure of light and dark survives.
  2. Drop the colour. Convert to grayscale. Structure is what matters, and colour is what changes most between two shots of the same scene.
  3. Reduce it to a fingerprint — a short string of bits describing the relationships between those regions. Brighter here than there, and so on.

Now comparison is cheap. Two fingerprints differing in only a few bits describe images with nearly the same structure. That is the entire trick: a resized copy, a re-saved copy, a screenshot of a photo, or the next frame of a burst all produce nearly the same fingerprint, while a photograph of something else does not.

The threshold is a dial. Set it tight and you find only near-identical images. Set it loose and you start grouping photos that merely resemble each other, which is worse than useless — a review list full of false groupings trains you to tap through without looking, which is precisely how people lose photos.

What that method finds, and what it misses

Being specific about the boundary is more useful than a claim of completeness. Structure-based matching reliably finds:

It does not reliably find:

That last group is a feature rather than a shortcoming. A tool that grouped every photo of your kitchen taken over five years would be presenting you with a list nobody can usefully review.

Bursts, and the pile iOS hides from you

There is one case where iOS does group similar photos, and it is worth knowing because it hides the size of the problem rather than solving it.

When you hold the shutter (or hold Volume Up, if you have enabled that in Settings → Camera), the camera fires a burst — ten frames a second for as long as you hold it. In your library that appears as one photo with a stacked edge. Every frame is still there.

To deal with a burst properly:

  1. Open it and tap Select… at the bottom.
  2. Swipe through the frames. iOS marks its own suggestions with a grey dot.
  3. Tick the ones you want and tap Done.
  4. Choose Keep Only Favourites — the rest are deleted.

The trap is that a burst looks like a single photo in every count your phone shows you. A library that reports twelve thousand photos might hold twenty thousand frames. If your photo count seems low relative to how full the phone is, unexpanded bursts are a strong candidate.

Picking the keeper out of nine frames

Once a group exists, something has to nominate a survivor, and there are only a few honest signals available on-device:

What no on-device tool can assess is whether anyone was blinking, whether the frame it chose is the one where your father was laughing, or whether the technically sharpest photo is the one you would have picked. Those need a person, which is why the only defensible design is a suggestion you can overrule — not a decision.

The round trip that makes a worse copy

One pattern is worth calling out because it is both extremely common and the one where deleting the wrong file has a permanent cost.

You photograph something. You send it to someone through a messaging app, which recompresses it to a fraction of the size. They send it back, or you save it from the thread. Now your library holds two images that look identical at thumbnail size: your original at 4 MB and a 180 KB copy that has been through a lossy encoder.

Structure-based matching pairs them correctly — they have the same fingerprint, which is exactly what the technique is for. What matters is which one survives. A tool that keeps the most recent file keeps the degraded one. A tool that keeps the largest or highest-resolution file keeps the photograph.

This is the strongest practical argument for reviewing rather than automating. The two frames look the same in a grid, they are not the same, and the difference only becomes visible when you crop, print or zoom — long after the 30-day recovery window closed.

Why this category deserves more caution, not less

The similar-photo category is where cleaner apps earn their money and where they do their damage. It involves more files, more judgement, and more opportunity to be wrong than exact duplicates ever do.

The correct response to that risk is not to avoid the category — it is genuinely where your storage went — but to insist on the safeguards:

An app that offers a single button to clean everything is offering to make thousands of photographic judgements on your behalf, at speed, with no review. That is not efficiency.

What to do about it

Run the built-in Duplicates album first. It is free and it is the safest pass available. Then decide whether the similar-photo problem is large enough on your phone to be worth a tool at all — on a library of two thousand photos, probably not; on forty thousand with a decade of bursts in it, almost certainly.

SweeperX groups similar photos by structure on the device, marks the sharpest frame in each group as the keeper, and shows you the group's size so you can skip the ones not worth your time. It says on the screen that the same subject shot from a different angle may not be grouped, because a tool that overstates what it found is a tool you will eventually stop reading. Everything it selects goes to a review bin you confirm once, and from there to Recently Deleted for 30 days.

Next: How to Compress a Video on iPhone Without Losing Quality — where the gigabytes really are, and what a codec actually throws away.