What a region heatmap shows that a score cannot

Whether the whole frame is synthetic, or one part of a real photograph was changed. Those are different problems with different consequences, and a single score reads them the same.

The problem with one number

Consider two images. The first is generated end to end and every part of it is synthetic. The second is a real photograph of a real room with a crack added to one wall.

A whole-frame model reduces each to a single figure. The first scores high, correctly. The second scores low, because eight ninths of it came from a camera, and the average buries the edit.

The second image is the one that costs somebody money. It is the insurance claim, the altered receipt, the doctored listing photo, the news image with a detail changed. A tool that reports one number will clear it.

The three patterns

Once the frame is scored in tiles, three shapes appear, and each means something different.

Every tile high

91 94 89 96 93 90 88 95 92

Consistency is the signature. Real photographs almost never score this evenly, because real scenes have varied texture.

A wholly generated frame. No part of this image came from a camera, so no tile is cold.

One tile high, the rest cold

14 11 16 12 88 13 9 15 10

The cold tiles are genuine camera capture. The hot one is not. This is the pattern a single number destroys.

A real photograph with an inserted object. The whole-frame score was 31, which would have read as probably real.

Evenly warm, nothing hot

58 61 54 63 59 57 55 62 60

Uniformly middling. Global processing raises everything a little and nothing a lot, which is the fingerprint of software rather than fabrication.

A genuine photograph that has been heavily processed: night mode, noise reduction, then recompression.

How to read a map properly

Look at the spread before the peak
The gap between the highest and lowest tile matters more than the highest value. A wide gap points to a local edit; a narrow one points to whatever happened to the whole frame.
Check whether hot tiles are adjacent
Real edits are contiguous, because objects are contiguous. Two hot tiles next to each other is a plausible insert. Two at opposite corners usually is not.
Compare against the picture
Look at what is actually in the hot region. If it contains the damage, the price, the face or the object under dispute, that is meaningful. If it contains empty sky, it probably is not.
Discount tiles with little detail
Flat areas of sky, wall or water carry less signal and score less reliably. A hot tile covering a blank wall deserves less weight than one covering a detailed object.
Treat conflict as uncertainty
A low frame score with several scattered hot tiles is usually compression rather than editing. That is a reason to find a better copy, not to split the difference.

When no map is produced

Below roughly 576 pixels on the shorter side, the frame cannot be divided into regions with enough detail to score independently. You get a whole-frame number only.

That is a real limitation on thumbnails, profile pictures and images pulled from chat apps, which is unfortunately where a lot of suspicious images arrive. Finding a larger copy of the same picture is usually the single most useful next step.

Why this matters more than accuracy

Improving a whole-frame model from 91 to 93 percent changes very little for somebody checking a claim photo. Being able to see that one corner of the frame behaves differently from the rest changes the finding entirely.

The region view also makes a result explainable, which matters wherever a decision has to be justified later. A moderator, a claims handler or an editor can point at a specific part of an image and say why they looked harder at it. Nobody can do that with a percentage.

Why overlapping tiles rather than a plain grid

A naive grid has a weakness that shows up immediately in practice. An edit sitting on a boundary is split between two tiles, and each one sees half the evidence against a full tile of genuine pixels. Both come back moderate, and the finding disappears.

Overlapping the crops fixes it. Every point in the frame falls fully inside at least one region, so a small insert is scored against a tile it dominates rather than one it shares. The cost is running the model more times, which on a local device is a second rather than a bill.

It also explains why adjacent tiles often light up together. A single inserted object that sits across two overlapping regions raises both, which is why adjacency is a supporting signal rather than a sign of two separate edits.

The practical consequence is that tile counts vary with the shape of an image. A wide panorama is divided differently from a square portrait, and the grid dimensions are reported with every result so you can see how finely the frame was cut.

Questions people ask

What does a heatmap on an AI image detector show?

How the synthetic signal is distributed across the frame rather than averaged over it. Each tile is scored on its own by the same model, so you can see whether the whole picture looks generated or only one part of it does. Those two findings mean very different things.

Does a hot region mean that part was definitely edited?

No. It means the model found the strongest signal there. Flat areas, heavy compression in one corner, or a region with unusual texture can all raise a tile without any editing. Treat it as a pointer to look with your own eyes, not as a conclusion.

Why is the whole-frame score lower than the highest tile?

Because it is computed on the whole frame rather than being the maximum of the tiles. A large genuine area pulls the overall reading down. That is exactly why both numbers are reported instead of one.

How many regions are scored?

It depends on the shape and size of the image, typically between six and twelve overlapping tiles. The grid dimensions are shown with the result so you can see how finely the frame was divided.

Can I use the map to prove which part was edited?

No, and this is the most common overreach. The map shows where a statistical signal is concentrated, which is not the same as an outline of a manipulation. It supports a human looking closely at a specific area; it does not stand as a finding on its own.

What if every tile is different with no pattern?

That usually means the image has been through enough processing that the regional signal is noise. Scattered values with no adjacency and no relationship to the content are a sign the file cannot support a confident reading at all.