Why scores run high here
A detector looks for signs that pixels were produced or reshaped by software rather than sampled from the world. Instagram is a place where almost every image has been reshaped by software on purpose, and often several times.
A typical published photograph has been through a camera app that stacks exposures, an editing app that adjusts tone and removes blemishes, a preset that shifts colour globally, and finally the platform's own recompression on upload.
Each of those steps rewrites pixel relationships. None of them makes the picture fake, and all of them look, to a texture model, a little like the thing it was trained to find. The result is a genuine photograph sitting comfortably in the forties.
This is not a flaw specific to one tool. Any detector reading pixel statistics faces the same problem, and any that claims high accuracy on heavily edited social content is measuring on a benchmark that does not contain much of it.
The fourth row is the interesting one, because a replaced sky genuinely is a generative edit. The score rose for a real reason there, and the region map would show it confined to the top of the frame rather than spread across it.
The distinction that matters: edited or generated
Most Instagram images that trip a detector are photographs that have been edited, not pictures that were invented. Those are different claims with different consequences, and the region map is what separates them.
A generated image produces heat across the whole grid, because every part of it came from the same process. An edited photograph produces a cool frame with hot patches where the generative tool did its work: a removed object, a replaced background, an extended edge.
Generative fill is now built into mainstream editing software, so this pattern is common and usually harmless. Removing a stranger from the corner of a holiday photo is not deception, and a tool that reports it as such is answering a question nobody asked.
Getting a usable file
Instagram makes this harder than most platforms, because it does not offer a download and the displayed image is already resized. The best available copy is usually the one the browser loads on the web version rather than anything the app shows.
Pasting the post URL is generally the most reliable route. It retrieves the served image directly rather than a re-render of your screen, which avoids adding a second round of compression on top of the platform's own.
Where the account is private, there is no public file to fetch and no honest way around that. A screenshot is the only option, and a screenshot of an already-filtered, already-recompressed image is about as degraded as a test subject gets.
What actually indicates a generated account
For accounts built entirely on synthetic imagery, the pattern across the grid tells you more than any single image. Consistency is the giveaway, and it is the opposite of what people expect.
- Impossible consistency. The same face, lit identically, across supposedly different days, cities and seasons.
- No incidental photographs. Real accounts accumulate badly framed, badly lit, unimportant pictures. Fabricated ones are all hero shots.
- Backgrounds that repeat. The same wall texture or foliage pattern appearing in unrelated posts.
- No other people's cameras. Nobody has ever tagged them in a photograph they did not take themselves.
- Comments that do not fit. Engagement that never references anything specific in the picture.
Every one of those is about the set rather than the frame, which is why a single check on a single post is such a weak instrument here. Run several images from the same account and look at the spread before drawing any conclusion.
If you are the creator being questioned
This happens increasingly to photographers and illustrators, and the defence is provenance rather than argument. Keep the raw file, keep the layered project file, and keep the version history your editing software already records.
A raw file with camera metadata and a matching edit history is far more persuasive than any detector result, in either direction. It is also the only evidence that scales, because it does not depend on somebody else's model agreeing with you.
Sponsored posts and product photography
Commercial content on the platform is where generated imagery has moved fastest, and it is also where the stakes are highest for a viewer. A product shot that shows an item which does not exist as photographed is a consumer problem, not an aesthetic one.
The tells differ from portraits. Look at the surface a product sits on, the shadow it casts, and whether the reflections in packaging agree with the room implied by the background. Generated product photography tends to get the object right and the physics of its setting approximately right.
Text is the other reliable check here. Packaging, labels and small print are still where generation struggles, and a product whose own label does not survive close reading is worth a second look regardless of what any score says.
For a purchase decision, the check that matters most is not detection at all. Ask the seller for an unedited photograph of the item on a plain surface, taken today. That request is trivial for a real seller and awkward for anyone whose imagery came from a prompt.