Listing photography
Why AI listing photos look fake
You have probably noticed it without being able to name it. The reason is not the model being bad at its job — it is being asked to do the wrong job.
Almost every tool redraws the whole frame
Give a generative model a photograph of a living room and ask it to tidy the room, and what comes back is not your photograph with the clutter gone. It is a new photograph of a room that resembles yours. Every pixel has been generated, including all the ones you never asked anybody to touch.
That is why the failures below are so consistent across tools. They are not bugs in one product. They are what regeneration does.
Brickwork turns to smear
A model asked to redraw a wall produces something brick-like. At thumbnail size that reads as brick; at full size it reads as texture, because there was never a course of bricks — only an impression of one.
Window frames drift
Straight lines are the first thing to go. A mullion that was one continuous piece comes back very slightly kinked, and the eye notices long before it can explain why.
Floor grain repeats
Regenerated wood tends to tile. Once you see the same knot twice in one room you cannot unsee it, and neither can a buyer looking at their fourth listing that evening.
Text becomes nonsense
House numbers, appliance dials, spines on a shelf. Anything with letters is where a redrawn frame gives itself away fastest.
The light stops agreeing with itself
Shadows fall one way, the window falls another. Individually plausible, collectively impossible — which is the exact feeling people describe as looking fake without being able to point at it.
The part that matters commercially
It survives the thumbnail and fails the buyer
A regenerated frame looks fine in a search results grid, which is where most people judge it. It falls apart at the two moments that actually matter: when a buyer opens it full-screen on a laptop, and when it goes into print for a brochure or a window card.
Portals serve large images and buyers pinch in. Detail that was never really there has nowhere to hide.
The difference
Correcting a photograph is not the same as regenerating one
Geometry and lighting do not need a model at all. Lens distortion, a tilted horizon, leaning verticals, exposure, white balance — all of those are arithmetic on the pixels you already have. They can be computed and applied without inventing anything.
Only the parts that genuinely require new pixels — the gap where a laundry basket used to be — need generating. If generation is confined to those regions and composited back through a mask, then the walls, windows, floors and rooflines in the result are the ones your camera recorded. Bit for bit.
That is the whole difference, and it is architectural rather than a quality setting. A tool that redraws the frame cannot be configured into a tool that does not.