A model does not know that your backpack has one zip. It knows that backpacks usually have two, and when the second one is hidden in shadow, or out of frame, or simply absent, it draws it — neatly, in the right place, with a plausible pull. Nothing about the image says "invented". The customer orders the bag with two zips and receives the bag with one, and the return is not a photography error in their eyes; it is a lie.
This page is a method for preventing invented details in AI product photography. It publishes no accuracy figure for any tool. What it gives is the reason invention happens — a gap the model fills with the typical — the source and workflow choices that leave no gap, and a check that counts features rather than judging the picture.
How do you stop AI from adding details a product does not have?
Leave it nothing to complete. Photograph the product so that every feature it has is visible and every feature it lacks is visibly absent: every side, the underside where a channel shows it, closures and ports in focus, nothing hidden by shadow, a hand or a prop. Use workflows that keep the product's own pixels — a cut and a placement, a scene generated around the product — and never a re-render from a description, which invites the typical. Change one thing per image so an invented feature can be traced to the step that produced it. Then check by counting: closures, seams, pockets, ports, straps, buttons, holes, stitches where they matter, compared to the photograph, and reject any image where the count differs.
What is an invented detail?
An invented detail is any feature in a generated image that the physical product does not have: an extra closure, seam, pocket, strap, port, button, hinge, vent, logo or panel; a symmetrical twin of a feature that exists on one side only; a completion of a part that was hidden in the source. It differs from a changed detail — a button moved, a neckline deepened — in adding something rather than altering it, and from a scene error in being on the product. It is the error a customer can hold you to.
Why models invent
- They complete from the typical. A gap in the source — shadow, crop, occlusion — is filled with what such products usually have there.
- They prefer symmetry. A feature on one side is mirrored to the other; an asymmetric product becomes symmetric.
- They regularise counts. Three buttons where four would be evenly spaced; two zips where one looks unfinished.
- Descriptions invite them. "A leather backpack on a wooden bench" as a one-step request draws a leather backpack — a typical one, with typical features — not yours.
- Several changes at once hide the culprit. A new scene, a new angle and a new light in one generation give three chances to invent and no way to tell which did it.
- Hidden parts are open questions. The inside of a bag, the back of a device, the underside of a shoe: anything not photographed is anything.
Photograph so nothing is missing
- Every side the listing will show. Front, back, both sides, top, and the underside where the channel shows it. A side not photographed is invented if it ever appears.
- Closures, ports and hardware in focus. Each one photographed so its count and position are obvious on the source.
- Nothing hidden. No hand over a clasp, no strap across a pocket, no shadow across a seam. What the source hides, the tool completes.
- The product as it is sold. Straps attached, flaps closed, the configuration the customer receives; a product shown in a way it is not sold is a promise.
- Detail photographs. Close views of anything small — a zip pull, a port row, a stitch line — for the inspection.
- Even light, full resolution. A shadowed feature is a gap; upload at the camera's original size, up to 10 MB in Photta.
Generate so nothing is invited
- Keep the product's pixels. Cut and place, or generate the scene around the product; in Photta the background remover, studio product photography and the lifestyle scenes work this way. The product in the output is the photograph.
- Never describe the product. Describe the scene, the surface, the light. A request that names the product's features asks for a product drawn from those words.
- One change per image. White ground first; then the scene; then a new angle only from a source photographed at that angle. An invention then has a step to be traced to.
- Match the source's angle. A top-down source cannot become a three-quarter view without the tool inventing the sides.
- Top resolution for the inspection copy. Counting a stitch line needs the pixels.
Show, keep, count
The loop the checklist below serves, named so the other accuracy pages can point at it.
- Show — a source in which every feature is visible and every absence is visible, from every side the listing uses.
- Keep — the product's pixels carried into the image; the scene generated around them; one change per image.
- Count — closures, seams, pockets, ports, straps and marks counted on the source and on the output, with a mismatch rejecting the image.
Invented-detail checklist
| Count on the source, count on the output | Pass | Fail — regenerate |
|---|---|---|
| Closures | Same number, type and position of zips, buttons, clasps, snaps | One added, mirrored or regularised |
| Pockets and panels | Same count and placement | A pocket added for symmetry; a panel line invented |
| Straps, handles, loops | Same number, attached where they attach | A second strap; a loop that is not there |
| Ports, switches, holes | Same count, position and shape | A port row completed; a hole added |
| Seams and stitching | Where the source has them, at the same width | A seam added; stitching invented on a smooth panel |
| Logos and marks | Only the marks the product carries | A mark added where marks usually go |
| Hidden sides | Shown only from a source photographed there | A back or underside invented from the front |
| Symmetry | Asymmetric where the product is asymmetric | Mirrored to symmetry |
| Configuration | As sold: flaps, straps, accessories | An accessory shown that is not included |
Count, do not judge: an invented zip looks right, which is why it survives a glance and fails a count.
One backpack, one afternoon
A bag brand photographs a canvas backpack on white with a grey card: front, back, both sides, top, base, with the single main zip, the two side pockets and the one chest strap all visible and in focus, plus close photographs of the zip and the strap clip. The main image comes from the studio workflow; the count against the source is one zip, two pockets, one strap, one clip, and it passes. The lifestyle scene — the bag on a wooden bench in soft light, generated around the kept product — passes the same count. A three-quarter view requested from the front source shows a second side pocket where the hidden side would be; the count fails, and that view is regenerated from the side photograph instead, where the real pocket is. Every image published shows the bag that ships.
Frequently asked questions
Why does AI add features my product does not have?
Because it completes what it cannot see with what such products usually have: a hidden side, a shadowed closure, a cropped strap becomes the typical version. It also prefers symmetry and regular counts, and a request that describes the product invites a typical product to be drawn. Show every feature and every absence in the source, keep the product's pixels, and describe only the scene.
How do I check a generated product image for invented details?
By counting, not judging: closures, pockets, straps, ports, seams and marks counted on the source photograph and on the output, with any mismatch rejecting the image. Check hidden sides only against a photograph taken there, and compare at the top resolution so stitching and hardware can be counted.
Is re-rendering a product ever acceptable?
Not for a product image. A re-render draws the product from a description and adds the typical features of its category. It is acceptable for a mood board where no customer will compare the picture to the delivery.
What should I photograph to stop invented details?
Every side the listing will show, including the underside where the channel shows it; every closure, port and piece of hardware in focus; nothing hidden by a hand, a strap or a shadow; the product in the configuration it is sold in; close photographs of small parts for the inspection.
Can I remove an invented detail in an editor?
Remove, cautiously, if the surrounding pixels are the photograph's; but a patch on a generated area is a fabrication on a fabrication. The safer fix is to regenerate from a source that shows the area, with one change per image, and count again.
What this page does not claim
That any tool avoids invention at a given rate, or that accurate images lower returns by any amount; this page measures neither. What it claims is that invention fills gaps, that a source with no gaps and a workflow that keeps the product's pixels leave nothing to fill, and that a count catches what a glance approves.
To keep a product's pixels through a scene change, see AI product photography; for the measurement behind the counting, product geometry accuracy in AI photography; for the same discipline on garments, how to prevent AI from changing clothing details.