Most rejected product images are rejected with the same sentence: it looks off. That sentence sends the operator back to the tool to try again, and the next image looks off in a different way. What is missing is a name. A reflection that shows the old studio, a second zip, a shadow falling the wrong way, a soft edge from a soft source and a portrait crop on a square channel are five different failures with five different fixes, and none of them is fixed by trying again.
This page is a taxonomy: every way a generated product image fails, sorted into five classes by what went wrong, with the cause and the fix for each. It publishes no failure rate for any tool or class. What it gives is a vocabulary for the review — so a rejected image is rejected with a reason that points at a step — and an order in which to look.
What are the main ways AI product photography fails?
Five: the product changed (its colour, shape, texture, proportion or a detail it has is different from the photograph); something was invented (a feature, a mark or a side the product does not have); the scene disagrees with the product (light, shadow, scale, contact, reflection or perspective that do not match); the source was not good enough (blur, shadow, a hidden side, a low resolution, the wrong ground, so the tool had to guess); and the image is wrong for the channel (crop, ground, size or a rule the marketplace enforces). Each class has its own cause and its own fix, and the review looks for them in that order because the first two are the ones a customer can hold you to.
Why classify failures at all?
A failure with a name has a fix: a changed colour points at the colour reference and the check; an invented pocket points at the source's hidden side; a wrong shadow points at the scene's light. A failure without a name points at nothing, so the operator regenerates with the same source and the same request and gets the same class of failure with different details. Classifying also makes a batch reviewable: ten images rejected as "off" are ten problems; ten images rejected as "class 3, light disagreement" are one problem in one step.
Class 1 — the product changed
The product in the image is not the product in the photograph.
- What it looks like: a colour shifted against the reference; a proportion stretched; a texture smoothed or sharpened; a detail moved — a button, a seam, a neckline; a pattern re-spaced; a logo redrawn; a label's letters replaced with letter shapes.
- Cause: the product's pixels were redrawn — a re-render from a description, an angle the source did not have, an upscale asked to invent detail, several changes in one step.
- Fix: a workflow that keeps the product's pixels and generates around them; one change per image; a colour reference in the master; a comparison against the master at full size.
- Severity: a customer receives a different product. Reject.
Class 2 — something was invented
The image shows a feature, a mark or a side the product does not have.
- What it looks like: an extra closure, pocket, strap, port or seam; a feature mirrored for symmetry; a mark where marks usually go; a back or an underside drawn from the front; an accessory that is not included.
- Cause: a gap in the source — a hidden side, a shadowed closure, a crop — that the tool completed with what such products usually have; a request that described the product; a new angle from a source that did not show it.
- Fix: a source in which every feature and every absence is visible, from every side the listing uses; describe the scene, never the product; count features on the source and the output.
- Severity: a customer orders what does not exist. Reject.
Class 3 — the scene disagrees with the product
The product is right, the scene is right, and they do not belong together.
- What it looks like: the shadow falls toward the light; a highlight on the product from a side the scene has no light on; a reflection showing the old studio; a product floating above the surface or sunk into it; a product the size of a building or a toy; a perspective where the ground recedes and the product does not; a colour cast the scene cannot explain.
- Cause: the scene's light chosen without regard to the master's; a reflective or transparent product carrying its old room; a contact shadow generated apart from the product; a scale the request never fixed.
- Fix: scenes lit from the master's highlight side; a neutral surround for reflective products, the publishing ground for transparent ones; a shadow checked for direction, softness and contact; a size reference in the scene where scale matters.
- Severity: the image reads as fake before anyone can say why; trust, not the product, is what the customer loses. Reject, and fix the scene, not the product.
Class 4 — the source was not good enough
The tool had to guess because the photograph did not say.
- What it looks like: soft edges where the source was soft; a shadow side smoothed into nothing; a low-resolution product in a sharp scene; a colour that cannot be checked because there was no reference; a hidden side that becomes class 2 the moment a new angle is asked for; a white master pressed into a dark scene for a transparent product.
- Cause: the photograph — blur, uneven light, a hidden part, a small file, the wrong ground — not the generation.
- Fix: a reshoot. Even light, full resolution, every side, a colour reference, the ground the image will use. In Photta, uploads up to 10 MB keep the camera's original detail.
- Severity: every downstream image inherits it. Fix the source before generating anything else.
Class 5 — right image, wrong channel
The image is correct and the channel will not accept it or will show it badly.
- What it looks like: a main image on a scene where the marketplace requires pure white; a portrait crop shown in a square tile with the product cut; the product too small in the frame for the channel's zoom; a shadow or a prop the category rules forbid; a file size or format the channel rejects.
- Cause: the image made before the channel's rules were read.
- Fix: the channel's current requirements read from the channel itself, and a per-channel export: white for the main image where required, the crop and the fill the tile needs, the format and size the upload accepts.
- Severity: a rejection or a bad tile, and no harm to the customer. Re-export, do not regenerate.
Name, place, fix
The loop the checklist below serves, named so the other accuracy pages can point at it.
- Name — every rejected image tagged with its class and the row that failed.
- Place — the class mapped to the step: source, workflow, scene, check or export.
- Fix — that step changed, and the image regenerated once; a second failure in the same class means the step, not the image, is wrong.
Triage order for a batch
| Look for | Class | Ask | If it fails |
|---|---|---|---|
| Colour, shape, texture, detail against the master | 1 — product changed | Is this the product in the photograph? | Reject; keep pixels; one change per image |
| Counts of closures, seams, marks, sides | 2 — invented | Does the product have this? | Reject; show every side and absence in the source |
| Light, shadow, reflection, contact, scale | 3 — scene disagrees | Do the product and the scene share a light? | Reject; relight or reshoot the scene, not the product |
| Sharpness, resolution, hidden parts, reference | 4 — source | Could the tool have known? | Reshoot; every later image inherits this |
| Ground, crop, size, category rules | 5 — channel | Will the channel show it? | Re-export; do not regenerate |
Look in this order because the first two are the failures a customer can hold you to; the third loses trust; the fourth explains the rest; the fifth costs an export.
Reviewing at scale: sampling, exceptions and the record
The taxonomy is written for a reviewer looking at one image. A catalogue of thousands changes the question from "is this image right" to "which images does a person need to see", and the answer is a sampling rule, an exception rule and a record.
- Sample by product, not by image. Every product is seen at least once — its main image against the master — because the first two classes are per product and a changed colour or an invented feature repeats across every image of that product. Scenes and variants are sampled.
- Route the exceptions in full. Products whose category fails quietly — reflective, transparent, fine text, jewellery, layered garments — go to review in full regardless of the sample, because their failures are the ones a glance approves.
- Widen the sample when a class appears. One class-two rejection in a batch means the source setup that produced it is suspect; review every product photographed the same way.
- Name every rejection. A rejection without a class is a regeneration with the same result; a rejection with a class is a change to the source, the recipe, the scene or the export.
- Keep the record. Which images were seen, by whom, against which master, with which class and fix. It is what makes the next batch's sample smaller and the one after that trusted.
A review that sees every image is not a stricter review; it is one that will be skipped the first week the catalogue is late. A sample that always sees the product and always sees the exceptions is the review that survives.
One batch, five names
A homeware brand reviews a batch of a stoneware mug: twelve images from one white master with a grey card. Two show the glaze a shade warmer than the card — class 1, the scene's light was allowed to touch the product; the workflow is changed to keep the pixels and both are regenerated. One shows a handle on the far side that the mug does not have — class 2, the source never showed the back; the back is photographed and the angle regenerated from it. Three have a shadow falling toward the window — class 3, the scene was lit from the wrong side; the scenes are relit from the left. One is soft at the rim — class 4, the master's focus; the master is reshot and the batch repeated. Two have the mug too small for the marketplace's square tile — class 5; they are re-cropped. Three pass. Every rejection has a step, and the next batch fails less because the steps, not the images, were fixed.
Frequently asked questions
What is the most serious failure in AI product photography?
A changed or invented product — classes 1 and 2 — because the customer can compare the image to what arrives. A scene that disagrees with the product loses trust without misrepresenting the product; a poor source explains the others; a channel mismatch costs an export.
Why do my AI product images "look off" when nothing is obviously wrong?
Almost always class 3: the product and the scene do not share a light. The shadow falls the wrong way, the highlight comes from nowhere, a reflective surface shows the old studio, the product floats. Check the light direction on the product against the scene's shadow before anything else.
Should I regenerate a failed image or fix the source?
Depends on the class. Classes 1 and 2 are fixed in the workflow and the source; class 3 in the scene; class 4 only by reshooting; class 5 by re-exporting. Regenerating with the same source and request repeats the class.
How do I review a large batch quickly?
In triage order, with names. Compare each image to its master at full size, tag it with the first class it fails, and group the rejections by class: a batch that fails mostly in one class has one step to fix.
Does this taxonomy apply to virtual try-on and model images too?
Yes, with the garment or the product as the thing that must not change and the model as part of the scene. The classes are the same; the counts in class 2 become buttons, seams, prints and hems, and class 3 adds the fit and the drape agreeing with the body.
What this page does not claim
That any class occurs at any rate, in any tool; this page measures nothing. What it claims is that every failure belongs to one of five classes, that each class points at one step of the workflow, and that a review which names the class fixes the step instead of the image.
To keep the product's pixels through a scene change, see AI product photography; for the measurement behind class 1, how to measure AI product image accuracy; for class 2 in depth, how to prevent AI from adding nonexistent product details.