Why generating the whole thing is not the answer
It would be tempting to hand the product over to the model as well: describe it, get a photo of it. There are two reasons we do not do that. The first is that the buyer receives what they saw in the picture — if the generated product is a shade off, the handle sits somewhere else or the label is wrong, that means returns and complaints. The second is that the small markings on a product, the logo and the text on the packaging are exactly what models handle worst.
So the workflow that works starts from a real photograph, and AI only adds the surroundings: background, surface, light, mood, season. The product stays honest, and the variations take minutes.
What the measurements show
In the summer of 2026 a larger measurement of this task was published: 850 products, more than four thousand generations, and the question was how many images clear a webshop quality threshold.
- The best models had a pass rate of around 29 percent — three usable images out of every ten generations.
- The weaker, cheaper models came in at around 17 percent.
- The most frequent cause of failure was distortion of logos and lettering, accounting for roughly a fifth of all errors.
- The measurement was published by a company that sells a solution to exactly this problem — so treat the numbers as a direction, not as absolute truth.
What is worth taking from it: the output of generation is not finished material but a pool of candidates. If you use three images out of ten, that is not a failure, it is the normal ratio — and the workflow has to be planned around it.
The recipe
This order yields the most usable images from the fewest attempts.
1. ONE good base photo. Diffuse light, neutral background, sharp focus.
A phone is fine, as long as there is no strong coloured cast on it.
2. Cut the product out and keep it together with its mask.
This becomes the basis of every later image.
3. Generate surroundings, not products:
"bright kitchen counter, blurred window behind it, morning light
from the left, clear space around the object"
4. Ask for contact shadow and reflection, in a separate sentence.
Without that the product will float.
5. Check at full zoom: logo, lettering, stitching, material texture.
If any of them is distorted, the image goes in the bin — do not patch it.
6. For a series, regenerate the same background for every product
instead of inventing a new one each time. That makes the shelf line up.
The marking that is mandatory
This is the part most webshop owners skip, and it is not optional. If you upload your product images into an advertising or product-listing system, the large platforms require the image metadata to state that it was made with AI — and they also state that this marking must not be removed.
Technically this is a standard field in the image metadata that most generation tools write in by themselves. The danger is not the writing but the dropping out: many image editors, compressors and webshop engines throw the metadata away on save. So at the end of the workflow you have to check that it is still in the uploaded file. Exactly what each system marks, and how much the marking survives, is set out by watermarks and provenance in its own post.
What you need to know about people
If a person is wearing or using the product, different rules kick in. From the summer of 2026 one large marketplace requires a separate label on images that show a photorealistic, AI-generated person — retouching the product does not fall under this, an invented model does.
In Europe there is the general principle on top of that: artificial content capable of misleading has to be marked. A retouched product photo, or one moved to a different background, does not normally fall into that; a non-existent person praising the product very much does. The practical rule is simple: do not invent the product, and do label the person.
Where it fails most often
- Reflections. On glossy packaging, on glass and on metal the colour of the old background stays put. That has to be fixed separately, otherwise it is visible to the naked eye.
- Scale. In a generated environment the product often ends up larger or smaller than it really is. If there is a reference point in the frame — a hand, a mug, a table — it jumps out immediately.
- Lettering. Text on packaging is the most frequent source of error, which is also why the product should not be generated. The background to that is explained by text rendered into images in a post of its own.
- Consistency across the series. If every product gets its own mood, the category page looks untidy. The answer is the same reference for every item — for that technique, product and character consistency is the one to read, and for the background swap itself, inpainting and background replacement has the specifics.
What to take away
The recipe for good AI product photography in one sentence: the product is real, the surroundings are generated, and the marking stays on the finished file. Those three together are what make your images look good, make the buyer receive what they saw, and keep the platforms from bouncing your upload.