AI product photo accuracy: stop invented details

Accurate is not the same as good-looking

Plenty of AI images look better than the photography they replace. That is not the bar. The bar is whether the thing in the frame is the thing in the box. Four parts of an image can fail independently, and we score all four separately before a visual leaves our studio:

• •Geometry: shape, proportions and how the parts relate to each other
• •Texture: surface, weave, gloss and grain
• •Colour: hue, saturation and how it behaves under the light in the scene
• •Markings: logos, print, stitching, controls and any text on the product or pack

Why four scores instead of one? Because an average hides the failure that matters. An image that nails colour and texture but puts your logo on the wrong panel is not seven out of ten. It is unusable.

The cost lands at unboxing. Either the product matches the image it was bought from, or you get a return, a review, or both. Sell through dealers and retailers and the damage multiplies: one wrong image can sit on dozens of websites you do not control within a week.

Where AI quietly gets it wrong

It invents parts

A seam that does not exist, a vent on a panel that is solid in real life, a second logo on the side. This happens when the model falls back on its training data because your input did not tell it enough. It is an input problem, not a prompt problem, and nothing further down your process fixes it.

Proportions drift

Two versions of the same fault. Inside the product: a cap too large for the bottle, a table leg too slender under a heavy top. Against the room: a bath that reads as a hip bath, a cabinet that suddenly reaches the ceiling. The second is the bigger problem in lifestyle imagery, because people register it without being able to name it. They just trust the image a little less.

Colour shifts

Half a shade off is invisible in isolation and obvious next to the real swatch, or next to the same product on a dealer site. If you sell colour variants, this is where returns come from, because the colour was the thing the customer chose.

Text becomes almost-text

Pack copy, size markings, ingredient lists. Many models still produce something shaped like writing that falls apart the moment anyone reads it, and that makes a buyer question everything else on the page.

The one rule: let AI build the world, not the product

If you keep one thing from this article, keep this. Your packshot is the evidence of what the product actually looks like. Everything around it is fair game for AI: background, light, the setting, a different crop per channel. The product itself does not get regenerated. Keep the original unedited file, so you always have something to hold every generated image against.

That is also the practical difference between a packshot and a lifestyle image, and why you need both of them for different jobs. We wrote about that in packshot or lifestyle: which image sells where.

Eight checks that keep a product true

1. Start from a real photograph, not a description

The more the model can see of your product, the less it has to invent. Supply several angles, a close-up of the material and a clean shot of the logo, sharp and at full resolution. A 600 pixel web export is a poor starting point. How many pixels each channel needs is in product image resolution: what you need per channel.

2. Hand over your product as facts, not adjectives

Words like charcoal, compact or matte leave too much room. Build a short spec sheet per product covering:

• •dimensions in millimetres: height, width and depth
• •colour as a hex or Pantone value, not as a description
• •material and finish: matte, gloss, brushed, oiled
• •the position and size of the logo
• •the literal text printed on the product or the pack
• •the parts that are explicitly not on it

3. Put a known size in the frame and do the arithmetic

Place something in the scene whose dimensions are fixed, then check the ratio rather than trusting your eye. A 170 cm bath in a bathroom tiled at 60 by 60 should span almost three tiles. A dining table at 90 cm sits just above hip height on an adult. If the ratio is wrong, the image is wrong, however good it looks.

4. Test your hardest product first

Gloss, transparency, fine pattern, deep black, small printed type and brand-critical colour break a model fastest. Start there. If it holds up on your worst case, the rest of the catalogue usually follows. Start with your easiest product and you learn nothing.

5. Generate the same image four times

One good result proves nothing. Four runs of the same brief with the same settings tell you whether you have a method. If proportions move between attempts, you got lucky once, and luck does not scale to three hundred SKUs.

6. Check colour by value, not by feeling

Pick the hex value out of the generated file and hold it against your brand or product colour. Agree in advance how much deviation is acceptable, so the conversation afterwards is about a number rather than taste.

7. Read every word on the image out loud

Literally, character by character. Where there is any doubt, lay text, logos and size markings over the image as vector afterwards instead of asking a model to render them.

8. Publish the dimensions as their own image

A plain dimension drawing with height, width and depth removes a whole category of complaint. Take those numbers from your product data, not off the generated image, or you are publishing an estimate of your own product.

How to test a tool in an afternoon

To know how accurate a tool is for your assortment, treat it as a sample rather than an impression. A workable setup:

• •Take around thirty products from your real catalogue, including the difficult categories
• •Keep the settings identical for every product so you are comparing like with like
• •Score the four parts separately for each image
• •Use three levels: true, off but recognisable, and wrong
• •Have two people score independently and record where they disagree
• •Report the spread per part, not one combined number

The closing question for every image is the same: would a customer unpacking this product feel misled? If the answer is yes, the image is unusable, however good it looks in the grid.

What marketplaces and the law expect

Accuracy is no longer only a matter of craft. It is increasingly a condition of selling:

• •Amazon expects the main image to show the actual product. AI around the product, such as a background or a styled scene, belongs mainly in the supporting images.
• •Google Merchant Center accepts AI imagery but expects the metadata that flags it as generative AI (the IPTC DigitalSourceType tag) to survive. Do not strip it on export.
• •Since 2 August 2026, the transparency obligations in article 50 of the EU AI Act apply to images that are generated or manipulated with AI and could pass for real.

Rules differ per channel and change quickly, so check what applies to your own sales channels before you replace a whole catalogue. The European Commission has published guidance on article 50. On rights and usage policy for AI imagery we wrote this piece.

How we handle this at LensLab

Every visual we make starts from your own product image. That is not an intention but a condition in the platform: without a product photo you cannot generate, and a Visual Studio Pro request opens with your reference images or the link to your product page. Who sits at the controls after that is your choice. With Visual Studio Pro a creator from our own team builds the visuals and nothing reaches you without being checked, with human review through to delivery. Work in the Visual Generator yourself and that check is yours.

One thing we would rather say up front, because it bears directly on this subject. Your product is re-rendered, not cut out of your photograph and pasted in. That is why your reference material decides so much, and why checking the failures listed above stays necessary. More angles, a close-up of the awkward part and your dimensions in the brief move the needle further than any turn of phrase in the prompt.

Alongside that we work from your Brand DNA, so style and tone stay the same across the assortment instead of moving with whoever wrote the brief that day. Mind the distinction though: your brand colours drive your identity, not the colour of the product in the frame. That comes from the image you supply. A hex value in your brand book is no guarantee of product colour, which is exactly why check six above is there. More on consistency at scale in product photography at scale.

We also deliver more options than you need, and you only pay for the visuals you accept. Options you turn down cost nothing. That is what makes accuracy enforceable: you can reject an image because the proportions are off without it costing you anything.

A concrete example of why proportion matters. BosGrill sells kamado barbecues: heavy, awkward to move, and no longer sellable as new once they have been through a shoot. Every size and colour variant needs its own imagery, and a compact kamado belongs in a different setting than an XXL model. Put them in the same garden at the same scale and both images start lying.

As their owner Sjouke put it: "LensLab gave us a complete, consistent library of lifestyle visuals that actually elevate the brand. Every kamado now has a setting that fits its size and character, without us ever having to move a single barbecue." More of what we make for brands is on our customer cases.

And to be straight about it, not everything comes easily on our side either. Jewellery with fine detail, transparent glass and clothing on a model stay harder than a flat product with a clear silhouette. We would rather say that up front than have you discover it at delivery. For those categories a hybrid approach usually wins: AI for scale and consistency, a camera for the shots where the feel of the material is the whole argument. That was one of five honest lessons from hundreds of AI product photos.

Once an image is right, you want your dealers and retailers using that exact file rather than an older one from their own folder. That is what our distribution module is for.

Want to try it on your most difficult product? You start with five free credits, good for five images. What it costs after that is on our pricing page, and the process itself is on see how it works.