The Boxing Day brief that arrived 19 December
I sell kitchen goods, small electronics and casual clothing through Shopify—ceramic bakeware, electric kettles, hoodies with screen-printed designs, the sort of mid-price items people buy as late gifts or January-sale treats. I'm a sole trader registered with Companies House, I photograph most products myself, and I'd built a decent library of shots over two years. Clean white backgrounds, good lighting, multiple angles.
Then on 19 December my supplier told me they'd added eighteen new SKUs for Boxing Day—teapot sets, a compact air fryer, two new hoodie colours—and could I list them by the 23rd? I had product samples arriving on the 21st and no time to shoot. I'd read about Stability AI generating product images but never tried it. I opened Kryotta, picked Stable Image Core, and spent three days making 42 images. Twenty-six went live on Boxing Day. By 15 January I'd tracked which ones led to clean sales and which triggered "not as described" returns.
This is what worked, what didn't, and the three mistakes that cost me £340 in refunds.
The prompt structure that passed for ceramic bakeware
The teapot sets and baking dishes were easiest. I needed images that looked like catalogue shots—neutral, trustworthy, no creative flourish. I used this prompt for a stoneware baking dish:
"Product photography of a rectangular stoneware baking dish, speckled cream glaze, sitting on white marble surface, soft natural window light from left, shallow depth of field, 45-degree angle, professional e-commerce style, high resolution"
Generated four variations. Picked two. Listed the dish at £34.99 including VAT, overlaid the price in Canva, uploaded to Shopify. Sold eleven between 26 December and 3 January. Zero returns. The images looked exactly like the physical product—same proportions, same glaze texture, same sense of weight.
The key was "professional e-commerce style" in the prompt. It told Stability AI I wanted boring and accurate, not artistic. I also specified the light direction and angle because I'd learned that vague prompts gave me dramatic shadows or odd perspectives that didn't match how the product actually looked on a kitchen counter.
For the teapot set I added "shown with matching cups, all items in frame, no hands, no props beyond a linen napkin". Sold eight sets at £42.50. One return, but the customer said the teapot was smaller than expected—that was a listing-copy problem, not an image problem.
The hoodie disaster and the fix
Clothing nearly sank me. I generated images for two new hoodie colours—sage green and charcoal grey—using this prompt:
"Lifestyle photo of a person wearing a sage green cotton hoodie, standing in a minimalist living room, hands in pockets, relaxed pose, natural light, clean background"
Stability AI gave me a handsome model in a hoodie that looked... fine. I cropped it, added the £38.99 price tag, listed it. Sold fourteen sage hoodies and nine charcoal. By 10 January I had six returns, all citing the same issue: the hoodie in the photo looked thicker and more structured than the actual garment. The AI had given the fabric a weight and drape that my mid-range cotton blend didn't have.
I'd made two mistakes. First, I didn't specify "lightweight cotton hoodie" or "relaxed fit". The model looked like he was wearing a premium fleece-lined piece. Second, I didn't compare the AI image to the physical sample before listing. I was rushing.
I pulled the AI images on 11 January, shot the hoodies myself with my phone against a grey wall, re-listed them. The refund rate dropped to zero. The lesson: clothing needs either a flat-lay shot or a very specific fabric-weight description in the prompt. "Cotton hoodie" isn't enough. Try "thin cotton jersey hoodie, loose fit, visible texture, casual drape" if that's what you're actually selling.
The air fryer that looked too good
The compact air fryer was my third mistake. I used this prompt:
"Product shot of a small white air fryer on a kitchen counter, basket pulled out slightly, modern kitchen background with blurred subway tiles, warm overhead light, clean and inviting"
The image was gorgeous. The air fryer looked like a £120 Philips model. I was selling a £54.99 budget version with a plastic handle and a smaller basket. I listed it anyway because I was behind schedule and the photo looked professional.
Sold six on Boxing Day. Four came back in the first week of January. One buyer left feedback: "Photo shows a much larger, sturdier unit. This is fine for the price but not what the picture suggested."
I'd violated the first rule of e-commerce: the image sets the expectation. If the AI makes your product look premium, you'll either attract the wrong buyers or disappoint the right ones. I should have added "compact budget air fryer, plastic handle, small 2-litre basket" to the prompt, or I should have dialled back the lighting and background to match the product's actual market position.
I re-shot it with my phone. The new image was duller but honest. Sold another eight in the January sale with zero returns.
What worked reliably: small goods with simple geometry
Kettles, utensil sets, ceramic mugs, enamel bread bins—anything with a defined shape and a single material—generated beautifully and sold cleanly. My best prompt for a stainless steel kettle:
"Studio product photo of a stainless steel electric kettle, brushed finish, black handle, sitting on white background, soft diffused light, no shadows, straight-on view, high detail"
Sold nine at £29.99. Zero returns. The image matched the product because the product was simple and the prompt was specific.
For a set of wooden spoons I used "flat lay of five wooden cooking spoons, light beech wood, arranged in a fan pattern on white linen, natural daylight, shallow focus on front spoon". Sold twelve sets. One return because a spoon had a knot the customer didn't like—nothing to do with the image.
The pattern: if the product has one material, one colour and a straightforward shape, Stability AI will give you a usable image in two or three tries. If it has multiple fabrics, complex textures or any kind of branding, you'll need to shoot it yourself or spend an hour refining prompts.
The VAT overlay workflow that saved time
Every price on my site includes VAT. I didn't want to generate 42 images, import them into Photoshop, add price tags and export them again. I used this workflow:
- Generate image in Kryotta (Stable Image Core, 1024×1024).
- Download as PNG.
- Open Canva, drop image into a Shopify product template.
- Add a white rounded rectangle in the top right corner, type "£34.99" in 18pt bold, export.
- Upload to Shopify.
Took about ninety seconds per image. The key was keeping the price overlay simple—white box, black text, no gradients or drop shadows. It looked like a sticker I could have added in real life.
For products sold in both the UK and Ireland (I ship to both), I created two versions: one with £, one with €. Shopify's market settings handled the rest.
The three mistakes that cost £340 in refunds
By 15 January I'd processed nine returns directly caused by AI image mismatches. The refunds totalled £340, and I ate the Royal Mail return postage on six of them.
Mistake one: generating clothing without specifying fabric weight. The sage hoodie looked like a £70 piece in the AI image. It was a £25 piece in reality. Five returns.
Mistake two: making a budget product look premium. The air fryer image suggested quality the product didn't have. Four returns.
Mistake three: not comparing the AI image to the physical sample before listing. I was in a hurry. I trusted the image because it looked professional. I should have held the sample next to my screen and asked, "Does this match?" On three products—the air fryer, the sage hoodie, the charcoal hoodie—the answer was no, and I listed them anyway.
The products that worked were the ones where I either shot them myself or wrote prompts so specific that Stability AI couldn't invent details. "Stoneware baking dish" plus six descriptive constraints gave me an accurate image. "Cotton hoodie" with no constraints gave me a fantasy.
Questions people ask
Can I use AI product images for everything I sell?
No. Simple goods with single materials—ceramics, metal, glass, basic wood—work well. Clothing, anything with branding, or products where texture matters (like fabric weight or surface finish) need real photos or extremely detailed prompts. If you're not sure, generate the image, compare it to the physical product, and ask whether a customer would feel misled.
How do I avoid returns when using AI images?
Write prompts that describe the product's actual quality level. If you're selling a budget item, don't add "premium", "luxury" or "high-end" to the prompt. If your hoodie is thin cotton, say so. If your air fryer is compact plastic, specify that. The AI will follow your lead—if you're vague, it'll invent details that might not match reality.
Which Stability AI model should I use for product shots?
I used Stable Image Core for all 42 images. It's fast, it handles product photography prompts well, and it doesn't add artistic flourishes unless you ask for them. SD 3.5 Large is overkill for this kind of work—you're not making hero banners, you're making catalogue shots.
Do I need to disclose that I used AI for product images?
There's no legal requirement in the UK, but if the image misrepresents the product, you're liable under consumer protection law whether you made it with AI, Photoshop or a camera. The image has to be accurate. How you made it doesn't matter to Trading Standards; what it shows does.
I'll keep using Stability AI for simple products—it saved me two days of shooting over Christmas and let me list new stock in time for Boxing Day. But I won't use it for clothing again until I've figured out how to describe fabric accurately enough that the AI doesn't upgrade it. If you're testing this yourself, start with your simplest products, compare every image to the physical item before you list it, and track your January returns. You'll know within two weeks which prompts you can trust.
Try it in Kryotta—Stable Image Core is fast, and you'll see within three or four generations whether your prompts are honest enough to sell from.
