I generated fifty-five shots before I realised the shadows were wrong
I sell homeware and small gifts through Shopify. Candles, ceramic mugs, linen tea towels, that sort of thing. I've been photographing everything myself with a lightbox and my phone, and the results are fine but slow. Each product takes twenty minutes to shoot, edit and upload, and I've got forty new items arriving from my supplier in Stoke next week that all need to be live before the January sales.
I decided to test AI-generated product shots with Stability AI in Kryotta, then turn the best ones into short Reels with Luma Ray 2 for Instagram and Facebook ads. I'd generate three angles per product, pick the ones that looked real enough to pass a customer's scrutiny on a phone screen, and make ten-second Reels showing the product rotating or sitting in a lifestyle scene. The goal was to get everything online in three days instead of three weeks, spend less than £100, and avoid the fake-looking AI sheen that makes people scroll past.
I made fifty-five product shots. Thirty-two went straight in the bin. Fourteen needed Photoshop fixes. Nine went live. Then I made six Reels, tested them on WhatsApp with five regular customers, and learned which prompts HMRC's advertising standards would flag and which ones just looked odd to British buyers who've seen a lot of product photography.
Here's the checklist I built. It's what I check now before I generate anything, and it's saved me about four hours of wasted renders and £40 in ad spend on visuals that didn't convert.
Before you generate: check the product actually suits AI rendering
Not every product works. I learned this after generating twelve shots of a textured linen tea towel that came out looking like a flat piece of printer paper with a noise filter. Stability AI is excellent at hard surfaces, reflections, glass, ceramics and anything with a defined edge. It struggles with fabric texture, anything handmade with visible irregularities, and products where the material itself is the selling point.
My ceramic mugs rendered beautifully. The linen towels looked wrong in every attempt, no matter how detailed the prompt. I photographed those myself and saved two days of frustration.
Check your product category before you commit. If you sell clothing, jewellery on a model, or anything where texture and drape matter, AI shots won't replace a real photo yet. If you sell candles, glassware, packaged goods, stationery or homeware with clean lines, you'll get usable results in one or two attempts.
The shadow direction test: send it to yourself on WhatsApp first
This is the mistake that killed thirty-two of my fifty-five shots. The shadows didn't match the light source, or they were too soft, or they were missing entirely. A candle sitting on a white surface with no shadow at all looks like a PNG someone dropped into a scene. A mug with a shadow pointing left while the highlight on the rim suggests light from the right makes your brain itch even if you can't articulate why.
I now generate every shot with a specific lighting instruction in the prompt: "single soft light source from top left, natural shadow cast to bottom right, white studio background". Then I send the image to myself on WhatsApp and look at it on my phone as if I were a customer scrolling Instagram. If the shadow looks wrong or absent, I don't use it. I don't try to fix it in Photoshop unless it's a tiny adjustment, because customers can tell.
The nine shots that went live all had consistent, believable shadows. The ones that converted best in ads had shadows that were visible but not harsh—about 40% opacity, soft edge, matching the product's base shape.
The "would I believe this in an eBay listing" filter
I sell on eBay as well as Shopify, and eBay customers are ruthless. They've seen every dropshipping trick, every stock photo passed off as a real item, every angle that hides a defect. If an image looks too perfect or too generic, they assume it's not the actual product and they move on.
I tested this by uploading three AI-generated candle shots to an eBay draft listing and asking my partner—who buys a lot on eBay—which ones she'd trust. She picked the one with a slight reflection on the surface and a background that wasn't pure white. The other two looked "like they came from AliExpress".
The fix: add a small imperfection or detail that signals "this is a real photo of the thing I'm selling". A faint dust mote in the background. A barely visible texture on the surface. A reflection that shows the studio setup. Stable Image Core can do this if you prompt for it: "small imperfections, slight surface texture, realistic studio photo". Don't overdo it, but don't make it look like a render from a 3D model library either.
Prompts that fail review or look fake to UK customers
I ran six Facebook ad campaigns with AI-generated product shots in November. Two were rejected in review, and one was approved but got a 0.4% click-through rate because the image looked uncanny.
The rejections were both for the same reason: the background included a blurred lifestyle scene that Facebook's reviewer thought was misleading. One showed a candle on a table with a blurred bookshelf behind it, and the reviewer flagged it as "digitally altered in a way that misrepresents the product". The other had a mug on a kitchen counter with a blurred window, same problem. I appealed both and lost.
The low-performing ad used a shot with perfect symmetry—a candle dead centre, perfectly round, no variation in the wax surface. It looked computer-generated because it was, and people scrolled past.
What works for UK ad review and UK customers: plain white or light grey background, product slightly off-centre, one clear shadow, no blurred lifestyle elements unless you're very confident they look real. If you want a lifestyle shot, generate the product on a solid surface with nothing else in frame, then composite it into a real photo yourself. Don't let the AI invent the scene.
The Luma Reel test: does it move like a real camera or a video game
I made six Reels with Luma Ray 2. Three were simple rotations—camera circling the product on a white surface. Three were "lifestyle" shots where the camera pushed in on a candle sitting on a wooden table. I tested all six by sending them to five customers on WhatsApp and asking which ones they'd stop scrolling for.
The rotations worked. The push-in shots looked like video game renders. The movement was too smooth, the focus shift was too perfect, and the lighting didn't change as the angle changed. One customer said "it looks like a Sims object".
Luma Ray 2 is very good at slow, simple camera moves. It's not good yet at complex moves or anything that requires the lighting to shift naturally. Stick to gentle rotations, slow zooms, or a static shot with the product itself moving slightly (a candle flame, steam rising, that sort of thing). Don't try to replicate a handheld product video or a dolly shot unless you're prepared to regenerate it five times.
I kept three Reels. They've been running in Instagram ads for six weeks and they're converting at 1.8%, which is slightly better than my best static shots. The cost to generate all six was about £8 in Kryotta credits. The three I kept cost me £4 total.
The January sales timing question: generate now, test before you spend
I'm writing this in early January. If you're reading it now, you've probably got sale stock you need to shift and you're wondering whether AI visuals are fast enough to be useful before the sale ends.
They are, but only if you don't waste time on products that won't render well. My checklist: pick your five best-selling hard-surface products, generate three angles each with Stable Image Core, test them on WhatsApp with three customers, pick the two best, make a simple Reel with Luma Ray 2, and get them into an ad by tomorrow. Budget £15 for generation, £30 for ad spend to test. If the click-through rate is above 1%, scale it. If it's below 0.8%, go back to your phone photos.
Don't generate fifty-five shots like I did. Generate fifteen, test five, use two. Speed matters more than volume when you've got a narrow sales window.
The VAT question no one mentions: are AI-generated images stock or created assets
This is boring but it cost me an hour on the phone with my accountant. If you're VAT-registered and you're paying for AI image generation, is that a stock purchase or a service? Does it go under software subscriptions or marketing materials?
My accountant's answer: it's a service, it goes under software or digital tools, and you can claim the VAT back on your next return if you're using standard VAT accounting. If you're on the Flat Rate Scheme, it's just an expense and you don't claim anything back. If you're not VAT-registered, ignore this entirely.
I'm mentioning it because I didn't think about it until I was reconciling November's expenses and I had £47 in Kryotta credits sitting in a spreadsheet with no category. It's not a big deal, but it's the kind of thing that creates a small mess if you don't label it correctly from the start.
Questions people ask
Can I use AI product shots on eBay without declaring they're AI-generated?
Yes. eBay's policy requires accurate representation of the item, not disclosure of how the photo was made. If the AI shot accurately shows what the customer will receive, you're fine. If it misrepresents colour, size or condition, you'll get returns and disputes regardless of whether it's AI or a phone photo.
Will Facebook reject my ad if I use an AI-generated Reel?
Not automatically, but they will if the Reel includes a blurred or fabricated background that looks like you're trying to imply a lifestyle context that isn't real. Stick to plain backgrounds or composite the product into a real photo yourself.
How much does it cost to generate fifty product shots and six Reels in Kryotta?
I spent about £35 in total. Stable Image Core costs roughly 60p per image, Luma Ray 2 costs about £1.20 per ten-second Reel. Your cost will vary depending on how many attempts you need, but budget £50 to be safe.
Do AI product shots convert worse than real photos?
In my six-week test, the AI shots that passed the WhatsApp test converted at the same rate as my best phone photos—around 1.8% click-through on Instagram ads. The ones that looked fake converted at 0.4% or got rejected in review. The quality gate is whether a customer on a phone screen believes it's real, not whether it came from a camera or a model.
I'm still using my lightbox for linen and anything textured. For ceramics, candles and glassware, I'm generating the shots in Kryotta and testing them on WhatsApp before they go live. It's faster, it's cheaper than a photographer, and it works as long as you check the shadows first.
If you want to try this for your own shop, the checklist is the same regardless of what you sell: check the product suits AI rendering, test the shadow direction on your phone, apply the eBay believability filter, avoid complex camera moves in Reels, and don't spend ad budget until you've shown the visual to a real customer. Start at app.kryotta.ai/auth and generate five shots tonight. You'll know in twenty minutes whether it'll work for your products.
