The 40-product test: what I tracked and why
I run product photography for three Shopify stores—one selling handmade candles, one selling vintage-style jewellery, and one selling organic cotton baby clothes. In September I generated 40 product shots using Stable Image Core and SD 3.5 Large, then shot the same 40 products myself with a Canon EOS R10 and a $180 softbox kit from B&H Photo. I split-tested them on the stores for 30 days, tracking click-through rate from collection pages, add-to-cart rate on the product page, and return rate after purchase.
The candles got 2,400 collection-page views, the jewellery 3,100, the baby clothes 1,800. I rotated AI and real photos every three days so time-of-week bias wouldn't skew the results. I used Shopify's built-in analytics and a $29/month app called Conversific to track which image was live when someone added to cart.
The hypothesis: AI would match or beat real photos for simple products on white backgrounds, but tank on anything that needed texture or fabric drape. I was half right.
Prompt formulas that worked (and three that looked good but killed conversion)
For the candles, this prompt in Stable Image Core gave me usable shots 7 times out of 10:
Product photography of [candle scent] soy candle in amber glass jar on pure white background, shot from 45-degree angle, soft diffused lighting from top left, shallow depth of field, commercial studio quality, sharp focus on label
The failures: any time I added "luxury" or "premium" to the prompt, the model gave me overlit, too-perfect shots that looked like 3D renders. Customers scrolled past them. When I tried "natural lighting" or "window light," I got muddy shadows that made the wax look discoloured. The 45-degree angle mattered—straight-on shots had no depth, and overhead shots hid the label.
For jewellery, I needed two goes at every piece. First prompt:
Flat lay product photo of sterling silver [item type] on white marble surface, soft shadows, even lighting, high resolution, editorial style
Then I'd take that image into Kryotta's editor, mask out the background, and regenerate just the shadows with:
Subtle shadow under jewellery piece on white background, natural and soft, no harsh edges
The ones that failed: anything with "sparkling" or "glittering" in the prompt. The model added fake light flares that looked cheap. Rings photographed at an angle showed warped bands. I had to shoot rings straight down, which meant the AI images had no sense of scale—customers couldn't tell if they were looking at a cocktail ring or a midi ring.
Baby clothes broke every prompt I tried. "Cotton onesie on white background" gave me flat, lifeless fabric. "Folded baby clothes styled on wooden surface" looked like stock photos from 2015. "Baby clothes laid flat with natural wrinkles" gave me creases in the wrong places, and one image had a phantom sleeve that didn't match the actual product. I scrapped all the clothing AI shots by day eight.
What tanked conversion: lighting, backgrounds, and the uncanny valley of product detail
The candles performed nearly identically—AI images converted at 4.1%, real photos at 4.3%. The difference wasn't statistically significant over 2,400 views. But the jewellery told a different story. AI shots had a 2.8% add-to-cart rate; real photos hit 5.1%. The return rate flipped it further: 18% of jewellery orders from AI images came back with "doesn't look like the photo" notes. Real photos had a 9% return rate, mostly sizing issues.
I looked at the images side by side. The AI versions had perfect, even lighting with no hotspots, which made the silver look flat. My real photos—shot with one softbox and a white reflector card—had a subtle gradient across the metal that gave it dimension. The AI also smoothed out tiny surface textures that customers apparently used as proof the piece was handmade. One reviewer wrote, "I thought this was going to be mass-produced based on the photo, but it's actually really nice." That's a compliment and a warning.
The baby clothes weren't close. Real photos converted at 6.7%, AI at 1.9%. I pulled the AI images after 12 days because the return rate hit 31%. Parents wanted to see how the fabric draped, where the seams sat, how the snap closures looked when fastened. AI couldn't fake that. I tried photographing the clothes on a mannequin and then using img2img in SD 3.5 Large to clean up the background, but the model changed the colour temperature and made the organic cotton look synthetic.
The surprise winner: where AI beat my camera by 40%
Candles on styled backgrounds. Halfway through the test I tried a different approach—I prompted Stable Image Core to generate lifestyle shots instead of clean product photos:
Soy candle in amber jar on wooden tray with eucalyptus stems and linen napkin, morning light from window, cosy minimal styling, magazine editorial style, shallow depth of field
These converted at 5.9% versus 4.2% for my real styled shots. I think it's because the AI gave me consistently good styling—my real photos had a coffee ring on the tray in one shot, and the eucalyptus looked wilted in another. The AI version was aspirational without being unattainable, and it didn't promise anything about the candle itself that wasn't true. The product was still the product; the context was just prettier.
I ran the same test with jewellery on a model's hand. AI failed completely—fingers looked waxy, nails had inconsistent polish, and one image gave the model six fingers. Real photos with my partner's hand (and $8 press-on nails from Target) converted 30% better than the solo jewellery shots.
What I'd do differently before Black Friday, and what I'm keeping
I'm keeping AI for candle lifestyle shots and any product that's symmetrical, one solid colour, and doesn't need texture detail. I'll shoot real photos for anything worn, anything with fabric, and any hero image that loads first on the product page. The workflow that's working now: I shoot one real photo of each product on white background, then use that as an img2img reference in Stable Image Core to generate three styled variations. I get the accuracy of a real product with the perfect styling of AI.
For Black Friday I'm generating 15 email header images with sale products arranged in a grid—something that would take me two hours to shoot and style, but takes 12 minutes in Kryotta with Stable Image Core. I'll use real product photos in the email body where people are deciding whether to click through.
The ROI: my real photo setup cost $490 (camera was a gift, but the lens, softbox, backdrop, and props added up). The AI images cost me $40 in Kryotta credits over the month. If I only sold candles, I'd go 80% AI. Because I sell jewellery and clothing, I'm landing at 40% AI, 60% real. The time saved on candle shots pays for the extra hour I spend on jewellery.
One more thing I wasn't expecting: AI images load faster. The file sizes from Stable Image Core averaged 320 KB versus 1.8 MB for my Canon shots, even after I compressed them. On mobile, where 60% of my traffic comes from, that half-second difference in image load time added up. My bounce rate on AI-image product pages was 2% lower. I don't know if that's the speed or the images themselves, but I'll take it.
Questions people ask
Can I use AI product photos if I'm selling on Etsy or Amazon?
Etsy's terms say your photos must accurately represent the item you're selling. AI images are fine as long as they don't mislead—so lifestyle shots are safer than trying to fake the product itself. Amazon requires at least one photo on a pure white background, which AI does well, but read the category rules; some require real photos of the actual item, especially in beauty and grocery.
Which Stability AI model is better for product photography, Stable Image Core or SD 3.5 Large?
I got more consistent results with Stable Image Core for clean product shots on white. SD 3.5 Large gave me more creative styled images, but it also invented details that weren't on the real product more often. Start with Core; use Large if you need a specific mood or setting.
Do I need to disclose that I used AI for product images?
There's no legal requirement in the US as of now, but if the AI image misrepresents the product and someone returns it, you're still liable. I'd avoid AI for anything where colour accuracy, texture, or exact dimensions matter to the buying decision.
How do I keep AI from changing my product's colours?
Shoot one reference photo in good light, upload it to Kryotta, and use img2img mode with a high influence setting (0.7 or higher). Your prompt should describe the styling and background, not the product itself—let the reference image control the product details.
I've been running these tests in Kryotta because I can compare Stable Image Core and SD 3.5 Large side by side without switching tools, and the image editor lets me fix small details without starting over. If you're prepping product images for Black Friday and want to try the AI-plus-real hybrid approach, you can start at app.kryotta.ai/auth and use the same prompts I've shared here.
