I spent $47.60 on 40 product images and learned which prompts customers trust
I run a small Shopify store selling ceramic planters and vases. I make them, I photograph them, I ship them. Last September I had 40 new pieces ready for the holiday push and exactly zero time to shoot them properly. My usual setup—DSLR, white backdrop, two softboxes—takes about 20 minutes per piece once you account for positioning, test shots, and cleanup in Lightroom. That's 13 hours I didn't have before Black Friday.
I'd been hearing about Stability AI for product shots. I tried it. Some images converted beautifully. Some got me refund requests. This is what I learned about which prompts work and which ones waste money.
The baseline: what I was replacing and what it costs
My old workflow cost me time, not cash. Equipment was already paid for. But 13 hours of my labour at even $25/hour is $325 in opportunity cost, plus I'd have to delay the launch or skip sleep. Stability AI's Stable Image Core model costs about $0.04 per image at standard resolution through Kryotta. At that rate, 40 images cost $1.60. I generated about 120 images total—three variations per product, kept the best—so actual spend was $4.80. Add another hour of my time reviewing and tweaking prompts at $25, and I'm at $29.80 all-in. I'll explain where the rest of the $47.60 went in a moment.
The real question wasn't cost. It was whether customers would buy from AI-generated images or bounce.
Prompt one: the control shot that worked immediately
I started with the simplest possible prompt for a terracotta planter with a matte glaze:
"Terracotta ceramic planter with matte white glaze, centered on pure white background, soft even lighting, studio product photography, no shadows, front view, commercial quality"
Stable Image Core gave me a clean, trustworthy shot in four seconds. The planter looked real. The glaze texture was visible. The background was genuinely white, not grey or cream. I uploaded it to Shopify, launched the product, and it sold within two days at $42. No questions, no refund. That image is still live.
What made it work: "pure white background" and "no shadows" are specific enough that the model doesn't improvise. "Soft even lighting" avoids the harsh spotlight look that screams AI. "Front view" and "centered" mean the product fills the frame without weird cropping. I used this exact structure for 14 other pieces. Conversion rate on those listings was 3.8 per cent over the next month, which is slightly above my store average of 3.4 per cent.
Prompt two: the artistic disaster that got me a refund request
I got confident. I tried to make the images more interesting. For a tall blue vase, I wrote:
"Elegant blue ceramic vase on rustic wooden table, morning sunlight streaming through window, cozy home interior, depth of field, lifestyle photography"
The image looked gorgeous. Soft focus background, warm light, very Pinterest. I uploaded it, customer bought the vase for $68, vase arrived, customer requested a refund. The complaint: "The blue is much darker than the photo and there's no texture visible."
The model had smoothed out the glaze completely and shifted the colour toward a bright sky blue to match "morning sunlight." The real vase was cobalt with a slightly rough surface. The lifestyle setting made it impossible to see the actual product clearly. I refunded $68, kept the vase, re-shot it with the control prompt, relisted it. It sold again three weeks later with no complaint. That one mistake cost me $68 plus shipping both ways, another $18. That's where most of the $47.60 overage came from.
Lesson: lifestyle shots work for marketing emails or Instagram, but they don't work for the main Shopify listing image. Customers need to see the product as it will arrive, not as it might look in a fantasy kitchen.
Prompt three: adding one detail that doubled my click-through rate
I noticed my terracotta planters were getting clicks but not sales. The images were accurate, but they looked flat. I added one phrase to the control prompt:
"Terracotta ceramic planter with matte white glaze, centered on pure white background, soft even lighting from above and left, studio product photography, no shadows, slight front-left angle, commercial quality"
"Slight front-left angle" and "lighting from above and left" gave the images subtle dimension. The curve of the rim became visible. The glaze caught light in a way that showed texture. Click-through rate from search to product page went from 2.1 per cent to 4.3 per cent. Sales followed. I applied this tweak to every remaining product.
The difference between "soft even lighting" and "soft even lighting from above and left" is that the second version gives the model a clear instruction instead of leaving it to guess. Guessing produces flat, boring images.
What didn't work: trying to fake scale or context
I tried adding a hand holding a small vase to show scale. Prompt was:
"Small ceramic vase held in hand, white background, studio lighting, product photography"
The model gave me a vase, a hand, and five and a half fingers. The wrist bent at an angle no human wrist bends. I tried again with "realistic hand proportions"—got four fingers. I tried "professional hand model"—got a weirdly smooth hand that looked like a mannequin. I gave up and shot a separate scale reference image myself with my phone. Took two minutes.
Hands, people, and anything that requires anatomical accuracy are still a problem. If you need a model or a hand in the shot, hire someone or do it yourself. Don't waste tokens trying to fix AI hands.
The checklist: what to include in every product prompt
After 40 products and 120 generated images, here's what consistently worked:
- Pure white background or "solid white seamless background"—anything less specific gets grey or off-white.
- Lighting direction—"from above and left" or "from front and slightly above." Avoids flat lighting.
- Angle—"front view" or "slight front-left angle." Never "side view" unless you actually want side-only.
- "No shadows" if you want clean ecommerce shots. Shadows can look dramatic but they complicate background removal later.
- "Studio product photography" and "commercial quality"—signals you want a professional result, not an artistic interpretation.
- Specific product details—"matte glaze," "terracotta," "cobalt blue," "hand-thrown texture." The more specific, the more accurate.
What to leave out: lifestyle context, people, hands, props, "artistic," "moody," "dramatic," "rustic table," "window light." All of those push the model toward invention instead of documentation.
The one image I kept that I thought would fail
I had a crackle-glaze vase—white glaze with intentional fine cracks. I was sure the model would either ignore the cracks or turn them into damage. I wrote:
"White ceramic vase with crackle glaze finish, fine intentional cracks in surface, centered on pure white background, soft even lighting from above, front view, studio product photography, commercial quality"
It worked. The cracks showed clearly. The glaze looked deliberate, not broken. Customer bought it, left a five-star review mentioning the "beautiful crackle detail." I'd underestimated how well the model handles texture when you name it specifically.
Cost per image vs cost per mistake
Final tally: 120 images generated at $0.04 each = $4.80. One refund from the lifestyle-shot mistake = $68. Shipping both ways = $18. One hour of prompt tweaking = $25. Total real cost: $115.80 for 40 products, or $2.90 per product. If I'd stuck to the control prompt from the start and skipped the lifestyle experiment, cost would've been $29.80 total, or $0.75 per product.
The expensive part isn't the image generation. It's the refund when you get it wrong. One bad image cost me more than 1,700 good images would have.
What I'd do differently for the next 40 products
I'd generate two images per product from the start: one pure front view, one slight front-left angle. Upload both to Shopify and let the customer toggle between them. Total cost per product: $0.08 in generation, maybe five minutes of my time. I'd skip all lifestyle shots for the main listing and save those for email campaigns or Instagram, where a return policy doesn't apply.
I'd also test Stable Diffusion 3.5 Large for products where texture really matters—hand-thrown ceramics, anything with visible brushstrokes or glaze variation. It's slower and costs more ($0.10 per image), but the detail rendering is noticeably better. For simple shapes and solid colours, Stable Image Core is fine.
Questions people ask
Can I use AI product images on Shopify without disclosing they're AI-generated?
Yes. There's no legal requirement in the US to label AI-generated product photos, and Shopify doesn't require it either. The image just needs to accurately represent what the customer will receive. If it does, you're fine. If it doesn't, you'll get refunds regardless of how it was made.
Do AI product images hurt conversion compared to real photos?
Not if they're accurate. My control-prompt images converted at 3.8 per cent vs my usual 3.4 per cent. The difference is small enough that it could be noise, but they definitely didn't hurt. The lifestyle-shot image hurt badly, but that's because it misrepresented the product, not because it was AI.
How do I handle products with complex details like embroidery or engraving?
Photograph those yourself. AI models struggle with fine text, intricate patterns, and anything that requires exact replication. If the detail is a selling point, don't trust the model to get it right. Shoot it with your phone if you have to—accurate beats pretty.
What's the fastest way to generate product images in bulk?
Write one good prompt, test it on three products, refine it until it works consistently, then batch the rest. In Kryotta you can queue up multiple image generations with slight variations and let them run while you do other work. Don't try to write a custom prompt for every product—you'll waste time and introduce inconsistency.
I've kept using AI for new product shots. The time savings are real and the results are good enough that customers don't ask questions. I still shoot anything with text, complex texture, or where colour accuracy is critical. For everything else, a solid prompt and Stable Image Core get me 90 per cent of the way there in four seconds. If you're staring at 40 unshot products and a Black Friday deadline, that's good enough. Try it at Kryotta—you'll know in the first five images whether it'll work for your store.
