I made 45 product shots with Stability AI then hired a photographer: which got more Shopify sales

I generated 45 AI product shots with Stability and compared them head-to-head against a professional photographer on my Shopify store. Here's what actually sold more—and why it wasn't obvious.

KKryotta TeamProduct & research · · 9 min read
Flat-lay product photography showing a folded linen shirt, concrete candle with eucalyptus, and lip gloss on marble with natural side lighting
Flat-lay product photography showing a folded linen shirt, concrete candle with eucalyptus, and lip gloss on marble with natural side lighting

I made 45 shots of a linen shirt, a candle, and a lip gloss

I run a small Shopify store selling three things: linen apparel (shirts, dresses, trousers), hand-poured candles in concrete vessels, and a line of five lip glosses. Revenue last year was $87,000. I handle photography myself because hiring a pro every time I launch a new colour or restock means $200–400 per session, and I launch something new every six weeks.

In March, I tested Stability AI (Stable Image Core and SD 3.5 Large, both available in Kryotta) against a local photographer I've used twice before. Same fifteen products: five apparel pieces, five candles, five cosmetics. Three shots per product. The AI batch cost me $0 in generation fees because I was already paying for workspace access. The photographer charged $350 for the session plus $120 for edits, so $470 total.

I ran both sets on the same product pages for two weeks each, tracked add-to-cart rates and conversion, then switched them. I wanted to know if AI product photography could actually move sales or if it just looked "good enough" while secretly costing me money in lost conversions.

The results weren't clean. AI won for candles, lost badly for cosmetics, and tied for apparel—but with a caveat that matters if you're shooting clothing.

What I asked Stability AI to generate and what it gave me

I used Stable Image Core for the first round because it's faster and I wanted to test whether a general-purpose model could handle product work. Prompt for the linen shirt: "Linen button-down shirt in sage green, folded on white marble surface, natural window light from the left, soft shadows, clean minimal background, product photography style, sharp focus on fabric texture."

It gave me a decent shot. The fabric looked like linen—you could see the weave—and the colour was close. The fold was a bit too perfect, almost geometric, but for a thumbnail it worked. I generated three variations by tweaking the lighting direction and fold style. Total time: four minutes.

For the candles, I asked for "Hand-poured candle in round concrete vessel, wick visible, placed on wooden table with eucalyptus branch beside it, warm afternoon light, shallow depth of field, product photography." This was where Stability surprised me. The concrete texture looked real, the wick was centered, and the eucalyptus didn't look like clip art. I used SD 3.5 Large for one variation to see if the upgrade mattered—it did. The depth of field was smoother, and the wood grain on the table had actual variation instead of that repeating AI pattern.

Cosmetics were a disaster. Prompt: "Lip gloss tube in rose gold, standing upright on pink marble, soft diffused light, luxury beauty product photography, sharp label detail." The tube shape was fine, but the label text was gibberish. I tried six variations, adding "no text" and "blank label," but Stability kept inventing letters. I ended up with clean shots of unlabeled tubes, which is useless when your branding is 60% of the product.

The photographer shot everything in two hours and I still had to fix three images

The photographer came to my garage studio (a corner with a white backdrop and two softboxes I bought on Amazon for $140). She shot all fifteen products in two hours, gave me RAWs the next day, and delivered edited JPEGs three days later.

The linen shirt shots were better than AI. She draped it over a chair, styled it with a ceramic mug (my brand sells "slow morning" vibes), and the fabric had natural wrinkles that made it look real. The AI fold was too stiff by comparison.

Candles were nearly identical to the AI shots. She used the same wooden table and eucalyptus branch, same lighting setup. When I put them side by side, I couldn't tell which was which until I zoomed in—the photographer's version had a tiny reflection of the window in the concrete, and the AI version didn't. Neither my customers nor I cared.

Cosmetics: the photographer nailed it. The label was sharp, the rose gold caught the light, and she shot one version with the cap off and applicator wand visible, which I hadn't thought to ask AI for (and probably couldn't have, because AI struggles with small product mechanisms).

But I still had to ask for re-edits on three images. One candle shot had a dust spot on the concrete. One dress had a wrinkle near the hem that looked like a stain. One lip gloss had a shadow that made the tube look dented. She fixed them, but it added two days.

Add-to-cart rates told me which images people trusted

I ran AI images for two weeks, then swapped in the photographer's shots for two weeks. Same products, same prices, same traffic sources (Instagram, Google Shopping, a small email list of 1,200 people).

Linen apparel: AI 3.2% add-to-cart, photographer 3.8%. The photographer won, but not by much. I think the styled shot (shirt on chair with mug) gave people a reason to imagine using it, while the AI fold was just a product on a surface. If I'd prompted AI to include styling props, it might've closed the gap, but I didn't think of that until after.

Candles: AI 4.1%, photographer 4.0%. Essentially tied. I was shocked. The AI shots converted just as well as the pro ones, and no one mentioned image quality in customer emails or Instagram DMs. Candles are simple: vessel, wick, background. AI can do simple.

Cosmetics: AI 1.8%, photographer 5.3%. Brutal. The unlabeled AI tubes looked like mockups, not real products. People didn't trust them. The photographer's shots—especially the one with the applicator wand—converted nearly three times better. One customer replied to my email saying she almost didn't buy because the first image "looked like a render."

Conversion rates (add-to-cart to purchase) stayed consistent across both sets, so the difference was entirely in whether people clicked "add to cart" in the first place.

What failed and what I'd do differently

Hands. I tried one prompt with a hand holding the lip gloss tube. Stability gave me six fingers and a thumb that bent backward. I deleted it immediately. If your product needs to be shown in use—especially cosmetics, tools, or anything worn—don't use AI unless you're willing to spend an hour fixing it in Photoshop.

Text and logos. AI cannot reliably generate readable text. If your product has a label, logo, or packaging with words, you'll need to either shoot it yourself, use Photoshop to overlay the real label, or hire a photographer. I tried adding "accurate label text" to the prompt. It made the gibberish more confident.

Lighting consistency. When I generated three variations of the same candle, the light direction shifted slightly between shots even though I used the same prompt. It wasn't dramatic, but if you're building a cohesive product page with multiple angles, you'll notice. The photographer kept lighting identical across all fifteen products because she didn't move the softboxes.

Backgrounds. AI loves adding random objects. I asked for a "clean minimal background" and got a marble surface with a half-visible book spine in one corner. I cropped it out, but it added time. The photographer's backdrop was actually clean.

When I'd use AI vs when I'd pay the photographer

I'll use AI product photography for candles, simple home goods (coasters, planters, anything with one solid shape and no moving parts), and initial mockups when I'm testing a new product idea and don't want to spend $400 before I know it'll sell. AI is fast, and for simple products with no text, it's good enough that most customers won't notice or care.

I'll pay the photographer for anything worn, anything with visible branding, anything that benefits from styling (apparel, gift sets, products sold as part of a "lifestyle"), and cosmetics. The $470 session cost me $31 per product. If those products convert 2–3% better, the session pays for itself in about sixty sales. I hit that in ten days for the lip gloss alone.

If you're shooting ten products or fewer and you have decent natural light, a smartphone and a $30 white poster board will beat AI. If you're shooting fifty products and half of them are simple shapes, AI can handle the boring ones while you spend the photographer's time on the hero shots.

The hybrid workflow I'm using now

I generate AI shots for new candle colours and home goods, then hire the photographer once a quarter for apparel and cosmetics. I use Stable Image Core for quick tests and SD 3.5 Large when I need a final image for a product page. The quality gap is noticeable—Large handles shadows and textures better—but Core is fast enough for Instagram posts and email thumbnails.

I also started using AI for secondary images: the "lifestyle" shot of a candle on a bedside table, or a shirt styled with jeans and boots. I generate those, check that they look plausible, and use them as the third or fourth image in the carousel. The hero shot is still from the photographer, but AI fills out the page without adding cost.

One workflow that works: shoot the hero image with the photographer, then use AI to generate variations (different backgrounds, different lighting, different props). You get the quality and trust of a real photo up front, and AI gives you options for A/B testing or seasonal campaigns without booking another session.

Questions people ask

Can AI generate product photos with my actual logo on them?
Not reliably. You can generate a clean shot of the product shape, then use Photoshop or Canva to overlay your real logo as a separate layer. It adds five minutes per image, but it's faster than re-shooting.

Which Stability model should I use for product photography?
SD 3.5 Large if the image is going on your main product page or in a paid ad. Stable Image Core if you need something fast for Instagram Stories or email thumbnails. The difference is sharpness and how well it handles shadows and reflections.

Do AI product images hurt SEO or conversion on Shopify?
Google doesn't penalise AI images, but customers penalise images that look fake. If the AI shot is clean, well-lit, and shows the product clearly, it converts fine. If it has six-fingered hands or gibberish text, people leave. Test both and watch your add-to-cart rate.

How much does a local product photographer cost in the US?
$200–500 for a two-hour session, depending on your city and how many products you're shooting. Editing is usually included, but ask up front. If you're shooting more than twenty products, negotiate a per-product rate instead of hourly.

I'm still using both. AI hasn't replaced my photographer, but it's cut my annual photography budget by about 40% because I'm not paying for reshoot sessions every time I launch a new candle scent. If you're testing AI image generation for selling on Shopify, start with your simplest products and compare the add-to-cart rate. That number will tell you faster than any article whether it's working.

You can try Stable Image Core and SD 3.5 Large in Kryotta alongside other models if you want to test a few options without switching tools. I keep it open in a tab next to Shopify and generate images between packing orders.

K
Written by
Kryotta Team
Product & research

Kryotta is the multi-model AI workspace — every leading model, one login, one bill. Try it free →

Related reading