I made 25 Jumia product photos with Stability AI — which prompts got clicks and which got returns

I listed 25 phone cases on Jumia with AI images—here's what sold and what came back

I sell phone accessories on Jumia. Cases, screen protectors, charging cables, power banks. Most of my stock comes from suppliers in Industrial Area and Luthuli Avenue. I photograph products myself with my phone, usually on my desk with whatever light is coming through the window. The photos are fine, but they're not great. White balance is off half the time, shadows fall in weird places, and the background is always my desk or a wrinkled bedsheet I use as a backdrop.

Last month I decided to test Stability AI to generate product images. Not to replace photography entirely, but to see if I could make my listings look more professional without hiring a photographer or buying a lightbox. I picked 25 phone cases—mix of plain silicone, patterned TPU, and leather wallet cases—and generated new listing photos using Stable Image Core on Kryotta. I kept my original photos as backups and tracked what happened over four weeks: which listings got more clicks, which ones converted to sales, and which ones triggered returns or complaints.

Thirteen of the 25 worked well. They got more views than my old photos, converted at roughly the same rate, and no one complained. Twelve were disasters. Four came back as returns with customers saying the product "looked nothing like the photo." Three got angry WhatsApp messages before the customer even confirmed delivery. Five just sat there with high bounce rates and zero sales.

Here's what I learned about which prompts work for actual e-commerce in Kenya and which ones get you returns.

The twelve images that failed all had the same three problems

I started by trying to make my products look premium. I wrote prompts like "luxury phone case on marble surface, studio lighting, high-end product photography, bokeh background." The images looked beautiful. Soft focus, elegant shadows, the case sitting on a marble countertop that could've been in a Westlands apartment. I uploaded them to Jumia, feeling quite pleased with myself.

The first return came back within three days. Customer said the case was "cheap plastic, not the leather shown in the photo." I looked at the listing again. The AI had rendered the silicone case with a texture that looked like stitched leather. The colour was close—navy blue—but the material looked completely different. The marble surface and the bokeh made it worse, because the whole image screamed expensive and the actual product was a KSh 800 silicone case.

I went back through the twelve failed images and found the same problems in all of them:

Problem one: the AI changed the material. Silicone cases came out looking like leather or brushed metal. TPU cases looked like glass. The model tried to make everything look premium, which meant it invented textures that didn't exist.

Problem two: colours shifted. I sell a coral pink case. The AI rendered it as salmon, then as dusty rose, then as a peachy orange that was nowhere near the real product. Customers ordering coral pink and receiving coral pink still complained, because the listing photo was orange.

Problem three: fake text and logos. I prompted "phone case with brand logo" on a few listings where the real case had a small brand mark embossed on the back. The AI generated crisp, clear logos that looked nothing like the real embossed mark. One customer thought they were getting a designer case and filed a return request before I even shipped it, just based on comparing my photo to photos they found online.

All three problems came from me trying to make the images look better than the product. The AI obliged. It made them look much better. Then real customers received real products and felt misled.

The thirteen that worked followed three boring rules

The successful images were boring. I mean that as a compliment. They showed the product clearly, accurately, and with nothing else competing for attention.

Here's the prompt structure that worked:

"[Product name] on plain white background, even lighting, no shadows, accurate colour, simple product shot, front view"

For a patterned case I added "showing [specific pattern]" and for wallet cases I added "closed, showing exterior only." I kept every prompt under twenty words. I didn't ask for studio lighting, bokeh, marble surfaces, or premium anything.

The images came out clean. White background, product centred, colours that matched the real item when I held it up to the screen. I uploaded them alongside my phone photos—sometimes I used the AI image as the main listing photo, sometimes as the second or third image in the gallery.

These thirteen listings got 40% more clicks than my old photos (I checked Jumia's seller dashboard). Conversion rate stayed about the same—around 3%, which is normal for my category. No returns, no complaints. A few customers sent WhatsApp messages asking about colours, and when I sent them my original phone photo they said "yes, that's what I wanted" and completed the order.

The boring prompts worked because they didn't promise anything the product couldn't deliver.

What each type of case needed in the prompt

Not every product needs the same prompt. I tested different variations and tracked what worked for each category.

The key was to describe what the product is, not what I wanted it to look like. When I wrote "soft-touch silicone case," the AI rendered a case with a matte texture that looked expensive. The real case was glossy. When I wrote "silicone phone case, plain black," the AI gave me glossy black silicone. Perfect.

For patterned cases, I had to name the pattern exactly. I sell a case with a marble print. First prompt: "phone case with marble pattern." The AI generated a case that looked like it was carved from actual marble. Second prompt: "phone case with printed marble pattern, white background." Much better—it looked like a print, which is what it is.

When to use AI images and when to just use your phone

I still use my phone for most product photos. AI images work for specific situations, and they don't work for others.

Use AI when: you need a clean white-background image and your phone photos have messy backgrounds or bad lighting. The AI is faster than setting up a lightbox, and it's more consistent. I now generate AI images for the main listing photo and use my phone photos as secondary images so customers can see the real texture and lighting.

Don't use AI when: the product has fine details, text, logos, or anything that needs to be exactly right. The AI will invent details. I tried generating images for screen protectors (which have brand names printed on the packaging) and charging cables (which have logos on the plug). Every single image had fake text or logos that didn't match the real product. I deleted them all and went back to phone photos.

Don't use AI when: you're selling anything where material and texture matter more than shape. Leather goods, fabric cases, anything with stitching or embossing. Customers need to see the real texture, and AI will always make it look smoother and more premium than it is.

I now use AI for about 30% of my listings—plain cases, simple designs, anything where shape and colour are the main features. For everything else, I use my phone and a white bedsheet.

The one Jumia-specific thing that matters

Jumia's search algorithm seems to favour listings with clean, white-background images. I can't prove this, but my AI-image listings consistently show up higher in search results than my old phone-photo listings, even when the titles and prices are identical.

I tested this by creating two listings for the same product—one with an AI image, one with my phone photo. Same title, same price, same description. The AI-image listing got 60% more impressions over two weeks. I don't know if Jumia's algorithm explicitly favours white backgrounds, or if customers just click on cleaner images more often and that signals quality to the algorithm. Either way, it's real.

The other thing: Jumia customers are suspicious of photos that look too good. I had one listing with a particularly beautiful AI image—perfect lighting, subtle shadow, the case looking like it cost KSh 5,000. I priced it at KSh 900. I got clicks, but almost no conversions. When I replaced it with a more boring AI image, conversions went back to normal. Customers in Kenya have seen enough misleading listings that they've learned to distrust photos that look too professional for the price.

Questions people ask

Can I use AI images for all my Jumia listings?
Only if you're selling simple products with no fine details. Plain cases, simple accessories, anything where shape and colour are the main features. Don't use AI for anything with text, logos, complex patterns, or materials that need to look realistic.

Which Stability AI model should I use?
Stable Image Core worked for all my product shots. I tried SD 3.5 Large for a few images and the results were sharper, but also more likely to invent details that didn't exist. Core is safer for e-commerce.

How do I stop the AI from making my product look too premium?
Don't use words like "luxury," "premium," "studio lighting," or "high-end" in your prompt. Describe the product plainly: material, colour, shape, white background. That's it.

Should I tell customers the photos are AI-generated?
I don't mention it in the listing, but I always include at least one real phone photo in the gallery. If a customer asks on WhatsApp, I send them the real photo immediately. Transparency avoids returns.

I'm still using AI images for new listings, but I'm more careful now. The goal isn't to make my products look better than they are—it's to show them clearly and accurately so customers know exactly what they're buying. If you want to try this yourself, start with five or ten products, keep the prompts boring, and track what comes back as returns. That'll tell you faster than anything else whether your prompts are working.

You can generate product images with Stable Image Core at Kryotta—same workspace where I write all my product descriptions and WhatsApp replies.