Which images actually got M-Pesa orders: the test setup
I run a small accessories business in Nairobi—phone cases, laptop sleeves, tote bags, wireless earbuds—selling through WhatsApp catalogue and Instagram Shop. My product photos were phone shots taken on my desk with whatever light came through the window. They were honest, but they weren't moving stock. My Instagram Shop conversion sat at 1.2 per cent, and half my WhatsApp catalogue browsers went silent after I sent the product link.
In February I ran a test. I picked twelve products—four electronics, four bags, four phone accessories—and created four versions of each listing. Version A: my original phone photo. Version B: AI-generated image from Stable Image Core. Version C: AI image from SD 3.5 Large. Version D: a "hybrid" where I shot the product on white paper and asked Stability AI to clean up the background and lighting. I rotated which version appeared first in each product's gallery, published all forty-eight images across WhatsApp and Instagram, and tracked which listings turned browsers into M-Pesa payments over three weeks.
One approach drove 4.1 per cent conversion and cut my time per listing from twenty minutes to three. Another got flagged by Instagram's commerce review and cost me two days of downtime. Here's what worked, what failed, and the cost in shillings per image versus hiring a photographer in Nairobi.
The AI-only images looked great and failed Instagram's review twice
I started with pure AI generation. I wrote detailed prompts describing each product—"black leather laptop sleeve with bronze zipper, lying flat on white marble surface, soft natural light from left, professional product photography style"—and ran them through Stable Image Core in Kryotta. The images came back gorgeous. Clean backgrounds, perfect lighting, no shadows or clutter. I was thrilled.
Instagram rejected eight of the twelve listings within six hours. The reason: "digitally created content that misrepresents the product." WhatsApp catalogue doesn't pre-screen images, so those went live, but three customers asked "is this the actual item?" before going quiet. One sent a laughing emoji and the word "Photoshop."
The problem wasn't quality; it was believability. The AI images were too perfect. The leather texture looked generic, the zipper pulls didn't match my actual bronze hardware, and the lighting was the kind you see in a Jumia banner ad, not a small business listing. Shoppers in Nairobi have seen enough oversold products to spot when an image doesn't match the seller's scale. I'd accidentally made my KSh 1,200 phone case look like a KSh 4,500 import.
I tried again with more specific prompts—"slight fabric texture visible, small crease near seam, realistic product photo not advertisement"—and SD 3.5 Large. Better, but Instagram still flagged four listings for review, and my WhatsApp conversion on those products stayed flat at 0.9 per cent.
The hybrid method—real photo, AI cleanup—converted at 4.1 per cent
I changed approach. I shot each product on a piece of white printer paper taped to my desk, using window light and my phone. The photos were rough: shadows under the product, a corner of my keyboard visible in one frame, uneven brightness. Then I uploaded each image to Kryotta and used Stable Image Core with this prompt:
"Clean product photo: remove background, replace with pure white, even out lighting, keep all product details and textures exactly as shown, professional e-commerce style."
The AI kept the actual product—every stitch, every scuff, the exact bronze zipper pull—and fixed everything around it. Shadows gone, background pure white, lighting consistent. The images looked professional but still real, because the product itself was real. Instagram approved all twelve on first submission. WhatsApp browsers stopped asking "is this what I'll get?"
Over three weeks those hybrid listings converted at 4.1 per cent on Instagram Shop and drove eleven M-Pesa payments from WhatsApp catalogue browsers who'd previously gone silent. The original phone photos converted at 1.3 per cent. The pure AI images that survived review converted at 0.8 per cent, worse than my starting point.
Six prompts that worked and three that got flagged
The hybrid approach only works if your cleanup prompt is specific enough to fix problems without inventing details. Here are the six prompts that gave me clean, honest images:
For removing messy backgrounds: "Remove background completely, replace with solid white (#FFFFFF), keep product exactly as photographed, no changes to colour or texture."
For fixing uneven phone lighting: "Even out lighting across product, remove harsh shadows, maintain natural look, do not brighten colours beyond original photo."
For products with reflective surfaces (earbuds, phone cases): "Clean reflections on surface, remove surrounding objects visible in reflection, keep product shape and colour unchanged."
For fabric items with wrinkles: "Smooth minor wrinkles in fabric, keep overall drape and texture realistic, do not make fabric look stiff or artificial."
For products on cluttered surfaces: "Isolate product, remove all background objects, place on clean white surface, preserve all product wear and details."
For fixing colour cast from bad window light: "Correct colour balance to neutral, remove yellow/blue tint, match product to true colour, do not oversaturate."
Three prompts got me in trouble. "Professional studio lighting" made a KSh 800 tote bag look like a KSh 5,000 designer piece and Instagram flagged it. "Perfect product photography" added a subtle glow I didn't ask for and changed the leather tone. "High-end commercial style" gave me an image that looked like a stock photo, and two WhatsApp customers asked if I was dropshipping. Keep the prompt focused on fixing specific problems, not achieving a vibe.
Cost breakdown: KSh 8 per image versus KSh 1,500 for a photographer
Stable Image Core in Kryotta costs about KSh 8 per image at current rates. I generated forty-eight hybrid images (one original photo + one AI cleanup per product, plus a few retries) for roughly KSh 400 total. Shooting the original phone photos cost me nothing but fifteen minutes per product—three hours total.
I'd priced a product photographer in Nairobi the month before: KSh 1,500 per product for a basic setup, KSh 18,000 minimum for twelve items. Some photographers quote per hour (KSh 3,000–5,000) and can shoot faster if you're organized, but you're still paying for travel, setup time, and edits. For twelve products the quote came to KSh 22,000 after edits and file delivery.
The hybrid method saved me KSh 21,600 and three days of scheduling. It didn't save time versus my original phone photos—shooting plus AI cleanup took about the same fifteen minutes per product—but the conversion rate jumped enough that I made back the KSh 400 image cost on the first two sales. A photographer would've given me better images than my hybrid method, but not three times better, and definitely not KSh 21,600 better for a business doing six to eight sales a week.
What didn't work: lifestyle shots, model hands, and context scenes
I tried generating lifestyle images—someone holding the tote bag on a Nairobi street, earbuds in use during a workout, laptop sleeve on a café table. Every single attempt looked wrong. The AI-generated hands had weird finger joints, the Nairobi street backgrounds looked like stock photos of Lagos or Accra, and the café table had that generic "AI coffee shop" look that doesn't match any actual Nairobi café I've seen.
One image showed a woman's hand holding a phone case against a blurred background that was clearly meant to be urban Kenya but had street signs with unreadable text and a boda that looked slightly off. A customer replied "ulitengeneza wapi hii picha?" (where did you make this image?) and I knew I'd lost the sale. Lifestyle shots need to be real photos of real places, or shoppers in Nairobi will clock the fakeness immediately.
I also tried adding context to plain product shots—"place laptop sleeve on wooden desk next to coffee cup and notebook." The AI added objects that didn't match Kenyan context: a ceramic mug that looked European, a notebook with English cursive writing, a desk surface that was too polished. These details sound minor, but they add up to an image that feels foreign, and that kills trust for a local seller.
The one-line pitch that made customers okay with AI cleanup
After the "Photoshop" comment and the "is this real?" questions, I added one line to my WhatsApp catalogue descriptions and Instagram captions: "Photo cleaned up for clarity—actual product shown, colours and details accurate."
Responses changed. Customers stopped questioning the images and started asking normal questions about size, colour options, and delivery time. Two people replied "smart" and placed orders the same day. One asked "which app?" and I sent her a voice note explaining the process, which turned into a fifteen-minute conversation about her own business and a referral to her chama group.
Being honest about using AI to clean up photos didn't hurt trust; it built it. Trying to pass off AI images as untouched product photography destroyed trust. The difference is one sentence, and that sentence saved me more sales than any prompt refinement.
Questions people ask
Can I use AI images for Instagram Shop in Kenya or will they get rejected?
Pure AI-generated images get flagged often. Hybrid images—real photo of your actual product, AI cleanup of background and lighting—pass review consistently. Instagram's commerce policy wants to see the real item, not a representation. If you're cleaning up a real photo and adding the disclosure line, you're fine.
How much does it cost to generate product images with Stability AI versus hiring a Nairobi photographer?
Stable Image Core costs about KSh 8 per image in Kryotta. Photographers in Nairobi quote KSh 1,500–3,000 per product or KSh 3,000–5,000 per hour. For twelve products I spent KSh 400 on AI cleanup versus a KSh 22,000 photographer quote. The photographer's images would be better, but the hybrid method gave me 4.1 per cent conversion, which was enough.
Do AI product photos actually increase M-Pesa conversions on WhatsApp catalogue?
In my test, hybrid images (real product + AI cleanup) converted at 4.1 per cent versus 1.3 per cent for my original phone photos. Pure AI images converted at 0.8 per cent, worse than my starting point. The cleanup worked because it kept the product real while fixing amateur lighting and backgrounds. WhatsApp browsers trust clean images of real items.
Which Stability AI model is better for product photos, Core or SD 3.5 Large?
For hybrid cleanup I got nearly identical results from both. Stable Image Core is faster and cheaper; SD 3.5 Large gave slightly sharper edges on fabric texture but not enough to matter for WhatsApp or Instagram. For pure generation (which I don't recommend for commerce), SD 3.5 Large produced more realistic materials, but both still failed the "does this look like a real small business product?" test.
I still shoot products on my desk with my phone. I just don't post those photos raw anymore. Three minutes of AI cleanup turns an amateur shot into something that converts, costs me eight shillings, and doesn't make customers ask if I'm dropshipping. If you're selling on WhatsApp or Instagram Shop and your product photos look like mine used to, try the hybrid method once. Shoot your actual product, clean it up in Kryotta, add the one-line disclosure, and watch what happens to your M-Pesa notifications.
