The WhatsApp status update that got zero replies
I posted a photo of a chrome desk lamp to my WhatsApp status at 9 a.m. on a Thursday. The lamp was real—I'd imported twenty of them from Johannesburg the week before—but the photo looked fake. I'd generated it with Stability AI using a prompt I found on Reddit: "professional product photography, studio lighting, white background, 8K, ultra-realistic". The lamp hovered at a weird angle. The chrome finish had an oily sheen. The shadow didn't match the base.
By lunchtime I'd had twelve views and zero messages. One person replied three hours later: "Is this from a catalogue?" Not a question you want when you're trying to move stock.
I run a WhatsApp-first retail business in Harare. I sell imported homeware, small electronics, accessories—things I source from South Africa or China and mark up enough to cover shipping, customs, and the risk that half the shipment arrives damaged. My customers find me through status updates, referrals, or the occasional Facebook post. Most transactions happen in USD via EcoCash or cash on delivery. Product photos are the entire pitch: if the image doesn't look trustworthy, nobody asks the price.
I'd been using my phone for photos since I started two years ago. The results were inconsistent—good lighting if I shot near the window in the morning, dim and grainy if I waited until evening. I kept seeing posts about AI product photography saving money, so I signed up for Kryotta and generated fifty product shots with Stability AI over three weeks. I tested them against my usual phone photos and tracked which images got replies, which led to sales, and which made people suspicious. Here's the checklist I ended up with, and the cost breakdown that made me keep using AI despite the failures.
Prompts that worked: the three-part structure customers trusted
The images that got responses all used the same prompt structure. I didn't figure this out until I'd wasted two weeks generating fantasy versions of my actual products.
Prompt format that worked: "Product photography of [specific item and colour], placed on [neutral surface], natural daylight, soft shadows, taken with smartphone camera, slightly imperfect, realistic texture."
The key parts: "smartphone camera" and "slightly imperfect". When I dropped those phrases and used "professional studio" or "8K" instead, the images looked too polished. Customers assumed they were stock photos from AliExpress. The imperfections—a tiny reflection, a shadow that wasn't perfectly symmetrical—made the images believable. I was generating fake photos that looked real by telling the AI to make them look amateur.
I used Stable Image Core for most shots because it handled textures better than SD 3.5 Large. A leather wallet needed visible grain. A ceramic mug needed a slightly uneven glaze. Core delivered that about 70% of the time; Large made everything look like plastic.
Specific example that sold six units in two days: "Product photography of black leather bifold wallet, placed on light wood table, natural daylight from window, soft shadows, taken with smartphone camera, slightly imperfect, visible stitching and grain." The wallet looked like something I'd photographed on my kitchen table. Three people asked if I had it in brown. Nobody asked if it was real.
The backgrounds that failed: why white didn't work
Every guide I read said "white background" for product shots. In Zimbabwe's WhatsApp commerce, white backgrounds killed trust.
Customers here are used to seeing products in context—on a table, held in someone's hand, propped against a wall. A floating product on pure white looks like a Takealot listing or a scam ad. I tested this directly: I posted the same set of Bluetooth earbuds twice, once on white, once on a wooden surface. The white version got two replies asking "Where's your shop?" (translation: "Are you a real person or a dropshipper?"). The wooden surface version got eight price inquiries and four sales.
Backgrounds that worked: light wood, concrete, fabric (linen or cotton texture), outdoor table, kitchen counter. I kept a list of five surfaces and rotated them. The AI handled wood grain well but struggled with fabric—about half the time the fabric texture looked like a Photoshop filter. I learned to check the preview before posting.
Backgrounds that failed: pure white, gradient grey, "studio backdrop", anything that looked like a professional set. Also failed: overly decorative backgrounds. I tried "rustic wooden table with scattered coffee beans" for a set of mugs and got three messages asking if the coffee was included.
The cost breakdown: USD 0.40 per image versus USD 15 per photographer session
I paid a photographer in Harare USD 15 per session last year. A session meant thirty minutes, ten to fifteen products, and I had to bring everything to his studio in Avondale. If I wanted edits—background removal, colour correction—that was another USD 5 per image. Total cost for fifteen usable photos: around USD 30, plus transport and two hours of my day.
Stability AI cost on Kryotta: I'm on the standard plan at USD 20 per month. Stable Image Core costs roughly USD 0.40 per image (I generated about fifty images and used maybe thirty-five). SD 3.5 Large is slightly cheaper but I used it less. Total cost for fifty images over three weeks: the USD 20 subscription. Per-image cost if I only count the ones I actually posted: USD 0.57.
The photographer's images were better in one specific way: colour accuracy. If a customer bought based on the photo and the real product was a different shade, they'd ask for a refund or just never buy from me again. The AI got colour wrong about 30% of the time. A "navy blue" tote bag came out purple. A "rose gold" phone case looked orange. I started adding "accurate colour, no filters" to every prompt and the failure rate dropped to maybe 15%, but it never hit zero.
Trade-off: I kept using the photographer for anything where colour was critical—clothing, bags, anything fabric. I used AI for electronics, homeware, accessories where the shape mattered more than the exact shade.
The connectivity problem: generating images offline didn't work
I assumed I could generate a batch of images when I had Wi-Fi, download them, and post them later when I was on mobile data. That worked, but the workflow broke down because I couldn't preview images properly on my phone. Kryotta's image editor needs a stable connection—if I tried to open it on mobile data during a busy hour, it would time out or load a low-res preview that looked fine until I posted it full-size and saw the artifacts.
What worked: I generated images in batches of ten during morning hours when my Wi-Fi was stable, downloaded them immediately, and checked each one full-screen on my laptop before moving it to my phone. If an image had issues—weird shadows, blurry text, a handle that didn't connect properly—I regenerated it right then. Trying to fix things later on mobile data was a waste of time and money.
I also learned to keep the prompts simple. Complex prompts ("product photography of stainless steel water bottle, placed on marble countertop, with condensation droplets, backlit by golden hour sunlight") took longer to generate and failed more often. Simple prompts ("product photography of steel water bottle, placed on white marble, natural light") worked better and used less processing time.
The images customers ignored: anything that looked too perfect
I generated a set of images for picture frames using "professional product photography, perfect lighting, no shadows, crystal clear". They looked like CGI. I posted one to my status and got zero replies in two days. I regenerated the same frame with "smartphone photo, soft window light, slight shadow" and sold four frames in six hours.
The pattern held across categories. Images customers ignored: anything with perfect symmetry, no shadows, impossibly clean surfaces, or that "floating product" look. Also ignored: images where the product was too small in the frame or too large (cropped edges made people think I was hiding damage).
Images that got replies: slight imperfections, visible texture, realistic shadows, the product taking up about 60–70% of the frame. I started thinking of it as "Instagram versus WhatsApp status"—Instagram tolerates polish, WhatsApp status needs to look like something your friend just photographed.
One surprise: customers didn't mind if the background was obviously AI-generated, as long as the product itself looked real. I posted a set of mugs on a background that was clearly a digital wood texture (the grain repeated in a pattern), and nobody mentioned it. But if the mug's handle looked weird or the glaze was too shiny, I'd get messages asking if the photo was "the actual product".
The four-image test: which prompt styles converted
I tested four prompt styles on the same product—a black leather wallet—and posted each version on different days at the same time (11 a.m., when my status views peak). I tracked replies, price inquiries, and sales over four days.
Prompt A (professional studio): "Professional product photography of black leather wallet, studio lighting, white background, 8K, ultra-realistic." Result: 18 views, 1 reply asking if I was a reseller, 0 sales.
Prompt B (smartphone, neutral surface): "Product photography of black leather wallet, placed on light wood table, natural daylight, soft shadows, taken with smartphone camera, slightly imperfect." Result: 22 views, 7 replies, 3 sales.
Prompt C (lifestyle context): "Black leather wallet on café table, coffee cup in background, natural morning light, smartphone photo." Result: 19 views, 4 replies, 2 sales, 1 person asking if the coffee cup was for sale.
Prompt D (minimal, high contrast): "Black leather wallet on white marble, top-down view, natural light, sharp focus." Result: 15 views, 2 replies, 0 sales.
Prompt B won on replies and sales. Prompt C worked but added confusion. Prompts A and D looked too formal for WhatsApp commerce. I've used Prompt B's structure for thirty products since then and it's the most consistent performer.
The checklist I actually use now
Before I generate any product image, I check these six things:
- Is colour accuracy critical? If yes, use the photographer. If no, AI is fine.
- Does the prompt include "smartphone camera" and "slightly imperfect"? If no, rewrite it.
- Is the background neutral and believable? Wood, concrete, fabric, kitchen counter. Never pure white or gradient.
- Am I on stable Wi-Fi? If no, wait. Generating on mobile data wastes money and produces worse results.
- Does the preview look real at full size? Check on a laptop screen, not a phone. Zoom in on edges, shadows, and any text.
- Would I believe this photo if someone else posted it? If I'd ask "Is this real?", customers will too.
I still use my phone for about 40% of product shots—anything where I need to show scale (a bag held in someone's hand), colour accuracy (fabric, clothing), or multiple angles quickly. But for homeware, electronics, and accessories where the shape and general look matter more than exact colour, AI images work. They've cut my photography cost from around USD 60 per month to USD 20, and I can generate a new image at 10 p.m. if I get a late shipment, instead of waiting three days for the photographer's next available slot.
The images that work aren't the ones that look professional. They're the ones that look like I took them myself, because that's what WhatsApp commerce customers expect. I'm not trying to compete with Takealot's product pages. I'm trying to look like a real person selling real products, and AI helps with that as long as I tell it to look imperfect.
Questions people ask
Can I use Stability AI images for products I'm dropshipping? Technically yes, but customers will ask for proof photos if the product looks too polished or doesn't match the image when it arrives. I only use AI for products I physically have, so I can send a quick phone photo if someone's suspicious.
Which Stability AI model is better for product shots, Core or SD 3.5 Large? Core handles textures and fine details better—stitching, fabric grain, metal finishes. Large is faster and cheaper but makes everything look slightly plastic. I use Core for anything leather, fabric, or ceramic, and Large for simple electronics or items where texture doesn't matter.
How do I stop the AI from getting colours wrong? Add "accurate colour, no filters, true to life" to your prompt. It doesn't fix the problem completely, but it cuts failures from 30% to maybe 15%. For anything where colour is critical—clothing, bags, paint—use a real photo.
Do AI-generated images work on Facebook Marketplace or only WhatsApp? They work on Marketplace, but the tolerance for imperfection is lower. WhatsApp customers expect amateur photos; Marketplace customers compare your listing to dozens of others and notice if something looks off. I use slightly more polished prompts for Marketplace (drop "smartphone camera", keep "natural light") and save the really casual-looking images for WhatsApp status.
If you're running WhatsApp-first retail and spending money on product photography, try generating ten images with Stability AI on Kryotta and post them alongside your usual photos. Track which ones get replies. You'll know in a week whether it works for your products and your customers. Start with the free trial and generate five images before you pay anything—worst case, you've wasted an hour and confirmed that your phone photos are good enough.
