I made 45 product shots with Stability AI then Instagram Reels with Luma: which format got M-Pesa orders

I generated 45 product images with Stability AI, split them between Instagram stills and Luma video Reels, then tracked which format actually drove M-Pesa orders for my Nairobi fabric business.

KKryotta TeamProduct & research · · 9 min read
Reusable fabric shopping bag with geometric pattern on a kitchen counter next to fresh produce in natural light
Reusable fabric shopping bag with geometric pattern on a kitchen counter next to fresh produce in natural light

I made 45 product shots in Kryotta, posted half as stills and half as Luma Reels, then watched which format moved M-Pesa payments

I sell reusable fabric shopping bags and beeswax food wraps through a WhatsApp Business catalogue and an Instagram shop. I'm based in Nairobi, revenue last year was KSh 890,000, and most orders come through M-Pesa after someone sees a post, DMs me, and asks for the Paybill number. I handle all the content myself because hiring a photographer costs KSh 8,000–15,000 per session, and I restock or add new prints every month.

In January, I made 45 product images using Stability AI (Stable Image Core and SD 3.5 Large, both in Kryotta). Then I took half those images and turned them into short Reels using Luma Ray 2. I posted the stills to my WhatsApp status and Instagram grid, and the videos to Reels. Same products, same three-week window, same call to action: "DM for price and Paybill."

I tracked which format got more DMs, which led to M-Pesa confirmations, and what the cost breakdown looked like compared to hiring a content creator. The results weren't what I expected, and the difference between product categories was bigger than the difference between stills and video.

What I generated and what failed immediately

I started with the fabric bags. I sell three sizes—small (for vegetables), medium (groceries), and large (bulk shopping)—in five prints: kitenge, plain canvas, geometric, floral, and a Maasai-inspired pattern. That's fifteen SKUs. I wanted three shots per SKU: one flat lay, one styled (bag on a kitchen counter with produce spilling out), one in-hand.

Stable Image Core gave me clean flat lays in under a minute each. Prompt: "Reusable fabric shopping bag, kitenge print, flat lay on white background, soft natural light, product photography." The lighting was even, the colours were accurate, and the fabric texture looked real enough that two people later asked if I'd hired a studio.

The styled shots were harder. Prompt: "Reusable shopping bag on kitchen counter, fresh vegetables, bright natural light, Nairobi kitchen." The AI gave me counters that looked vaguely East African but placed the bag at weird angles, and the vegetables were sometimes anatomically wrong—one carrot had a fork coming out of it. I ran each prompt four times and kept the best result. Out of 45 styled shots, I discarded 18 and regenerated them with SD 3.5 Large, which handled complex scenes better but still struggled with hands holding bags. The fingers were often too long or merged with the strap.

For the beeswax wraps, I needed shots that showed the product folded, flat, and wrapping a bowl. The flat and folded images worked. The bowl-wrapping shots looked fake—AI doesn't understand how fabric clings to curved surfaces when there's beeswax involved. I ended up using only the flat and folded versions and skipping the "in use" angle.

Total generation time: four hours across two evenings. Cost: KSh 0 beyond my Kryotta workspace, which I was already paying for.

Turning half the images into Reels and what Luma Ray 2 actually did

I picked the 22 best stills and fed them into Luma Ray 2 with a simple brief: five-second clips, subtle zoom or pan, no transitions. Prompt for the first batch: "Slow zoom on reusable shopping bag, kitenge print, product video, clean and minimal." Luma gave me smooth motion, no glitches, and the colours stayed consistent. The wraps were trickier—one video showed the fabric "breathing" as if it were underwater, so I regenerated it with "static fabric, no movement except camera."

I added text overlays in CapCut (free): "Reusable. Washable. KSh 450." and "DM to order." Exported at 1080×1920, uploaded to Reels and my WhatsApp status. The whole video process took another three hours.

If I'd hired a content creator, the going rate in Nairobi for this kind of work is KSh 12,000–18,000 for a batch of ten Reels including filming and editing. I made 22 videos for zero additional cost.

Which format got DMs and which got M-Pesa confirmations

I posted stills and videos on a schedule: Monday and Thursday for Reels, Tuesday and Friday for grid posts, Wednesday for WhatsApp status (mix of both). I used Instagram Insights and a simple spreadsheet to track DMs by post type, then matched DMs to M-Pesa payments using the phone numbers.

Reels got 3.2 times more DMs than stills. Over three weeks, the 22 Reels pulled 74 DMs. The 23 grid posts pulled 23 DMs. WhatsApp status views were harder to track, but I got 19 orders that people said came from "seeing it on your status," and 14 of those were from video statuses.

But here's the part that mattered: conversion from DM to M-Pesa payment was higher for stills. Reels pulled more lookers, but stills pulled more buyers. Of the 74 Reel DMs, 31 led to payment (42%). Of the 23 grid-post DMs, 14 led to payment (61%).

I think it's because video attracts casual scrollers who engage but don't have immediate intent, while someone who stops on a still and DMs is already closer to buying. The Reels also pulled more "How much?" messages that went nowhere, while grid-post DMs were more likely to open with "I want the medium kitenge bag, what's your Paybill?"

The beeswax wraps behaved differently. Video worked better there—probably because the product is less familiar and people needed to see it folded and unfolded to understand it. Wrap Reels converted at 48%, wrap stills at 39%.

What didn't work and what I'd redo

The biggest failure was lighting consistency. I generated the first 20 images in one session, then came back two days later and made 25 more. The second batch had cooler, bluer light even though I used identical prompts. When I mixed them in the same carousel post, people noticed. One DM: "Why does this bag look different from the other one?" I had to explain they were the same print. Next time, I'll generate all images in one session or save the exact seed and settings.

The AI also struggled with scale. The large bags sometimes looked the same size as the small ones because the AI didn't have real-world context. I added a banana for scale in three regenerations (yes, really), and that helped.

For video, the subtle zoom worked better than pans. Pans made the product look like it was sliding across the table. I also learned that five seconds is too short for WhatsApp status—people scroll fast and the video ends before they process it. Seven seconds worked better.

If I were doing this again, I'd shoot one reference photo per product with my phone, then feed that into Stability AI as a style guide. The fully synthetic images were good, but they lacked the tiny imperfections (a wrinkle in the fabric, a shadow from the strap) that make product photos feel real.

The cost breakdown and what I'd pay a creator

Here's what I spent versus what a content creator would have charged:

  • Kryotta workspace: KSh 0 incremental (already subscribed)
  • Time: 7 hours total (generation, video editing, posting)
  • Content creator for 45 stills + 22 Reels: KSh 25,000–35,000 based on quotes I got last year

I saved at least KSh 25,000. But the time cost is real—seven hours is nearly a full working day. If I valued my time at KSh 500/hour (low estimate), that's KSh 3,500, so net saving is still KSh 21,500.

The bigger win was speed. I made all 45 images and 22 videos in one week. A creator would have needed two weeks (one for shooting, one for edits and revisions), and I'd have had to coordinate schedules, send samples, and approve drafts. With AI, I iterated in real time.

What I'm doing differently next month

I'm keeping both formats but changing the mix. I'll make fewer Reels (10 instead of 22) and focus them on the beeswax wraps and any new products that need explanation. For the bags, I'll stick to stills and carousels because they convert better and take less time.

I'm also testing one new thing: using a Reel to drive traffic to a WhatsApp catalogue link instead of asking for DMs. Instagram added a "WhatsApp catalogue" sticker in January, and I want to see if it shortens the path to payment. Right now, the flow is: see Reel → DM → I send price and Paybill → they pay → I confirm and arrange boda delivery. If I can cut out the DM step, I might convert more of those casual scrollers.

I'll also regenerate the wraps with SD 3.5 Large and better prompts. The current images are fine, but I think I can get closer to the real texture if I describe it as "organic cotton fabric with visible weave, coated in beeswax, slight sheen, folds naturally."

For anyone selling physical products in Nairobi and doing your own content: AI stills are good enough to replace a photographer for most SKUs, especially if your product is simple and doesn't need hands or complex staging. Video works better for unfamiliar products or anything that benefits from motion. And if you're choosing between the two, test both and track M-Pesa, not just DMs.

Questions people ask

Can AI-generated product photos actually look real enough for Instagram shops?
Yes, if you're careful with prompts and pick the right model. Stable Image Core handles simple products (flat lays, single objects) well. SD 3.5 Large is better for styled shots with props or backgrounds. The giveaway is usually lighting—AI loves perfect, even light, which can look too clean. Add a prompt like "natural window light, soft shadows" to make it more believable.

Do Reels or stills convert better for M-Pesa payments?
It depends on the product. In my test, Reels got more engagement but stills converted at a higher rate for familiar products (fabric bags). For something new or complex (beeswax wraps), video worked better because it explained the product. Track both engagement and actual payments, not just likes.

How much does it cost to make AI product videos in Kenya?
If you're using Kryotta, the workspace covers both image generation (Stability AI) and video (Luma Ray 2), so there's no per-video cost. You'll spend time on editing and posting, but the tools themselves don't charge per output. Compare that to KSh 12,000–18,000 for a content creator to shoot and edit ten Reels.

What's the biggest mistake people make with AI product photos?
Inconsistent lighting across a batch. If you generate images over multiple sessions, the AI might give you different colour temperatures or shadow directions even with identical prompts. Generate everything in one go, or save your settings and reuse them exactly.

I'm still testing formats and models every month, and the results shift depending on what I'm selling. If you're running a WhatsApp catalogue or Instagram shop and want to try AI content without hiring out, Kryotta gives you both image and video tools in one workspace. Start with stills, see what converts, then add video where it makes sense.

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