Myth one: Stability AI images pass for real product photos in paid ads
They don't. I've tested this across twenty-odd Facebook and Instagram campaigns since October, selling everything from scented candles to ceramic planters. The images look fine in your Shopify product grid—clean background, good lighting, no weird fingers because there are no fingers. Then you push them into a carousel ad with a £12 daily budget and the click-through rate sits at 0.4 per cent. Real photos of the same products, taken on my phone against a white bedsheet, get 1.1 per cent.
The problem isn't quality. Stable Image Core and SD 3.5 Large both produce sharp, well-composed shots. The problem is that people scrolling Instagram in January have seen a million AI images in the past six months. The lighting is too even. The shadows are too soft. There's a faint plasticky sheen on anything reflective. It doesn't look fake enough to be obviously AI, but it doesn't look real enough to make someone stop and think "I want that mug."
Here's what actually works: use Stability AI for background replacement, not the whole image. Take a real photo of your product on your kitchen table. Upload it to Kryotta's image editor, mask out the background, and let Stable Image Core generate a clean white or pastel backdrop. You keep the texture, the imperfect lighting, the tiny details that make it look like something you can actually buy. Then run that in your ad. My CTR on those hybrid shots averages 0.9 per cent, and the cost per click dropped from £0.87 to £0.61 across three campaigns in November.
The other workaround: use AI images for concept testing, not final ads. Generate five versions of a product in different settings—kitchen counter, bathroom shelf, bedside table—and run them as organic posts on your business Instagram. See which one gets the most saves. Then recreate that setting with a real photo and put ad spend behind it. You've spent nothing on a photographer and you've got data on what your audience actually wants to see.
Myth two: Luma Reels work for every product category
They work brilliantly for some products and fail completely for others, and no one tells you which is which until you've burned £40 testing it. I sell homeware. I tried Luma Ray 2 on a linen tea towel, a ceramic mug, a soy wax candle, and a brass drawer pull. The tea towel looked fine—gentle fabric movement, soft focus, the kind of thing you'd see in a lifestyle brand's Instagram story. The mug looked like it was melting. The candle wick flickered in a way that didn't match the (generated) flame. The brass pull just sat there, and because nothing else in the frame moved, it looked like a still image that someone had accidentally uploaded as a video.
Luma Reels need motion to look convincing. Fabric, plants, anything with hair or fur, liquids, smoke—those all work because the AI has learned how they move and can fake it well enough that your brain doesn't flag it as wrong. Hard, static objects with reflective surfaces don't work. The model tries to add subtle movement—a shift in the light, a faint wobble—and it just looks off.
Here's the step-by-step workaround I use now. First, decide if your product needs video. If you're selling drawer pulls or enamel pins or anything small and static, a carousel of still images will outperform video every time. If you're selling something people want to see in motion—a dress, a candle burning, a plant swaying on a shelf—then test this: shoot five seconds of real video on your phone. Just hold the product in your hand and rotate it slowly, or film the candle burning for five seconds. Upload that to Kryotta, use Nova Reel to extend it to fifteen seconds, and let the AI generate the middle ten seconds. The start and end are real, so it anchors the viewer's brain in reality, and the AI-generated middle section looks plausible because it's interpolating between two real frames.
I ran this test on three candle ads in December. Real video all the way through: £0.74 per click. AI video all the way through: £1.12 per click. Hybrid (real start and end, AI middle): £0.79 per click. The hybrid took me three minutes to set up and saved me from filming fifteen seconds of smooth, steady footage, which I can't do without a tripod I don't own.
Myth three: you can use the same prompt for every product and get consistent results
You can't. I learned this the expensive way. I'd written a prompt for Stable Image Core that worked beautifully for my ceramic mugs: "Product photo of [item] on white background, soft natural lighting, slight shadow, studio quality, sharp focus." I used it for twenty listings. Then I tried it on a brass candle holder and got something that looked like a 3D render from 2007. The lighting was harsh, the metal had no depth, and the shadow was a perfect gradient that no real light source would ever produce.
Different materials need different prompts. Ceramics and matte surfaces are forgiving—"soft natural lighting" works fine. Metal and glass need you to specify where the light is coming from and what it's reflecting. I now use "Product photo of [brass item] on white background, diffused window light from left, subtle reflections, studio quality" for anything shiny, and the results look like something a human lit and shot. Fabric needs "slightly textured background" or it looks like you've Photoshopped it onto a piece of printer paper. Candles need "warm lighting" or they look cold and medical.
Here's the process that's saved me the most time. Pick one product from your store. Write a prompt, generate an image, and look at it on your phone, not your laptop. If it looks wrong on a phone screen, it'll look wrong in an ad. Adjust the prompt—add a lighting direction, change the background texture, specify the type of shadow—and generate again. Do this five times. Once you've got a version that looks right, save that prompt as a template for that material type, not that product. Then when you shoot a new ceramic mug or a new brass drawer pull, you've already got the prompt that works for that surface.
Myth four: AI-generated video is cheaper than shooting it yourself, so you should use it for everything
It's cheaper in money and more expensive in time, and the trade-off only makes sense for specific use cases. A fifteen-second Luma Reel costs me nothing except the Kryotta subscription I'm already paying. Shooting fifteen seconds of product video costs me nothing except the phone I already own. But generating a Reel that doesn't look immediately fake takes me six to eight attempts, and each attempt is three minutes of waiting for the render. That's twenty-four minutes. I can shoot and edit a fifteen-second video in twelve minutes, and it'll look better.
The place AI video saves you money is when you need variations. I ran a Boxing Day sale on ceramic mugs. I had five designs. I wanted a video ad for each one, showing the mug rotating on a wooden table with a hand reaching in to pick it up. Shooting that five times would have taken me an hour—set up the table, film the rotation, film the hand, do it again for each mug, transfer the files, edit them. Instead I shot it once with a plain white mug, then used Kryotta's video editor to mask the mug in each frame and had Stable Image Core replace it with each of the five designs. Then I used Nova Reel to smooth out the motion where the pattern changed. Total time: thirty minutes. Total cost: £0, because I didn't hire anyone and I didn't buy five mugs to destroy in the filming process.
Here's the decision tree I use now. If you need one hero video for your homepage or your best-selling product, shoot it yourself or hire someone. If you need ten variations of the same video for testing, or if you need to show a product in a setting you don't have access to (a bathroom, a café, a hotel lobby), use AI. If the product is under £15 and the ad will run for less than a week, use AI. If the product is over £50 and the ad is your main acquisition channel, shoot it properly.
Myth five: if an AI image or video fails, the model is wrong and you should switch to a different one
The model is rarely the problem. I've tested Claude Sonnet 4.5, Gemini Flash, and GPT-OSS 120B for writing image prompts, and they all produce roughly the same results when you give them the same instructions. I've used Stable Image Core, SD 3.5 Large, and Amazon Nova Pro for generating product shots, and the differences are small—Nova Pro is slightly better at reflections, SD 3.5 Large is slightly faster, Stable Image Core is slightly more consistent with fabric textures. But none of them will save a bad prompt, and all of them will produce a decent image if you give them enough detail.
The thing that fails is usually your reference. You're asking the model to generate a photo of a product it's never seen, in a setting it's only seen in stock photos, lit in a way you haven't described clearly. It's guessing. The more you tell it, the less it has to guess. The best results I've had came from uploading a rough photo of the product, writing a prompt that describes the changes I want, and using Kryotta's image editor to mask the areas I want the AI to regenerate. That way the model isn't inventing the product from scratch—it's editing a real thing.
Here's the workflow that's cut my ad creative time from three hours to forty minutes. I shoot one reference photo of each product on my kitchen table, white background, phone camera, no special lighting. I upload it to Kryotta. I write a prompt: "Replace background with soft grey studio backdrop, add diffused natural light from top left, slight shadow bottom right, keep product exactly as shown." I generate three versions, pick the best one, and export it. If I need video, I shoot five seconds of the product sitting still, upload it to Kryotta, and use Nova Reel to add subtle motion—a slow zoom, a slight rotation. That's it. No model-hopping, no regenerating the same image fifteen times, no spending £200 on a photographer in Bristol who'll deliver the files in a WeTransfer link that expires before I've downloaded them.
Questions people ask
Can I use AI-generated images in Facebook ads without getting flagged?
Yes, as long as the image doesn't violate Facebook's ad policies—no misleading content, no before-and-after claims, no implied endorsements. The algorithm doesn't penalize AI images specifically, but users do. If your CTR is low, Facebook will stop showing the ad regardless of how it was made. Test AI images against real photos and let the data decide.
Which Stability AI model is best for product photography?
Stable Image Core for most products—it's fast and handles fabric, ceramics, and matte surfaces well. SD 3.5 Large if you need very high resolution or you're printing the images. Amazon Nova Pro if the product is shiny or reflective and you need accurate lighting. I keep all three in Kryotta and pick based on the material.
How long should a product video be for Instagram ads?
Fifteen seconds maximum. Instagram users scroll fast, and anything longer than that needs to justify its length with a story or a demo. For static product shots with subtle motion, aim for six to ten seconds. You'll pay less per view and the completion rate will be higher.
Do I need to disclose that an image is AI-generated in a Shopify listing?
Not legally, but if the image doesn't match the product you're shipping, you'll get returns and disputes. Use AI for backgrounds and settings, but make sure the product itself is accurate. If you're generating the entire product from scratch, test it against a real photo before you list it.
I've been using Kryotta to manage all of this since November—one workspace for image editing, video generation, and prompt testing, and I'm not switching between six different tools or paying for three separate subscriptions. If you're prepping for your January sale and you're tired of choosing between expensive product shoots and obviously AI images that don't convert, it's worth trying the hybrid approach. You can start at Kryotta and test it on five products before you commit to anything.
