Start with the free tier and know exactly what you're testing
Most UK sellers I talk to think they need to spend £40 a month on Midjourney or hire a product photographer before they can test whether AI visuals work for their ads. That's backwards. You should know whether an AI-generated lifestyle shot of your candle or your skincare set gets more clicks than your phone photo before you spend anything.
Stability AI's image models are free to use through Kryotta. You get Stable Image Core and SD 3.5 Large with no credit card, no trial countdown, and no surprise invoice. The checklist starts here because you need to prove the concept works for your product category before you pay for video or hire anyone. Generate ten variations of your hero product in different settings. Post them as carousel ads on Facebook. If your cost per click stays the same or goes up, AI visuals aren't your problem and you've spent nothing. If CPC drops, you've just found a channel worth investing in.
The failure mode: jumping straight to a £40 Luma subscription because you saw a nice Reel, then realizing your audience doesn't engage with video, or your product (say, industrial cable ties) doesn't benefit from motion. Test stills first.
Use the right aspect ratio for the platform or the ad gets cropped to nonsense
Instagram Reels and Facebook feed ads don't use the same dimensions. Reels want 9:16 vertical. Facebook feed tolerates 1:1 square or 4:5 portrait. If you generate a beautiful 16:9 landscape shot of your product and upload it to Reels, Instagram will crop it into a thin vertical slice, and your product will be half out of frame. I've seen this kill ads for a soap company in Cork: they used a wide shot of five bars arranged in a line, Instagram cropped it to show two bars and a lot of empty wood table, and the ad got 0.3% engagement.
When you're prompting Stable Image Core, specify the output size in your settings before you generate. For Reels, set 1080×1920. For Facebook carousel or single image ads, use 1080×1080. Don't generate one image and try to crop it later; you lose composition control and end up with dead space or a cut-off product. Kryotta's image editor lets you adjust the canvas, but it's faster to generate the right ratio from the start.
The trade-off: vertical formats work beautifully for fashion, cosmetics, drinks, anything where the product has height. They're awkward for flat-lay items like stationery sets or board games. If your product is wider than it is tall, test 4:5 portrait for Facebook and skip Reels until you have enough budget to art-direct a vertical scene that makes sense.
Write prompts that describe the scene, not the feeling you want
"A luxurious skincare bottle in an elegant setting" will get you a blurry bottle on a marble slab with a light leak that looks like every other AI skincare ad. "A 100ml frosted glass bottle with a black pump, standing on a linen cloth next to two eucalyptus stems, soft north-facing window light, white background" will get you something you can actually use.
AI image models are literal. They need objects, materials, lighting direction and background. They don't understand "premium" or "cozy" unless you translate that into concrete details. I tested this with a Shopify seller in Manchester who makes soy wax melts. First prompt: "Beautiful wax melts in a calming scene." Result: a waxy blob on a pastel gradient, no definition, no product clarity. Second prompt: "Six wax melt cubes in a white ceramic dish, sitting on a wooden tray with a folded grey towel, daylight from the left, shallow depth of field." Result: a clean, usable product shot that worked in a Facebook carousel and brought CTR up from 1.1% to 1.8%.
The common failure is adding too much. If you ask for "a candle on a table with flowers and a book and a cup of tea and fairy lights," the model will muddle all of them. Pick two supporting objects at most, and make the product the largest element in the frame.
Check hands, text and symmetry before you use anything in a paid ad
AI image generators fail predictably. Hands grow extra fingers or melt into the product. Text on labels comes out as convincing-looking gibberish. Symmetrical objects (bottles, jars, tins) end up with one side wider than the other. You can't fix these in post without Photoshop skills, so you need to catch them at the generation stage.
Generate five versions of every concept and throw out the broken ones. It takes two minutes and costs nothing. I've seen a jewellery seller in Dublin waste £60 on a Facebook campaign because the hero image showed a ring on a hand with six fingers, and the comments were entirely people making jokes instead of clicking through. The ad was paused after one day.
For products with visible text (labels, packaging, book covers), don't ask the AI to render the text. Generate the scene without it, then add your label in Canva or Photoshop as a separate layer. Stable Image Core can sometimes get three-letter words right, but anything longer turns into alphabet soup. If your product is the text (a print, a card, a poster), AI image generation isn't the right tool yet. Use a phone photo on a clean background.
Move to video only when stills prove the concept and you know your hook
Luma Ray 2 and Nova Reel both cost money. Luma charges about £24/month for the Creator plan (billed in dollars but I'm converting for clarity). Nova Reel pricing sits inside AWS, so you're paying per generation, roughly £0.08–£0.15 per five-second clip depending on resolution. Neither is expensive, but both are wasted if you don't know what you're trying to show.
Video works for products that have motion, transformation or a use case that's easier to demonstrate than describe. A candle being lit. A skincare serum being dispensed onto a hand. A tote bag being packed with groceries. If your product just sits there (a bar of soap, a notebook, a static piece of jewellery), a three-second video loop won't outperform a still image, and you've spent money generating it and budget testing it.
The checklist item: before you generate any video, write down the one action you want to show in five seconds. If you can't describe it in one sentence, the video will be vague, and vague videos get scrolled past. I tested this with a seller on eBay UK who sells loose-leaf tea. First attempt: "A calming tea scene with steam." Luma gave us a slow pan over a cup with some steam that didn't move naturally. No clicks. Second attempt: "Hot water being poured into a glass teapot with loose Earl Grey, steam rising, close-up." Nova Reel nailed it in two generations, and the Reel got 4.2% engagement, which is high for a product with no inherent drama.
Use video for the hook, not the whole ad
A fifteen-second Reel that's just your product spinning on a turntable will get you completion rates below 20%. A three-second video loop of the product in use, followed by ten seconds of text overlay explaining the offer, will hold attention and drive clicks.
The mistake I see most often: sellers generate a beautiful AI video, post it as a Reel with no text, no voiceover and no call to action, then wonder why it got likes but no traffic to the Shopify store. Instagram rewards watch time, but your goal is revenue. The video is there to stop the scroll. The text, caption and link do the selling.
For Facebook ads, you can use a video thumbnail to get cheaper CPMs than a static image, then let the ad copy do the work. Generate a five-second loop in Nova Reel, set it to repeat, and Facebook will treat it as video inventory, which is often less competitive (and cheaper) than image inventory in the UK ad auction during Boxing Day and January sales. I've seen CPMs drop from £8 to £5 just by swapping a still image for a looping video with identical framing.
Test one variable at a time or you won't know what worked
You generate a new AI product shot, write new ad copy, and change your audience targeting all in the same campaign. The ad performs well. Which bit made the difference? You don't know, so you can't repeat it.
The checklist rule: if you're testing AI visuals, keep everything else the same. Same audience, same copy, same budget. Run the AI image against your current best-performing photo for three days. If the AI version wins, you know the visual was the variable. If it loses, you haven't wasted a week or a hundred pounds on a full campaign redesign.
This sounds obvious, but I've watched sellers burn through their monthly ad budget testing five things at once, then conclude "AI images don't work" when the real problem was the audience or the offer. One variable per test. It's slower, but it's the only way to know what you're paying for.
Know when to stop and hire a photographer instead
AI-generated visuals work well for lifestyle shots, scene-setting and concepts that are expensive to shoot (your candle on a marble bathroom shelf, your tea on a cafe table in Paris, your skincare on a hotel balcony). They don't work well for products where material texture is the selling point. If you're selling linen shirts, knitwear, leather bags or anything where the customer needs to see the weave or the grain, AI will give you a smoothed-out approximation that looks like a render. That's fine for testing, but it won't outperform a real photo once you're scaling.
The checklist decision: if your AI-generated ads are profitable and you're spending more than £500/month on Facebook or Instagram, take £150 of that and pay a product photographer in your city for two hours and twenty shots. Use the AI images to test concepts and audiences, use the real photos for your evergreen campaigns and your product pages. You'll get higher conversion rates on-site, and you'll have assets you can use for PR, stockists and anywhere else a visibly AI image would hurt your credibility.
Questions people ask
Can I use AI-generated images in Facebook ads without getting flagged?
Yes, as long as the image follows Facebook's ad policies: no misleading claims, no before-and-after health images, no shocking content. The fact that it's AI-generated doesn't matter to the algorithm. What gets ads rejected is broken policy, not the tool you used.
Do I need to tell customers the product photo is AI if I use it on my Shopify store?
There's no legal requirement in the UK or Ireland to disclose that a lifestyle or scene-setting image is AI-generated, as long as the product itself is accurately represented. If you're using AI to show your candle on a table and the candle in the image looks like the candle you're selling, you're fine. If you're using AI to fabricate a feature the product doesn't have, that's misrepresentation, AI or not.
Which is better for product videos: Luma Ray 2 or Nova Reel?
Luma Ray 2 gives you more control over camera movement and handles complex prompts better. Nova Reel is faster and cheaper per generation. For simple product loops (pouring, rotating, close-ups), Nova Reel is enough. For anything that needs a specific camera move or a scene transition, use Luma.
How many variations should I generate before I pick one for an ad?
Five to eight. Generate a batch, throw out anything with obvious flaws (wrong proportions, broken symmetry, nonsense text), and test the top two or three against each other in a small campaign. Don't spend an hour generating fifty versions. Diminishing returns kick in fast.
If you want to try this checklist without spending anything, Kryotta gives you free access to Stable Image Core and SD 3.5 Large, plus paid options for Luma Ray 2 and Nova Reel when you're ready to test video. No trial limits, no surprise billing. Start here and generate your first batch this afternoon.
