I Generate Ad Creatives for Shopify Products. Here's Exactly How the Pipeline Works.

I've been building an AI ad creative generator for the past few months. I've run the pipeline on 30+ products — skincare, sneakers, keyboards, perfume, candles, laptops, backpacks, watches. Here's exactly what happens between "upload photo" and "download ad."

No hand-waving. Actual steps, actual costs, actual failure modes.

Step 1: Product analysis (~3 seconds)

Before generating anything, GPT-4o looks at your product photo and extracts:

• Dominant colors (hex values)


• Material and finish (matte, glossy, metallic, fabric)


• Product shape and proportions


• Visual mood (premium, playful, minimal, rugged)

This sounds like a nice-to-have. It's not. Without it, the AI generates backgrounds that clash with your product — a warm amber scene for a cool blue tech product, or a minimal white backdrop for a rustic candle.

The color analysis feeds directly into the scene generation prompt. The background complements the product instead of fighting it.

Cost: a single GPT-4o vision call. Fast and cheap.

Step 2: Copy generation (~4 seconds)

One Claude API call generates the full ad copy for all selected angles. Each angle gets:

• A headline (platform-specific character limits — Facebook allows more, X needs punch)


• Body text (2-3 sentences, no fake social proof, no invented statistics)


• A CTA ("Shop Now", "Get Yours", whatever fits the angle)

The angles matter more than people think. A "pain point" ad for a skincare serum ("Tired of dull, uneven skin?") hits completely differently than a "benefit" ad for the same product ("Visible brightening in 2 weeks"). These aren't just different words on the same layout. They're different emotional triggers aimed at different parts of the buying funnel.

I use 5 angles: pain point, urgency, social proof, curiosity, benefit. The default is pain point + urgency because those consistently test best for cold traffic.

Cost: single LLM call for all angles. Trivial.

Step 3: Scene generation (~20 seconds)

This is where most tools and I diverge.

The common approach is to send the product photo and a prompt like "create a Facebook ad featuring this product on a marble countertop with the headline 'Your Skin Deserves Better'."

I tried that on 9 products. The scenes looked professional. The problem: the AI reimagines your product. It keeps the general shape and colors but invents its own version. Labels become unreadable. The GlowLab serum bottle came out with a completely different cap. The TypeForge keyboard lost half its keys.

So instead, I generate the scene WITHOUT the product. The prompt says: "Create a premium ad scene for a skincare product. Include the headline, CTA button, and promotional badge. Leave a prominent empty space for the product to be placed later."

The scene gets the product's color palette from Step 1, so it complements — but the product itself isn't in the image yet.

One GPT Image 2.5 Sunburst call.

Step 4: Product insertion (~25 seconds)

Now the AI takes the generated scene + your original product photo and composites them.

"Place this product naturally into the scene. Match lighting direction, add realistic shadows and reflections, adjust scale to fit the composition."

This is a second AI model call — GPT-4o with image output. It handles the product extraction (no separate background removal needed), perspective matching, shadow generation, and light integration.

The result: a scene that looks AI-generated (because it is) and a product that looks real (because it is).

Total time per image: ~50 seconds. The two-step approach takes roughly twice as long as a single-prompt method, but the product fidelity is the whole point.

Where it fails

I've run this on 30+ products. I won't pretend it works perfectly every time.

Text rendering is the weak spot. AI models have gotten much better at putting text in images, but they still occasionally misspell words or produce slightly wonky letterforms. GPT Image 2.5 Sunburst is the best at this — noticeably better than Flux or older DALL-E models — but it's not 100%. Very complex products get simplified. A keyboard with 84 individually labeled keys? The AI preserves the overall shape and layout but some key labels get mushy. A sneaker with intricate stitching patterns? It captures the big picture, loses fine detail. Phone photos with bad lighting still look like phone photos. The AI matches the product into the scene, but if your photo has harsh shadows or color cast, those carry through. A decent photo with even lighting produces dramatically better output. Small products get lost in wide scenes. A ring or earring stud in a 16:9 landscape banner can end up looking tiny. Instagram's 4:5 portrait format works much better for jewelry and small items.

The numbers

At $24/month for 100 images, a merchant can cover about 20 products (5 angles each). Total pipeline time: ~50 seconds per image when running parallel.

For comparison: a freelance designer charges $50-100 per ad set. An agency charges $3,000-5,000/month for ongoing creative. This is 100 images for $24.

Is every AI-generated image as good as a skilled designer's best work? No. But the designer produces 2-3 ads in the time this produces 20. And for testing — where you need volume to find what converts — 20 decent ads beats 2 perfect ones.

Try it

I built a free demo at the top of this site. Upload a product photo, see what comes out. No signup, no credit card, no email required. The output is the exact same pipeline described above.

If you're a Shopify merchant, we're launching this as a native Shopify app — pick a product from your catalog instead of uploading. Join the waitlist if you want early access.

Try it free →