Performance

AI ad creatives

Performance work is a volume game: the concept matters less than having twenty honest variants to test. This is how you get them without twenty briefs.

What a test set costs

Concept drafts, 10 framesFlux 2 Dev · 10 × 1 cr10
Four crops on a text-savvy modelGPT Image 1.5 · 4 × 20 cr80
Eight one-variable variantsNano Banana · 8 × 5 cr40
One motion version, fast clipSeedance 2 Fast · 1 × 60 cr60
Total190 cr ≈ 53¢

Credit costs are live from the catalogue; the cash figure is Starter-plan pricing.

Creative testing dies on production cost. If each variant needs a designer, you test three; if each variant costs a credit, you test thirty and let the numbers pick.

The trick is to fix the concept first, then vary one thing at a time — crop, background, wardrobe, headline placement — so the test tells you something rather than just producing noise.

The workflow

How it actually runs.

Each step links to the part of the studio that does it.

01

Settle the concept cheap

Ten 1-credit drafts to find the frame worth scaling. Ten credits, not ten hours.

AI image generator
02

Render the crops

The ad recipe renders the same concept across the aspect ratios a campaign needs.

Ad variations recipe
03

Vary one variable

Feed the winner to the editor and change exactly one thing per run — surface, colour, wardrobe.

AI image editor
04

Make the motion version

Animate the best static for the placements that reward video.

Image to video

Watch out

Where this goes wrong.

  • Headline type inside a generated image is still risky. Generate the art, set the type in your layout tool.
  • Changing three things at once makes the test unreadable. One variable per variant.
  • Platform policy still applies to generated creative — the model has no idea what your ad account rules are.
  • Hands, logos and small UI details are where reviewers spot a generated ad first.

FAQ

What people ask.

Yes, with a commercial licence — included on Creator and Studio. The platform rules about claims, likeness and disclosure still apply exactly as they do to a photograph, and some ad networks now ask you to declare synthetic media, so check the policy for the network you are buying on.

Enough that the winner is not luck. On a 1-credit model, twenty drafts cost twenty credits, so the constraint is your analysis rather than your production budget. Fix the concept, then vary one element per variant so the result is attributable.

GPT Image 1.5 handles short type better than anything else here, and it is what the ad recipe defaults to. Even so, for a headline that has to be exactly right — legal copy, prices, a brand lockup — generate the art and set the type in your layout tool.

Ten concept drafts, four crops on a text-savvy model, eight one-variable variants and one motion version comes to 190 credits — about 53 cents at Starter pricing, against a working week of a designer's time for the same volume.

Run the first one today

22 credits a day on the free account, no card. That is enough to get through the draft stage of this workflow.

110+

AI models

Image, video & music — one prompt bar

22 / day

Free credits

Plus 0-credit models. No card to start.

Auto-refund

On failed renders

Credits reserved on start, refunded if it fails

One balance

For everything

Image, video and music on the same credits