Most teams treat an ad creative test like a folder of exported files. That works until you need to answer a simple question: which idea actually won, and what changed between versions?
A lightweight data model can make AI-assisted creative work much easier to compare. The goal is not to build a full ad-tech platform. It is to preserve the intent behind each asset before motion, music, and editing make the differences harder to see.
Start with an experiment, not an asset
An experiment should define the commercial question first:
{
"experiment_id": "summer-landing-page-01",
"goal": "product_page_visit",
"audience": "first-time skincare buyers",
"placement": "vertical_social",
"hypothesis": "a problem-first hook will outperform a product-first hook"
}
Every concept generated inside the experiment inherits the same goal, audience, and placement. This prevents a common problem: comparing two attractive images that were designed to solve different problems.
Model the creative angle separately
A creative angle is the strategic idea, while an asset is one execution of that idea.
{
"angle_id": "problem-first",
"hook": "Dry skin should not decide what makeup you can wear",
"proof": "hydrating formula and close-up texture",
"cta": "See how it wears",
"constraints": [
"product visible within one second",
"headline readable at 390px width",
"leave safe area for platform UI"
]
}
This separation matters because changing a background color is a variant; changing the promise is a new angle. Without that distinction, a team may believe it tested six ideas when it only tested one idea in six colors.
Generate stills before video
Still images are cheap checkpoints. Produce several image variants for each angle, then score them before adding motion.
A useful review object can stay small:
{
"asset_id": "problem-first-v03",
"angle_id": "problem-first",
"type": "image",
"scores": {
"clarity": 4,
"product_readability": 5,
"hook_strength": 4,
"brand_fit": 3
},
"decision": "upgrade_to_video",
"notes": "Strong hook; simplify background before animation"
}
The exact scale is less important than consistency. Review at phone size, and do not let polish compensate for an unclear promise.
Upgrade only the winner
When a still passes review, create a child asset that records what motion is supposed to do:
{
"asset_id": "problem-first-v03-motion-a",
"parent_asset_id": "problem-first-v03",
"type": "video",
"duration_seconds": 6,
"motion_job": "reveal product, then direct attention to benefit",
"sequence": [
"0-1s: visual interruption",
"1-4s: product and benefit",
"4-6s: proof and CTA"
]
}
The parent-child relationship is important. It lets you compare a video to the still concept it came from, instead of treating every MP4 as an unrelated file.
Keep generation metadata, but do not confuse it with strategy
Store the prompt, model, seed, aspect ratio, and generation cost. These fields help reproduce an asset. They do not explain why the asset should work.
The hypothesis, angle, hook, proof, and decision notes carry the strategic context. Both layers are necessary.
A simple review pipeline
- Define one experiment and measurable goal.
- Create two or three distinct creative angles.
- Generate several still variants per angle.
- Score clarity, readability, hook strength, and brand fit.
- Upgrade only the winning still to motion.
- Record delivery metrics beside the original experiment.
- Use the result to update the next hypothesis.
I have been applying this image-first structure in TrendVis, where the workflow moves from campaign intent to creative angles, image variants, and then video for the selected concept.
The broader lesson is tool-independent: an AI creative workflow becomes more useful when every asset has a reason to exist and a traceable parent decision. A small schema is often enough to turn a pile of generated media into a repeatable experiment system.
Top comments (0)