How do you build an AI system for generating and A/B testing ad creative variants?

Build it as a closed loop: lock brand context and a winning reference ad once, generate variants along one dimension at a time — hook, script, format, product, market — ship every variant, and scale the winner by ROAS. Documented teams produce 8–10 variants of a single winning ad per day at roughly $30–$145 per variant.

Set the system up as five stages, in order. invideo is an agentic video creation tool with all the current generation models available, so the whole loop below runs inside one project.

1. Load brand context once. Upload brand guidelines, product images from every angle, color palette, and website links into the invideo agent's context tab before generating anything. This takes 15–20 minutes and persists across every variant in the project, so you never re-explain the brand — and it protects the identity elements that must stay identical while everything else changes.

2. Lock a winning template. Upload your best-performing ad — or a winner pulled from Meta Ad Library — and have the invideo agent deconstruct it into structural DNA: shots, framing, voiceover, setting, pacing, and cut timestamps. It tells you what drove conversion and what is safe to swap. Keep the edit structure, music bed, and pacing constant across all variants; that is what makes downstream A/B results readable.

3. Generate variants one dimension at a time. Clean testing means each variant changes a single variable against the locked template. The dimensions with documented workflows: hooks — the invideo agent generates 7–8 hook directions per iteration, each with a rationale, and the first 3 seconds decide performance, so spend the most iteration here; scripts and voiceovers — brief three script variations in different tones (enthusiastic, understated, skeptic) and three VO takes, then lock one; formats — produce a product film, a faceless UGC, and a full-character UGC in parallel rather than pre-selecting one; products — a product swap costs about 115 credits ($30) and takes 30 minutes once the first workflow is set, up to 12 per day; markets — a full localization (new cast, voiceover, UI screens) runs about 570 credits ($145), with 6 achievable in a day after the first. Combined, that's 8–10 variations of one winning ad per day from a single senior creative. The invideo agent routes each job to the right model — Seedance 2.0 for video clips, GPT-Image-2 for character and location sheets — so you never leave the platform to switch models.

4. Control cost per variant. Iterate on still frames first and spend video credits only on locked frames; generate clips in batches of 5 so a bad direction is caught early. Budget for rejection — one documented localization run discarded ~85% of generated clips, and its cost figures include those rejects. Per-variant actuals across documented runs span $30 (product swap) to $145 (localization), with fresh UGC creatives at ~$75 versus $300–$600 per creator on marketplaces.

5. Ship everything and recycle the winner. You cannot predict which variant wins without shipping it, so launch all variants and let platform delivery data decide — in one documented Meta comparison, a person-led UGC variant hit 6.2x ROAS against 0.8x for the product-only version. Then close the loop: re-upload the top performer into the invideo agent to generate the next round of structural variants — different products, audiences, and moods — while preserving the format that won.

Watch some of these to see what works for you:

This is exactly how real brands and agencies scale their campaigns. They don't just create one ad. They build a variant system for AB testing different audiences, seasonal campaigns, and platform-specific edits while protecting the brand identity.

— invideo's creative team