Generate many ad variants cheaply and let the data pick the winner

Cheap variants change what you are allowed to test

When a new creative costs a booking, a shoot or a freelancer's afternoon, you test the two or three angles you are most confident about. That is the opposite of what a testing budget is for. Confidence is exactly the thing an experiment is supposed to challenge.

Drop the cost of a variant to a few credits and the math flips. Now you can afford the weird hook, the ugly background, the offer framing nobody in the room liked. Most of them will lose, and that is fine, because a loser you generated for the price of a coffee is not a sunk cost.

The goal is not more ads for their own sake. It is to make the kill decision cheap, so you retire underperformers without arguing about the money already spent on them.

Build a test matrix, one variable at a time

Volume without structure just gives you noise. The discipline that makes cheap generation pay off is changing one thing per creative, so that when a variant wins you know which lever moved it.

Pick your variables and cross them. A simple matrix is three hooks against two backgrounds against two formats, which is twelve creatives from one product. Keep the product shot fixed and vary the copy overlay in one row, then keep the copy fixed and vary the setting in the next.

• Hook: the first line or the first frame, the thing that stops the scroll.
• Setting: studio white, in-context lifestyle, bold flat color.
• Format: square feed, vertical story, wide placement.
• Offer framing: price-led, benefit-led, scarcity-led.
• Then hold everything constant except the one axis you are actually testing this week.

Plug it into the agent you already use

Figgshield is not another web app with its own login and its own tab. It is an MCP connector at https://api.figgshield.ai/mcp that your AI agent calls directly. You describe the batch in the chat you already work in, and the agent generates the files through the connector.

It works with Claude on desktop and in Claude Code, with Cursor and Windsurf, and with ChatGPT in Developer Mode on paid plans. Any MCP-capable agent can reach it.

Two models sit behind it. Nano Banana 2 from Google produces the stills, and Kling 3, with a faster Kling 3 Turbo tier, produces the motion. Both draw from a single credit balance, so a batch of static images and a batch of video clips bill against the same pool.

What a ten-variant batch actually costs

Pricing is at cost. Figgshield does not mark up Nano Banana 2 or Kling 3, so a credit buys the model's real rate and nothing on top. The only money the service makes is on credits people buy and never spend, which means nothing nudges you toward heavier renders than a test needs.

Run the numbers on a static test. Ten image variants at 1K or 2K are 8 credits each, so eighty credits for a full matrix of stills. Push a shortlist to 4K at 16 credits each only for the ones you are ready to run.

Add motion selectively. Turn a winning still into a five-second clip with Kling 3 for 35 credits by passing the image id straight into the video step, so the clip starts from a frame you already approved instead of describing it again.

The Pro trial costs $5, runs 7 days and includes 150 credits, enough for a full static matrix with room to animate the winners. There is no free tier. Cancel before it renews and you pay nothing more. See how credits work for the breakdown.

Reading the first 48 hours

Cheap generation solves supply. Your reporting decides which supply survives. With a dozen variants live, the early signal is rarely conversions, because you do not have the volume yet. It is the top of the funnel: thumb-stop rate on video, click-through on static, cost per click before cost per acquisition means anything.

Cut early and cut hard. A creative with a weak hook almost never recovers downstream, and paying to confirm it is the expensive way to learn a cheap lesson. Kill the bottom half in the first two days and move spend to the survivors.

One warning on the other side: do not crown a winner off a handful of clicks either. A variant that looks brilliant on forty impressions is a coin flip, not a result. Let the front-runners gather enough data to be real before you build the next batch around them.

A brief you can paste into your agent

The batch is only as good as the instruction. Treat the agent like a junior designer who is fast but literal: give it the fixed elements, the axis you are varying, and the count you want.

A brief that works looks like this: Generate 6 image variants of this product, 1080x1350, product centered, three with a clean studio background and three in a kitchen setting. Keep the logo bottom-left. Vary only the headline across these six: [your six hooks]. Send your product photo as a reference image so every variant stays on-brand. That reference is used in memory to guide the render and is never stored.

When you like a still, follow up with animate variant 3 as a five-second clip, slow push-in, keep the product sharp and the agent starts the video from that exact frame. Files land on a private link that stays live for seven days, so pull the keepers into your ad manager while they are there.

Point your agent at the connector, paste a brief like the one above, and you will have a launch-ready batch before your coffee is cold. The Pro trial is built for exactly this: real output, cancellable before it renews.