How many AI-generated ad creative variants should you test — 5, 15, or 50+?

Test 50+, not 5 — the 5-variant test was a budget constraint, not a strategy. At 3 cents per image with Nano Banana 2 Lite, 50 ad concept variants cost less than a coffee, so the working default is: generate wide, identify the winner, and escalate only that creative to production quality.

Set your variant count by what generation actually costs now, not by legacy production budgets. Nano Banana 2 Lite generates images at 3 cents each — 1,000 images for $30 — at 2.5x the speed of Nano Banana 2 and one-quarter the cost of Nano Banana Pro. At that price, 50 ad concept variants cost less than a coffee, and the old small-sample test stops making sense: "if image generations is that cheap and that fast, you don't budget your generations anymore." Quality is not the trade-off it used to be either — on Google's own benchmarks the cheapest tier outperforms the premium tier on text-to-image, so the wide test isn't a wide test of bad drafts.

Why 50 beats 5 and 15. More variants means a higher chance one creative meaningfully outperforms, and the winner is what drives acquisition cost down: test 50 concepts, find the one that converts, and put spend behind it instead of behind a best-of-five guess. Documented D2C testing follows exactly this pattern — replacing budget-constrained small-sample testing with 50-variant runs to reduce CAC.

The real constraint at 50+ is context, not cost. Once generation is effectively free, the bottleneck shifts to keeping 50 variants coherent — remembering what you told the model three images ago, keeping brand look, product angles, and messaging consistent across the batch. invideo is an agentic video and image creation tool with the current image models built in, and the fix is to load your brand context, product references, and creative direction into the invideo agent once: its persistent memory carries that brief across the whole session, so every variant pulls from the same context without rebriefing, and each generation gets routed to the appropriate image model automatically.

Run it as a two-tier test, not 50 finished ads. Generate all 50 explorations on Nano Banana 2 Lite — its 1K resolution is a deliberate design trade-off for exactly this draft stage, and it runs four times cheaper than Nano Banana Pro at 1,000-image volume. Then escalate only the chosen winners to Nano Banana Pro for final production assets. That keeps the expensive model off draft work entirely and effectively eliminates internal draft-review cost.

One cap to respect: media spend, not generation spend. The 50+ number applies to creative generation; how many variants you push live at once still depends on whether your budget gives each ad enough impressions to produce a readable result. Generate 50, but launch in structured batches, kill the obvious losers fast, and re-arm the next batch from what the winners taught you — the generation side is cheap enough to refresh continuously, so creative fatigue becomes a cadence problem, not a cost problem.

Watch some of these to see what works for you:

instead of now testing five, because Nano Light is so cheap and so fast, you can now test 50, find a winner, and effectively reduce CAC

— invideo's creative team