Batch Generation at Scale: Variations, A/B Tests & Sanity

9 min read

How to generate dozens of variations, run real A/B creative tests, and name and organize the flood so you can actually find your winners next week.

Amateurs generate one image and hope. Pros generate twelve and choose. When output is nearly free and nearly instant, the winning strategy flips: you stop trying to nail it in one shot and start producing a spread, then editing ruthlessly. But volume without a system is just a messy folder of image_final_FINAL_v3.png files. This lesson is about scaling *and* staying sane.

Variations: The Portfolio Mindset

There are two ways to produce variations. Same prompt, new seed gives you different takes on one idea — great for picking the best composition. Systematic prompt changes vary one element on purpose: same product, five backgrounds; same character, four outfits. The second kind is where money lives, because you are exploring a decision space, not gambling on luck. If you have not internalized how seed controls repeatability, revisit Iterating & Refining.

A/B Creative Testing, Done Right

An A/B test is only meaningful when you change *one variable* between versions. Two thumbnails where A has a shocked face and B a calm face — with identical text, crop, and colors — teach you something. Two thumbnails that differ in five ways teach you nothing, because you cannot attribute the result. Generate variants that isolate a single lever: hook, color, framing, or facial expression. Then let real numbers — click-through, watch time, saves — pick the winner, not your gut at 1 a.m.

TestVariable changedWhat you learn
Thumbnail A vs BOnly the facial expressionWhich emotion earns clicks
Ad 1 vs 2Only the background colorWhich palette stops the scroll
Post X vs YOnly the first-line hookWhich promise pulls readers in

Systematic background variation (swap the last clause)

a sleek matte-black wireless headphone, centered product shot, soft studio lighting, on a minimalist concrete surface

Model: nano-banana

Keep the product and lighting identical, then run the same prompt swapping only the final surface clause: concrete, warm oak, pastel gradient, marble. Now your five images differ by exactly one variable — a clean A/B set.

Naming & Organizing: Future-You Says Thanks

A naming convention is the cheapest productivity upgrade you will ever make. Adopt one pattern and never deviate: client_project_asset_variant_date, for example acme_summer_hero-headphone_v02_0704. Group work into VAR2 projects so a campaign stays in one place, tag your keepers, and delete the rejects the same day you make them — clutter compounds. When you can find any asset in ten seconds, you can quote a client confidently and reuse past wins instead of regenerating from scratch.

  1. Pick one naming pattern and apply it to every single asset
  2. Keep one campaign in one VAR2 project — never scatter it
  3. Cull rejects daily; a lean library is a fast library
  4. Save winning prompts as reusable snippets for next time

Related glossary terms: Seed, Iteration, Prompt Engineering, Inference

Set up your batch, isolate one variable, and let the data crown a winner. Start a batch generation