Three days into building our World Cup Final Edition cover, the design brief changed. The trophy had to come out of the hero art.
On a normal job that means going back to the illustrator, or back into the file and masking around something that was never meant to be removed. Instead I generated a trophy-free version of the same background, in the same style, and dropped the new elements straight in. Ten minutes, not two days.
That's the part worth writing down. Not that the tool makes pictures.
The art came out of Nano Banana Pro, which is the nickname for Gemini 3 Pro Image. I run it through Google AI Studio, which is free and lives in a browser tab. Navy and gold, engraved trophy illustration, 2480 x 3307. Print resolution, because the model outputs up to 4K and I wanted this to survive being printed rather than just look fine on a phone.
What it actually costs
The whole cover build ran on a few dollars of image credits.
Repeat covers now cost me close to nothing. Not because the model got cheaper, but because I stopped treating generation as the workflow. The background exists. The player headshots exist, a banked library of them, all in one cel-shaded house style, ready to composite. The template exists. A new matchday cover is an assembly job now, not a generation job.
Generate once, template forever.
I think this is where most people lose money on image models. They regenerate the same thing from scratch every time they need it, chasing a slightly better version, and the credits drain into work they already did last week. The model is cheap. Regenerating is expensive. Those are different problems and people keep solving the wrong one.
What still breaks
Text inside generated images is the big one. Google is right that Nano Banana Pro is a step change on legible text rendering, and compared to where image models were a year ago it genuinely is. But for brand typography it still isn't dependable enough to ship. Our masthead has to be exactly our masthead, at exactly the right weight, every single issue. Close doesn't count. So I generate the art and composite the brand type on top, separately, every time. That's the gap I keep hitting and I don't think it's closing soon.
Style consistency across a library is the second one. Getting one player headshot in the house style is easy. Getting the fortieth one to sit next to the first one without looking like a different artist drew it takes real prompt discipline and a fair amount of throwing work away. Nobody tells you that part.
And the iteration cost is real if you're undisciplined about it. Without a system you will burn credits exploring, and exploring feels productive right up until you check the bill.
Worth noting for anyone doing client work: every output carries a SynthID watermark, invisible but detectable. That's a disclosure conversation, not a problem, but have it before someone else has it for you.
The bit I don't understand
I've shown this workflow to a lot of people who make marketing assets for a living. Almost none of them had opened AI Studio.
Some of it is naming. "Nano Banana Pro" is a great internal codename and a terrible thing to search for when you don't already know it exists. Some of it is that AI Studio reads as a developer tool, so creative people bounce off it before they find the image generation. And a lot of it is inertia. People have a subscription they're already paying for and they use that.
None of which is a complaint. It's just a strange thing to watch: a capable tool sitting there, free tier and all, while the people who'd get the most out of it don't know where it lives.
We ship the next cover in two weeks. Same template, new match, no new artwork.
If you're building brand assets with generative tools, where's your workflow still breaking? I'm most curious whether anyone has solved the in-image typography problem, because I haven't.