QRocodile

I spent a week making AI QR codes. They're gorgeous and mostly useless.

I build QRocodile, a QR code designer, so I spend an unreasonable amount of time thinking about what makes a code actually visually unique — which is exactly why I went down the AI-QR-code rabbit hole.

You've seen them everywhere: the QR code that's secretly a coral reef, a frosted cake, a swirl of woodwork. Every few weeks another one goes viral, with a thread underneath explaining exactly how it was made — Stable Diffusion, ControlNet, feed it a code as a structural hint, prompt it, done. It looks like magic.

So we did what any curious builder does: we tried it ourselves. A lot. And I want to be honest about what we found, because the threads never mention this part.

The art is real. The problem is everything that happens after!

• It's a scannability gamble. The model optimizes for looking good, not reading reliably. Half the generations scanned on my phone in good light and died on someone else's. "Reads on one phone" is not "reads on every phone," and you only find out after you've printed 500 flyers.

• You can't reproduce it. No source file, just a prompt and a random seed. Change the destination URL and the whole artwork regenerates from scratch. That exact code you loved? Gone.

• It fights every layout. An AI code wants to be the hero. Try to shrink it into a business-card corner or a footer and it falls apart.

• It's raster, not vector, and takes seconds to minutes on a GPU per attempt — times however many tries until one both looks right and scans.

• It can't be dynamic. Most real QR use isn't one hero image — it's codes minted on the fly, one per product, per ticket, per campaign link, thousands of them through an API. A minute on a GPU per code makes that a non-starter. Nobody spins up a GPU farm to render 5,000 menu codes that each take 40 seconds and might scan.

None of this makes AI codes bad. It makes them art pieces — wonderful for a one-off you'll personally scan-test to death, genuinely risky for anything you print at scale, put your brand on, generate dynamically, or need to look identical next month. (We wrote the full breakdown here if you want it.)

The question we think everyone got backwards

The AI approach asks: how do we take an image and bend it into a code? Start with art, hope it scans.

Flip it. What if the code itself were the thing you designed? Start from a real, valid QR code — quiet zone intact, error correction respected, contrast preserved — and then style every part of it: module shapes, corner finders, colors, gradients, an optional halo, even motion. At no point does it stop being scannable, because you never left the structure behind.

That's the whole idea behind QRocodile.io. Swap the URL and the look stays identical. Save a design, reuse it across a campaign. Export vector for a billboard. Generate ten thousand on the fly, each in milliseconds. The beauty is intentional, embeddable, and repeatable — the opposite of a lucky seed.

Where I actually want QRocodile to go — and it's free

Here's the part I'm most excited about, and honestly the reason I'm posting.

Right now a "designed code" means picking from the module styles, finder styles, and halo effects we've built. But that's a library we wrote — which is a ceiling. What I really want is for the design modules themselves to be a community thing, free for everyone: anyone can build a new module style, a new finder shape, a new halo pattern, and drop it into the pool. The more people contribute, the wider the design space gets — and it composes, so contributions multiply instead of just adding up.

And this is where AI comes back in, doing the thing it's actually good at. Not painting a code and praying it scans — designing one. You describe the look you want in plain language, and instead of gambling pixels, the AI assembles from that community pool into an on-brand, reliably-scannable, fully reproducible code that renders instantly. You get the "I typed a sentence and got something beautiful" moment, but the output is a real design asset with every advantage above. Same magic, none of the strings.

We're extremely fresh, so I'll be upfront: we are still working hard on our ideas. But we've started posting new styles as we build them on Instagram (@qrocodile.io) if you want to watch the design space fill in over the next weeks.

The plain black square walked so the coral-reef code could run. I don't think the next step is picking between beautiful or reliable — that was always a false choice. It's codes that are beautiful and dependable because they were valid from the first pixel to the last.

Happy to answer anything about the render pipeline, the scannability engineering, or why we bet on "design it" over "generate it." Would genuinely love pushback from people who've shipped QR stuff at scale.