I spent 577 credits teaching an AI to spin a can. Then I built the can.

I built a website for a made-up energy drink called VOLTIC. The showpiece is the can: it turns as you scroll down the page, and you can grab it and spin it yourself. It's live at web-voltic.vercel.app — go spin the can before you read how it nearly didn't work.

Every picture on that site was made by AI. The splash shots, the close-ups of water droplets, the moody hero image — all of it. And it looks great.

The can is a different story. The can cost me 577 credits — the tokens you buy to use AI image and video tools, like tokens at an arcade — across four separate attempts. Every attempt failed. Understanding why turned out to be worth more than the website.

Why the AI kept failing

Here's the thing I didn't understand when I started.

When you ask an AI for a picture, it isn't photographing an object that exists somewhere. It's imagining a brand-new picture, every single time.

So I could show it a photo of the front of my can and say "match this" — and it would, beautifully. But then it had to draw the back of the can. It has never seen the back. There is no back. So it made one up. A different one every time.

Show it the back instead, and it invents the front. There's no clever request that fixes this, because the AI isn't turning a can around — it's dreaming up a fresh can on every try and hoping. Each new attempt is a pull on a slot machine, and I paid for every pull.

The second problem was worse.

AI can't write small text. The little nutrition label on the can came back as scribbles at first — panel-shaped squiggles where words should be. That at least looked wrong. On a later attempt, something sneakier happened: the label came back crisp and perfectly readable, with numbers the AI had invented. It said 170 calories. My drink is supposed to have zero.

That's the sentence I carried out of this whole mess: beautiful and wrong is worse than obviously fake. A blurry label looks like a rough draft. A crisp label with made-up numbers looks like a fact.

The fix took one sentence

A can is just a cylinder.

That's it. That was the entire insight. A can isn't a hard object like a face or a pair of hands — it's one of the simplest shapes there is, with a printed label wrapped around it.

So I changed what I was asking for. Instead of "draw me pictures of this can from different angles," I asked the AI to build the can: a real 3D object that lives on the web page — like a prop in a video game — wearing a label I wrote myself, with real fonts and the real numbers.

Now every angle is right automatically. The back of the can exists for the same reason a real can has a back. The logo appears once because the label carries it once. The nutrition panel says zero calories because I typed zero calories. Made-up numbers, duplicate logos, gibberish text — not just fixed, but impossible. And when I want to change the label, I edit the text. No more slot machine.

And here's the part I want to be clear about, because it's where I would have assumed I was stuck: I never opened any 3D software, and I didn't learn 3D modeling. The AI wrote all the technical parts. The only thing that changed was my request.

The playbook

Five decisions, in the order I'd make them next time.

1. Sort every image with one question

What: does this thing need to look the same twice?

How: go down your list of shots. A splashy background appears once — that's one pile. Your product shows up in six places from six angles — that's the other pile.

Why: "matching itself" is exactly the thing AI images can't promise. This two-minute sort decides everything else.

2. Let the AI make the scenery

What: everything in the appears-once pile goes to the AI.

How: exactly what you're already doing. My site is the proof — every scene on it is AI-made, and it's the best-looking part.

Why: a background has no wrong answer. Nothing has to match it, so the AI's imagination is pure upside there.

3. For the product, ask for the object — not pictures of the object

What: have the AI build the thing instead of drawing it.

How: ask one question first: is this basically a simple shape with a printed surface? A can is a cylinder. A cereal box is a box. A gift card is a rectangle. Most products are simpler than they look. If yes, tell the AI to build it that way — it handles the technical side; your job is the ask.

Why: a built object can't drift. Every side of it exists for the same reason the front does.

4. Words get typed, never generated

What: anything written on your product — the label, the ingredients, the fine print — is real text you wrote, placed onto the object.

How: I wrote my label the way you'd write any document, and the AI wrapped it around the can like a printed label.

Why: this is what makes wrong numbers impossible instead of just unlikely. An AI paints things that look like words. Typed text says what you typed. My 170-calorie surprise is the receipt.

5. When two versions look good, pick with numbers, not vibes

What: I ended up with two good cans — the one on the site, and one made in professional 3D animation software that had genuinely prettier lighting. I put them side by side on a comparison page where you can drag a divider between them. It's still up.

How: the prettier one lost on two things I could measure. Its images were about eight times heavier, which slows the page down. And it could only turn the can in small steps — a subtle stutter — where the other version turns smoothly and lets you grab it.

Why: my site's entire sales pitch is smooth motion. The version that couldn't be smooth couldn't win, no matter how nice its reflections were. Knowing what the page is selling turned a taste call into arithmetic.

The mistake that almost shipped

One honest one, because it's the part I'd want someone to tell me.

Near the end, I spotted two bad frames and asked the AI to fix them. It fixed both, ran its checks, and told me everything passed. But the animation between those two frames still didn't match — and nothing caught it, because every check was pointed at the parts we'd just repaired.

I caught it by eye, scrolling the page the way a visitor would.

The lesson sits right next to the big one: check the whole thing, not just the part that got fixed. The AI repairs what you point at. Walking through the finished thing like a customer is the job that stays yours.

Try this tonight

Find the image you've re-generated the most — the product shot, the logo mockup, the diagram that never comes back quite the same.

Ask one question about it: is this a simple shape with a printed surface?

If it is, stop asking your AI for pictures of it. Ask it to build the thing — and keep the picture-making for the scenery, where an AI that invents things is exactly what you want.