I Spent a Morning Breaking Google Flow’s Image AI — The Burn-In Problem Is Real (But Not How You Think)

The Itch I Needed to Scratch

Here’s what’s been bugging me about AI image generators: they remember too much, or not enough. You start with a great concept, iterate on it five or six times, and suddenly there’s a hat in every image that you never asked for and can’t get rid of. Or you nail a brand identity in image one, iterate to put it on a product, and the brand is unrecognizable by image three. See also: AI Agent Verification.

I wanted to understand the actual mechanics. Not the marketing copy — the real behavior. So I opened Google Flow, created a fake coffee brand, and started breaking things on purpose.

The Experiment

I generated a logo for “Summit Brew Coffee Company” — mountain peak, coffee cup, black and gold. Clean, distinctive, easy to track across changes. Then I did seven sequential iterations, deliberately adding and removing elements to see what sticks.

The sequence: logo → coffee bag mockup → added a red baseball cap → removed the cap → added the cap AND a mug → moved everything to a mountain campsite → tried to remove the cap and mug after they’d been in three straight images.

Why the red cap? Because it’s visually loud, easy to spot, and completely unrelated to coffee branding. If the AI “burns in” the cap after enough iterations, it’ll be obvious.

What I Found (And Didn’t Expect)

The brand held up beautifully. Seven iterations, radical context changes, and the Summit Brew identity was recognizable every single time. That mountain-and-cup icon traveled from flat logo to kraft bag to embroidered cap to ceramic mug to outdoor campsite and never lost its core shape. Score: 9/10 on brand persistence.

But here’s the part that surprised me: the burn-in threshold is higher than I expected. The red cap appeared in three consecutive iterations — and I still removed it cleanly with one well-written prompt. The mug too. Gone. Like they were never there.

The trick? Don’t just say what you don’t want. Say what you DO want. “Remove the cap” is weaker than “Show ONLY the coffee bag by itself on the table, no hat, no cap, no mug anywhere in the scene.” The model responds to positive specificity better than negative commands alone.

The Isolation Discovery

The finding that actually made me sit up: iteration chains are completely isolated. I’d been worrying that if I iterated a logo through seven rounds of caps and mugs, that “memory” would infect everything else I did in the project. Nope.

I went back to the project grid, clicked a different logo variant, and started a fresh chain. Asked for a city billboard. Got a pristine billboard with zero contamination from the cap-and-mug chain. No red hats lurking. No mugs ghosting in. Clean slate.

That means the project is just a folder. Each image’s edit history is its own isolated context. You can experiment wildly on one chain without poisoning your other work.

The Rule I’m Using Now

I call it the 3-5 Rule: keep any iteration chain to 3-5 steps for maximum control. At that depth, you can add things, remove things, change scenes, and the AI cooperates. Beyond 5, you’re accumulating context that starts influencing outputs in ways you didn’t ask for.

When you hit 5 iterations, save your best image. Start a new chain from it. Fresh context, same great image as your starting point. This is the equivalent of “Save As” in Photoshop — you’re preserving your work while clearing the undo history.

For anyone building brand assets, personal content, or product imagery: the AI isn’t as stubborn as you think. You just need to understand the architecture. Short chains. Fresh branches. Explicit prompts. That’s the whole game.

Frequently Asked Questions