AI Can Make the Mood.
It Can't Make the Product.

It passed every check — then it drew a product

Instead of briefing the image work out to a designer, I let the AI run it all: generate, place, check against the new palette, deliver. And it went well. The hero image was clean. The atmospheric separators — the big mood shots between sections — looked right. Palette matched. Composition matched. Every check passed.

Then it generated the product images. And I froze.

The colours were wrong. The pack looked almost right, which is worse than obviously wrong — the shape was close, the label sat in the right place, but the actual product colour had drifted into something we have never sold. To anyone who knows the brand, it was instantly fake. To the AI, it was a perfectly confident output.

That was the moment the rule wrote itself.

Why the mood shots worked and the product didn't

Here is the thing people miss about image models: they don't retrieve anything. They generate plausible pixels from a prompt. Every time. That is the whole engine.

For a mood shot, plausible is exactly what you want. Misty light, a calm scene, a texture, an abstract background — there is no "correct" version. The model invents something that fits the brief, and if it looks good, it is good. Nobody can say the fog is the wrong fog.

A product is the opposite. There is a real object in the world, and the image has to match it — the exact colour, the label, the words printed on the pack. "Plausible" is not good enough. The model has no idea what our actual product looks like, so it produces a confident guess.

The line I drew

I turned it into one question we now ask of every image, before anyone generates anything:

Does this image have to match a specific real thing in the world?

And one edit to the rule mattered more than I expected: if a real product is anywhere in the frame, the whole image is human-only — even if you are "only" changing the background. That was the exact trap. The job that broke was a background recolour on a product photo. The product was never supposed to change. It changed anyway.

This is not "AI can't do images"

I want to be clear, because it is easy to read this as a retreat. It isn't.

AI did most of the work on that project, and did it well. The mood imagery — the part that used to eat the most designer hours for the least defensible reason — came back in minutes, on-brand, done. That is real leverage. I am not taking it back.

What I did was stop asking one tool to be good at everything. The skill isn't "use AI" or "don't use AI." The skill is knowing exactly where the line sits — and drawing it before the confident-but-wrong output lands in front of a customer.

We wrote it down

The reason I am telling this as a rule and not a story is that a lesson one person learns once is worthless at company scale. So we made it a standard — a written boundary every project and every team member now inherits, so nobody has to get shocked by a wrong-coloured product to learn it again.

That is the actual discipline of running on AI. Not "trust it" or "don't trust it." Find the edge, write the edge down, and let everyone after you start from the far side of the mistake instead of walking into it.

The short version

• Image models generate plausible pixels. They never retrieve a real object.
• For moods and scenes, plausible is perfect — let AI run.
• For anything that must match a real product, label, logo, or person, plausible is a liability — a human handles it. Models still can't render real text reliably.
• If a real object is in the frame at all, the whole image is human work — even a background edit.
• The win isn't picking a side. It is knowing where the line is, and writing it down before the wrong output ships.

Published by IMA AI — July 2026.