The AI kept inventing football crests. The fix was a sentence.
Eighteen AI-generated ad images, eleven redos: hallucinated trademarks, drifting product geometry, and the QA ruleset that made an image model production-ready.
I needed eighteen ad images for a consumer brand I work with — six motifs, three formats each, for a seasonal campaign. A few years ago that's a brief, a shoot, a retoucher, and a couple of weeks. Last week it was one afternoon with Higgsfield, an AI image platform, for less than the cost of a stock photo subscription.
That's the part everyone already knows. The part worth writing down is that eleven of those eighteen images had to be redone, and not one of the redos was the tool's fault in the way you'd expect.
The model kept inventing football crests
The campaign has a football theme, so an early prompt asked for a person in a "fan jersey." The model delivered — including a beautifully rendered emblem on the chest that looked uncomfortably close to a national federation's official crest. Twice.
That's not an aesthetic note; that's a trademark problem that could cost real money if it ships.
The fix was a sentence: "plain solid-colour t-shirt, no badge, no crest." The model never did it again. But you have to know that a hallucinated crest is a legal risk and not set dressing, and you have to catch it at review. Nielsen Norman Group's piece on AI hallucinations makes the underlying point well: these aren't bugs, they're an artifact of how generative models work — plausible output, confidently delivered, with no concept of which details are non-negotiable.
The product kept drifting into a generic version of itself
The second failure was subtler and more dangerous. The product has a distinctive physical geometry — it's the brand, mechanically speaking. Generated from text alone, the model would drift it toward the generic category look within a couple of variations. Plausible furniture, wrong product.
The fix wasn't prompting harder. It was switching to image-to-image with the real product photography as an anchor reference in every single generation, plus a reference person for continuity across motifs. Text describes a category; a reference image pins an identity. Academic reviews of AI in the creative industries keep landing on the same conclusion — human oversight and curation remain essential in production creative work — and this is the concrete shape that oversight takes: knowing which asset is brand IP and refusing to let the model improvise on it.
The rules that came out of the afternoon
By the end, the feedback loop had condensed into a short standing ruleset: the product positioned the way it's actually used, never floating in the middle of a room. People oriented toward what they're supposed to be watching. Evening mood, but a bright, friendly interior — dark scenes don't sell comfort. Logical product proportions. No invented emblems, ever.
None of these rules are in any prompt guide. Every one of them came from twenty years of looking at ad creative and knowing what makes a buyer trust an image — and each redo cost pennies instead of a reshoot day.
The model gives you a plausible scene. The brand lives in the details the model doesn't know are non-negotiable.
This is the creative-department version of the number that was wrong in the most convincing way. The failure mode is identical: output that looks finished is not the same as output that is correct. With data you verify against the source system; with creative you verify against a QA ruleset that exists only in an operator's head until someone writes it down.
So that's what I did — the rules now live in the project's brain and apply to every future generation automatically, the same pattern as the anti-hallucination audit. The tool gets cheaper and faster every quarter. The ruleset is the part that compounds, and it's the part a novice with the same subscription doesn't have. That asymmetry is the whole game now.
Eighteen images shipped. The reshoot took an afternoon. The art director was the leverage.
Sources & further reading
External
Nielsen Norman Group — AI hallucinations: what designers need to know · arXiv — Advances in artificial intelligence: a review for the creative industries
Related posts
The AI handed me a number that was wrong in the most convincing way · The anti-hallucination audit · The death of generic AI