A hallucination is what happens when an AI image model generates something that's factually wrong but visually confident. In fashion and product photography, that usually looks like a logo duplicated onto a sleeve that shouldn't have one, a hand with six fingers, a zipper that doesn't actually close, or a print pattern that drifts from the reference garment within a single generation. The model doesn't flag any of it as uncertain. It renders the wrong detail with exactly the same confidence as the right one.
The term comes from language models, where it originally described an AI generating plausible-sounding text that's factually false. Image models hallucinate in the visual equivalent: plausible-looking pixels that don't correspond to reality.
Why it happens
Diffusion-based image models don't "know" facts the way a database does. They generate images by predicting what a coherent picture should look like based on patterns learned from training data, then refining that guess step by step. Most of the time that produces something that looks right. Occasionally the model's best guess for an ambiguous region, an overlapping fold of fabric, a logo half-obscured by a model's arm, a hand mid-gesture, lands on something structurally wrong. There's no internal checkpoint that catches this before the image is returned, because the model isn't checking against ground truth. It's generating the statistically likely next pixel, not verifying a fact.
Hallucination versus ordinary creative variance
Not every unexpected result is a hallucination. Ask a model to generate a product shot and you'll reasonably get some variation between runs: a slightly different pose, a fold falling differently, a marginally different angle. That's normal generation variance, the same way two photographers shooting the same brief won't produce identical frames. A hallucination is specifically when the output breaks physical or logical plausibility, a garment with a seam that goes nowhere, a hand with the wrong number of fingers, a logo that's mirrored or duplicated, not when it's simply different from what you expected.
Why it's the actual hard problem in AI photography
Getting an AI model to generate a photorealistic garment on a photorealistic model was the frontier problem a few years ago. That part now works reliably. What's left, and what actually determines whether an image is publishable, is catching and fixing the hallucinations: the duplicated logo, the malformed hand, the fabric texture that's subtly wrong in one specific area of an otherwise perfect shot. Ninety percent of a generation can be flawless, and the remaining ten percent, one hallucinated detail, can be the difference between an image you can ship and one you can't.
How it gets caught
There's no way to prevent hallucinations outright with current models; the mitigation is detection and correction, not prevention. That means automated quality scoring that flags likely problem regions, human or AI-assisted review gates before an image gets marked ready to export, and targeted spot correction tools that fix the hallucinated region without regenerating the whole image, and risking a new hallucination somewhere else. Brandmachine's Agent scores every output for exactly this kind of failure and routes anything below threshold back for a fix before it reaches a review queue, rather than leaving someone to spot the sixth finger by eye.