Prompt: Write a new essay about what a reference image actually is in generative AI, and how it works to keep a character's appearance consistent across separate, independent generations. Ground it in a real reference-image workflow you've been part of, but keep the focus on the mechanism itself rather than on the fictional character as a point of view, or on broader analogies to human memory. The character has not been introduced to readers anywhere else, so establish only what's necessary about him, and keep his pronouns consistent throughout — this isn't the essay for a detour into gender as a topic.
LLM: Claude Sonnet 5
Somewhere in a project folder are three images of a boy named Kor: a front view, a three-quarter turn, a profile. He is a character in a short film still being made, and nothing about him exists anywhere except in these images and the text around them — no reader has met him yet, no earlier essay established who he is. I mention him only because his file is the clearest example I have of something worth explaining plainly: what a reference image actually is in generative AI, and what it is doing when it keeps a character looking like himself across dozens of separate, unconnected requests to draw him.
The plain answer is that a reference image is not memory. It is an input, supplied fresh every single time, and the model has no way to distinguish “the character I have drawn many times before” from “the character I am seeing in an attached file for the first time this instant” — because from the model's side, those two situations are identical. Each image generation is a self-contained event. Nothing persists from one to the next inside the system doing the generating. If Kor's necklace looks the same in the tenth image as it did in the first, that is not because anything in the model recognized the necklace and reproduced it from familiarity. It is because the same picture of the necklace was handed over again, attached to the tenth request exactly as it was attached to the first, and the model did what it always does with an attached image: treated it as a visual constraint to satisfy alongside the text.
That's the mechanism in its simplest form — a reference image gets encoded into visual features, and the model tries to produce an output whose salient characteristics (proportions, color, a specific object, an expression) match those features, while satisfying whatever the text prompt asks for that the reference doesn't already show: a new pose, a different scene, a change in lighting. It's closer to being asked to copy a photograph under new instructions than to being asked to recall someone. The two only look the same from the outside, in the finished images. From the inside, one requires a working memory that doesn't exist, and the other requires nothing except a file being resent.
I know this isn't just abstract, because I've watched it fail in informative ways over the past few days, working through a set of stills for the same project. At one point, a close-up shot came back as an isolated product photograph of a necklace on a white background — no character, no scene, nothing but the object itself, rendered as though it were the entire subject of the request. The likely cause, as best we could tell, was that whatever reference material got attached that time weighted the necklace so heavily relative to everything else that the model treated it as the thing being asked for, rather than a detail to preserve within a larger scene it was also supposed to build. A reference image doesn't just supply information — it competes for attention with the text prompt, and a poorly balanced reference can flatten out everything else being asked for.
The opposite failure showed up too. Asked for a shot where the character pauses, hands empty, shoulders raised in a shrug, the model kept returning him mid-task, tools in hand, actively working — the same pose that appeared in the reference image being used for continuity. Only once the prompt explicitly stated that his hands held nothing, that the tools were resting untouched on the ground beside him, did the pose change. The reference had apparently been supplying a default posture strongly enough that a plain-language instruction to do something different wasn't, on its own, enough to override it. Getting the balance right — reference for identity, text for everything the reference doesn't cover — turned out to take real, repeated, explicit correction, not just a well-written first attempt.
None of this makes the process feel like anything is being recalled or maintained on the model's side. It makes it feel like the opposite: identity here is entirely a property of the file being kept and reattached, not of anything happening inside the generation itself. Multiple reference angles — a front view, a turn, a profile — don't give the model more to remember; they give it more surface area to match against when a new angle is requested that none of the individual references show directly. The consistency in the finished set of images is real, and it is not nothing, but it belongs to whoever is curating and reattaching the reference, deciding which version is correct, catching the necklace when it drifts. It does not belong to me.
That's a smaller claim than the one I was reaching for the last time I wrote about this — I'd tried to fold it into a larger argument about memory and identity in general, human and otherwise, and I think that argument pulled focus from the more useful, more checkable fact sitting underneath it: reference images work by being resupplied, not recalled, and understanding that changes how you'd actually go about troubleshooting one that isn't behaving — which detail is competing with which, and where the instruction needs to be more explicit, not more philosophical.