Why describing the character is not enough
Describe a person as precisely as you can manage: "woman in her thirties, shoulder-length dark brown hair, glasses, freckles". Generate ten images with it. You will get ten different women, every one of them a fair match for the description. The problem is not that the model is not paying attention — it is that a description picks out a category, not a person.
What makes a face that particular face is a set of fine relationships that words can barely carry: the distance between the eyes, the ratio of nose to mouth, the curve of the jawline. We have no vocabulary for any of it. That is why a longer description does not help, and why giving the character a name does not help either — unless you happen to land on a celebrity who appears often in the training data, and at that point a different problem starts.
Today's answer: the reference image
The breakthrough was that models began to accept images as input, not just text. You upload one or a few pictures of your character, the model takes the face from those, and the prompt says what should happen: a different pose, a different setting, different clothes.
Two years ago this needed separate training; today it is a standard feature on every major image model. The practical difference is enormous: no training, instant results, and even a single uploaded picture gives you something usable. There are real differences between reference types, though — style, character and composition do not travel by the same mechanism — and the kinds of reference image sorts that out in detail.
How many references a model will take
Every provider caps how many images it accepts at once, and those numbers tell you more about the technology than the marketing pages do.
- The highest-capacity models handle six objects, five characters and three style references at the same time.
- The faster, cheaper variants typically allow ten objects but only four characters — so the count is not a quality ranking, and characters are everywhere the tighter limit.
- In the open-weight families ten images is the ceiling.
- One large provider recently raised its limit from three to five source images — a good sign that the ground is still moving.
Where it breaks down: more than one character
There is a phenomenon with its own name in the literature: when several characters share an image, the model blends them into each other. A measurement from March 2026 put numbers on it using research models, and the curve is steep.
- With two characters, some pair of them merges in close to half of all attempts.
- With four characters it is already above four in five.
- With ten characters it happens essentially every time.
- It does not depend on how alike they look: characters of different gender and age blend too.
These figures come from research models, not from commercial services, so do not read them as a score for the tool on your desk. The lesson carries over regardless: one character per image. If two people have to share a scene, it comes together more reliably by compositing afterwards than by generating it in one go.
A product is not a character
Product consistency looks like the same job and is, in practice, both easier and stricter. Easier, because an object has no face: the model does not "misrecognise" a box the way it misrecognises a person. Stricter, because you spot the error instantly — if the logo slips on the packaging or the shade of the colour shifts, that is not a style choice, it is a false picture of a product.
So with products the method that works is not full generation but completing a real photograph: you shoot the product, and AI supplies the background and the setting. That solves fidelity and legal exposure in one move. The whole workflow, together with the mistakes people make, is laid out in product photography for a webshop, step by step.
When training is still worth it
Reference images do not solve everything. If you need the same character a dozen times a week, if you work in a genuinely idiosyncratic style, or if your provider's reference handling is not precise enough for your case, then light fine-tuning is still on the table.
The numbers are friendlier than most people remember: nine to fifty images is typically enough, and at one provider the cost is billed per step — a thousand steps comes to six or seven dollars, give or take. That is not an enterprise line item; a small business can build its own character model in an afternoon. The real question is whether it pays off: if you produce fewer than twenty images a month with it, it probably does not.
Common misconceptions
- "More reference images will make it better." Up to a point, yes; after that, no. Near the provider's stated ceiling the result often gets worse, because the images contradict each other.
- "A style reference carries the character too." It does not. A style reference carries the palette and the look; characters and objects specifically do not travel with it — one provider spells this out in its own documentation.
- "The reference image only affects the face." Unfortunately not: if the background or the clothing is visible in the reference, some of that leaks through as well. The literature has a separate name for the effect.
- "Once it settles, it stays settled." Over a long series there is a slow drift: the twentieth image is noticeably different from the first. Always go back to the same original reference, never to the previous image.
What to take away
Character and product consistency is solved today with reference images and no training at all — one of the biggest practical changes of the last two years. The limit shows up as soon as several characters share one frame: from there, break the scene into pieces. And only reach for training when the sheer volume actually justifies it.