Image models make every picture from scratch. There is no "memory" in them of what the face, the logo or the packaging looked like in the previous image — which is why the same prompt run twice gives you two different people. Character consistency is the work of pushing that scatter down.
In 2026 the answer is no longer training but a reference image: the model is handed the original picture and lines the next one up with it. It works for faces and for products, but not without limits — the further the prompt takes you from the original setup (a different angle, different light, a different age), the faster the likeness drifts.
The most common misconception is that this is a prompting mistake. It is not: the scatter is built into how the model works, and the best prompt in the world will not remove it entirely. Wherever being recognisable is worth money — a brand, a mascot, a series of product shots — a human still has to pick the keepers at the end.
The practical methods and where they stop working are gone through in Keeping a Character or Product Consistent, one by one. Directly related terms: reference image, seed and style reference.