Answer
How do you keep a character consistent across a long video?
Re-describing a character in every prompt does not hold them. What holds one across dozens of shots is reference conditioning plus frame chaining — explained, including when a LoRA is the better answer.
Re-describing a character in each prompt will not hold them, because a text description is not an identity — two generations from identical words produce two different people. What holds a character across a long video is conditioning every shot on the same locked reference images rather than on words, and opening each shot on the previous shot's final frame so wardrobe, lighting and position carry forward. Unlimited Studio locks a character from eight reference angles, picks the angle that matches each shot, and scores the result against the reference before delivering it. If you need one specific person across hundreds of generations over months and can supply twenty-plus images of them, training a LoRA is still the stronger answer; reference conditioning is better when you have one or two images, or no real person at all.
Detail
Why prompts alone cannot do it
A prompt describes a type, not a person. "A woman in her thirties with dark hair" names a space of millions of faces, and each generation samples from it independently. Adding adjectives narrows the space but never collapses it to one point, which is why the face drifts no matter how precise the words get.
The fix is to stop describing and start conditioning: give the model the actual pixels of the person and ask it to place that person in a new scene.
Detail
What eight angles buy you
One front-facing photo holds a character until the shot needs them in profile, at which point the model has to invent the side of a head it has never seen — and invents a different one each time. A reference set covering the main angles means whichever way a shot turns, there is a real reference for it.
Each generated result is then scored against the reference set, and a result that has drifted is regenerated with stronger conditioning rather than delivered. That check is the difference between a system that usually holds and one you can plan a film around.
Detail
When to use something else
If you have twenty or more images of one real person and will generate them hundreds of times over months, train a LoRA. It bakes the identity into weights and holds tighter than any conditioning approach, and the training cost amortises.
Reference conditioning wins when you have one or two images, when the character does not exist, or when you need a new character working within minutes rather than after a training run. It is also the only one of the two that extends to video chaining without retraining per character.
More of these: all answers, or the longer explainer guides.