What it conditions, and what it does not
Most of the difficulty in producing more than one shot of the same thing comes from a single fact: each generation starts fresh. Describe a character in words, run the prompt twice, and you get two different people who both match the description. A character reference exists to break that, by supplying pixels instead of adjectives.
What it constrains is identity: face structure, hair, distinguishing marks, often clothing and colour. What it deliberately does not constrain is the shot. Pose, framing, camera angle, lighting and environment stay under the prompt's control, which is the whole point. If the reference dictated composition too, you could not put the same person in a new scene.
This division is also where most failures come from. Hand a model a stylised painting as a character reference and it will carry the brushwork across, because it cannot tell which properties of the image you meant. Hand it a photograph shot in hard side light and the shadow becomes part of the identity.
Choosing reference images
The rules are unglamorous and they matter more than prompt wording.
- Neutral, even light. Flat frontal or soft window light. Shadow gets baked in as a feature.
- Multiple angles. Front, three-quarter, profile. Three to five images beats one, because the model has genuine information about the sides of a head rather than a guess.
- Consistent subject across the set. Same haircut, same age, same wardrobe if wardrobe matters. A reference set that disagrees with itself averages out into a fourth person.
- Plain background. Other faces in frame are the single most reliable way to get a hybrid.
- Resolution and sharpness. Reference detail is the ceiling on output detail.
For a recurring subject, build the reference set once, keep it in the project, and reuse it for every shot. Consistency across a sequence comes from the reference set being fixed, not from any individual generation being good.
Prompting alongside a reference
The prompt's job changes when a reference is attached. Stop describing the subject and start describing everything else.
Write the scene, the framing and the light in full, then name the reference explicitly rather than redescribing its contents: the woman from the reference image, standing at a bus stop in the rain at night. Adding dark hair, green eyes, mid thirties on top of a reference does not reinforce the identity. It introduces a second, weaker description that competes with the pixels and pulls the face toward the average of both.
Then pin the attributes that must survive: keep her face, hair length and jacket exactly as in the reference. Models treat clothing as far more negotiable than faces, so wardrobe continuity usually needs to be named.
Where you can, keep the subject at roughly the scale it occupies in the reference. Identity holds well from medium shot to medium close-up and degrades quickly in wide shots, where a face is a few dozen pixels and there is nothing to condition.
When to move to a LoRA instead
A character reference is the right tool for a handful of shots, a pitch, or a one-off. It costs nothing, needs no training, and works immediately.
It has a ceiling. Identity is suggested rather than learned, so it loosens under unusual poses, profile turns the reference did not cover, and long video clips. If the same character appears across dozens of shots, training a LoRA on twenty to thirty images holds the likeness through angles and expressions no reference set can cover, and removes the per-shot drift entirely. The practical trigger for switching is repetition: once you are regenerating shots because the face moved, the training run is cheaper than the re-rolls.