Why more words do not help
The instinct when a face drifts between shots is to describe it harder. Add the eye colour. Add the jaw. Add "consistent character, same face, identical features". The result is a longer prompt and a face that is differently wrong.
This happens because a text prompt is a description of a distribution, not a reference to an instance. "Green-eyed woman, mid-thirties, sharp jaw, dark hair" narrows the set of faces the model will produce. It does not pick one and hold it, because there is nothing in a text string that can carry the identity of a specific face. Every additional adjective narrows the set a little more and costs you something else — prompt weight that was doing work elsewhere, usually the composition.
Past a certain length, extra description actively hurts: the tokens compete, the model averages, and you get a face that satisfies the adjectives and resembles nobody. This is the point most people reach before concluding the technology cannot do continuity.
What actually holds
In rough order of how much consistency they buy per unit of effort.
One reference image, reused as an input
Not described — passed. An image reference carries identity information that no amount of text can. If you take one thing from this post, it is that the reference should be an input to every shot rather than a thing you looked at while writing the prompt.
Lock the reference before you like it
The expensive mistake is generating twelve shots, deciding on shot nine that the face is finally right, and having eleven shots that predate the decision. Pick a face early, accept it is only the fourth-best face you saw, and get the consistency. A merely good face that appears in every shot beats an excellent face that appears in one.
Separate identity from everything else
The character prompt says who they are and nothing about where they are, what the light is doing, or what lens you are on. Once those are mixed into one string, changing the lighting perturbs the face, because you have given the model no way to know which words were load-bearing.
Hold the seed while you vary one thing
Changing the seed and the prompt together tells you nothing about which change did what. Vary one. This is slower per attempt and much faster per solved problem.
Fix the face in the frame it is worst in
Drift is not uniform. It concentrates in extreme angles, small head sizes and strong side light. Solve the hardest frame and the easy ones tend to come along; solve the easy ones and you learn nothing about the hard one.
What is realistic
Worth being plain about the ceiling, because the gap between demo and production is where people lose weeks.
A close and a medium of the same character, same lighting, adjacent angles: reliably consistent with a reference input, and this is the case demos are made of. A full sequence across different locations, lighting conditions and shot sizes: consistent enough to read as one person, with drift that a viewer will not notice and you will. Extreme angles, distant figures, complex occlusion, hands: still unreliable, and no prompt fixes it today.
The working assumption that survives contact with a deadline is that continuity is a budget, not a switch. You get a certain amount of it, you spend it where the audience is looking, and you design around the shots where you know it will fail — which is a familiar discipline to anyone who has worked around a difficult location or an actor who was only available on Tuesday.
Questions we get
Should I train a LoRA for my character?
If the character appears across a long project and you have twenty or more good images of them, a trained model is the strongest tool available and worth the setup. For a single sequence it is usually not worth it — you will spend longer on the training loop than on the shots, and a reference input gets you most of the way. The threshold in practice is whether the character outlives the project.
Why does my character change when I only changed the lighting?
Almost always because identity and lighting are in the same prompt, so the model has no way to know which words it is allowed to move. Split them: identity in the reference and the character node, lighting in the shot. If they are already split and it still drifts, the reference is probably too weakly weighted relative to the scene description.
Does a higher resolution improve consistency?
No, and it is an expensive place to look. Resolution changes how much detail is rendered, not which face is chosen. Generate at working resolution while you are solving identity, and upscale once it holds.
How many reference images should I use?
One strong frontal reference beats three inconsistent ones. If you use several, they must be the same person under similar conditions — mixing references from different generations reintroduces exactly the averaging problem you are trying to escape.