Character consistency is the problem that separates a nice clip from a campaign, and by 2026 there are three mature ways to attack it. They are not competing solutions to one problem. They solve the same symptom at three different levels of commitment.
The three methods
Reference image conditioning
Hand the model one or more images of the person and ask it to condition on them. Setup is minutes. Every modern video model supports it natively.
The weakness is that the reference is interpreted afresh each session, and interpretation drifts. Across two or three assets in one sitting it holds. Across nine assets made over three weeks, with a model version change in the middle, it does not.
LoRA or equivalent lightweight fine-tune
Train a small adapter on a set of images so the identity is baked into the generation and never supplied alongside it. Setup is hours plus an image set somebody has picked over. Drift falls hard.
The weaknesses here are practical. The adapter is tied to a base model, so a base model upgrade means retraining. And a badly curated training set bakes in whatever was over-represented in it, usually one lighting condition and one expression, which then appears in every output regardless of the prompt.
Trained identity as a production artefact
What we mean by this is the adapter plus everything around it: a documented image sheet somebody picked over, a written description of the identity in the terms the model responds to, a set of validated prompts, a record of which base model and version it was built against, the licence and consent file, and a per-shot check sheet.
Setup is a day or more. What you get is an artefact that survives a personnel change, a model upgrade and a legal review, which the other two do not.
Compared on what matters
| REFERENCE IMAGE | LORA | TRAINED IDENTITY | |
|---|---|---|---|
| Setup time | Minutes | Hours | A day or more |
| Holds across | 2 to 3 assets, one session | 10 to 30 assets, one base model | A campaign, across upgrades |
| Survives a model version change | No | Requires retraining | Yes, documented to rebuild |
| Survives the operator leaving | No | Partly | Yes |
| Consent file attached | Rarely | Sometimes | Always, by construction |
| Cost to redo | Nothing | A few hours | A day |
| Right when | One-off, one session | Volume, one campaign | A face you will use again |
The threshold
The break-even is around nine assets. Below that, reference conditioning plus tight gating is cheaper overall even accounting for the regeneration. Above it, the regeneration cost from drift exceeds the setup cost of a trained identity, and it keeps exceeding it.
Nine is not a universal constant. It moves down when the face has to appear at close range, when the assets are produced across weeks and not in a day, or when several operators are involved. It moves up when the face is small in frame or the pieces are short.
The licensing question, which is not optional
Every one of these methods is doing something to a person’s likeness, and the paperwork differs.
- A reference image of a real person needs a release for that use, and old photography releases almost never cover generative conditioning.
- A LoRA or trained identity built from a real person needs a release that explicitly permits derivative training, names the scope of output, states territory and duration, and says what happens to the trained artefact at the end of the term.
- A wholly synthetic identity needs no personal consent but does need a model licence cleared for commercial use, plus a reverse image search on the final face with the result recorded.
The consent file exists before the first render. Trained identity as a production artefact enforces by construction, and it is the main reason we treat it as a documented artefact and never as a file on somebody’s machine.
What to do this week
- Count the assets the face has to appear in over the next quarter. Most teams undercount by looking at one brief.
- If it is under nine and inside one session, use reference conditioning and spend the saved time on gating instead.
- If it is over nine, or spread across weeks, build the identity properly and document which base model it was built against.
- Either way, get the release signed before anything renders, covering derivative training explicitly.
- Log the presenter acceptance rate per batch. It is the number that tells you when you crossed the threshold, usually about one batch after you crossed it.
The full pipeline, published with its gates, its consent requirements and the points at which a person decides.
READ THE IDENTITY LOCK PIPELINE →