Television casts a lead once and protects the choice for seven seasons. A long AI series asks the same discipline from one person with a prompt box: cast deliberately, then defend the casting against every pressure to improvise.
Cast on paper first
Before frame one, fix the muse in writing: name, age, hair, two facial anchors, build, one signature detail — the necklace, the beauty mark, the way the hair is always parted. This identity paragraph opens every prompt in the series, pasted unchanged. You will get bored of reading it. That boredom is the sound of consistency working.
The reference set
Frames one to three exist to be checked against: the canonical portrait, the signature look, her home setting. Every later frame is judged beside them before it joins the series. Frames that fail the check are not “close enough” — they are recasts waiting at the face swap bench, where one job restores the canonical face onto a frame whose body, clothes and room were already right.
Repair at the right bench
Identity maintenance has a division of labour. Face drifted: face swap from the canonical portrait. Prop changed colour, necklace vanished: the editor, at a third of a premium re-render. Composition wrong: only then the generator, with the brief clause that failed named and fixed. Creators who re-render for every defect pay generator prices for editor problems — and introduce fresh drift with every rescue.
Retire nothing silently
Keep a one-line log per frame: number, variable changed, repair applied. When a series ends, the log is the casting bible for the next one. Leads get recast between projects all the time — but here, as on television, it should be a decision, never an accident.