Direct answer first: you cannot keep AI characters consistent with prompt text alone - you keep them consistent by generating from reference images every single time. The working method is a casting sheet: a multi-angle reference set per character, generated once, then attached as image input to every keyframe that character appears in. Describing the character again in words ("red curly hair, green vest...") drifts within ten shots; a locked reference set holds for a whole season.
The casting-sheet method
- Generate a 3x3 multishot, not nine separate images. One generation call that asks for "a 3x3 sheet of the same person - front, back, side, three-quarter, face close-up, expressions - physically consistent" produces angles that actually match, because they were made together. Nine separate generations produce nine cousins.
- Split the grid and store the cells. Each cell becomes a labeled reference (front, profile, expression) you can attach per shot.
- Attach references on every keyframe. Every image generation that includes the character takes the sheet cells as image input plus a short identity descriptor. The reference does the work; the text only sets pose and scene.
Where it still breaks - and the fix
- Wardrobe changes. The single biggest drift source is describing outfit changes in prompt text. Do not. Make a named look variant sheet per state - "ep1-base", "injured", "rain-soaked" - and bind each scene to a variant. We learned this the hard way when a character's outfit silently changed mid-episode; the fix was banning prompt-text wardrobe and regenerating from variant sheets.
- Special physical states. Hanging upside down, soaked, wounded - models fight you if you prompt the state. Bake the state into a new reference sheet instead, then generate from that. Change the input, not the instruction.
- Video generation. Image-to-video inherits the keyframe's face, so consistency is won or lost at the keyframe stage. Get the keyframe right from references and the video follows.
Does it hold at production scale?
This is the method behind every series on our watch catalog - 30+ short dramas, anime, and webtoons made by one person, including a full reproduction where 288 shots were re-generated against the same casting sheets and cut back into the same episodes. Reference-locked generation is also why a 10-episode thriller can survive a mid-season model swap: the sheets stay, the model changes.
In Owntent the whole flow is built in - sheet generation, per-scene look binding, and automatic reference attachment on every keyframe - but the method itself works in any tool that accepts image references. If you are assembling it by hand: one multishot per character, split, label, attach every time, and never describe clothes in text.
What we could not verify
We have not benchmarked cross-model identity retention numerically (same sheet, different video models, measured face similarity) - our evidence is production-level: full seasons shipping without visible identity drift. A measured benchmark is on our list.