Define who exists in this scene. Write the sheet once, in full, and it gets pasted verbatim into every shot that character appears in. If your video tool supports reference images, note the image filename here so every generation uses the same one.
Environment Lock
One description of the location: lighting, time of day, colors, materials. It repeats unchanged in every prompt so the world never drifts between shots.
EMPTY
Camera Lock
The look of the film: lens, depth of field, lighting style, texture, color treatment. Exact same wording every shot, so the whole sequence feels shot on one camera.
EMPTY
Beat Sequence
Plan every shot before generating anything. Each beat is one shot: what happens, how it is framed, who is in it, and what source motion drives it. Keep beats connected so the model preserves identity and spatial continuity.
The Method
Consistent AI video comes from references, planning, and repeated rules, not from a newer model. Direct a scene instead of generating isolated clips.
1
Build a full character sheet first
Front, side, back, facial expressions, wardrobe details, proportions, eyes, hair, distinguishing features. Reference the exact sheet in every shot.
2
Lock the environment
One detailed description of location, lighting, time of day, colors, materials. Paste it unchanged into every prompt.
3
Lock the camera style
Lens, depth of field, lighting style, film texture, color treatment. Repeat the exact wording in every shot.
4
Separate identity from motion
The character sheet defines who is moving. A depth map of source footage defines how they move. The original footage never influences face, clothing, or environment.
5
Plan the entire sequence before generating
Write the storyboard with every shot, action, framing change, and payoff before making any video.
6
Generate connected beats whenever possible
Create the scene as one continuous sequence instead of isolated shots. This preserves identity, clothing, timing, and spatial continuity.
7
Reduce the decisions left to the model
Lock every important variable so the model is mainly deciding performance and movement. No rewriting descriptions hoping for luck.