When generated images drift, a story world loses trust. The same character changes age, a familiar street becomes a different climate, or a symbolic prop disappears. The fix is not simply to make prompts longer. It is to keep a visual record connected to the narrative.
Build a Context Card Before You Generate
For every recurring character or place, maintain a short card with three groups of detail:
| Fixed | Variable | Narrative |
|---|---|---|
| features, clothing logic, architecture, era, props | weather, time, pose, expression, temporary damage | what the image reveals and what changes in the scene |
Fixed details protect recognition. Variable details keep the world alive. Narrative details stop the image from becoming a static character sheet.
Describe What the Shot Must Do
Do not begin with style. Start with the moment. Is the character powerless, concealing a decision, or seeing the cost of an earlier choice? Then select a camera position that makes that beat legible.
A high angle can reduce a character’s apparent power; a close-up can make a concealed reaction unavoidable. The choice should come from the scene, not from a random collection of fashionable visual terms.
Run a Continuity Check
Before accepting a new image, compare it with the context card:
- Does the character still read as the same person?
- Does the location still obey its established geography and social character?
- Does the lighting serve the current event rather than merely looking dramatic?
- Has the image introduced a new prop, rule or relationship that needs approval?
If an image creates a compelling new idea, treat it as a proposal. Decide whether to add it to canon before repeating it in later work. This makes visual exploration a source of discovery without allowing accidental world changes.
Connect Assets to Scenes
Name or catalogue the asset by scene and purpose, not only by visual style. “Harbour confrontation—Mara realises the relay is gone” is more useful than “purple city 07.” The asset then remains attached to the beat it was made to support, making later editing and video planning much easier.