World Bibles: Locking Environments So Sets Don't Drift
A character bible locks who's in frame. A world bible locks where they are — without it, your set redecorates itself shot to shot.
The set has an identity too
If you've worked through The Character Bible, you already know the failure mode by heart: generate a character across a dozen shots without a locked reference, and by shot six the jawline has shifted, the jacket has changed color, and the actor is quietly becoming someone else. The fix was a master reference sheet — one artifact, generated once, referenced everywhere.
The set has the exact same problem, and almost nobody builds the exact same fix for it.
Watch any AI-generated action sequence closely enough and you'll see it: a warehouse that has three doors in shot 2 and five in shot 5. A rooftop ledge that's on the north side of frame in one shot and the south side in the next, even though the character never turned around. A sky that's blown-out midday in one shot and moody overcast two shots later, with no explanation for the jump cut in weather. None of this is a rendering bug. It's the same drift that eats character identity, just happening to the environment instead of the face.
The world bible is the companion fix. It does for the location what the character bible does for the person in front of the camera: it turns "wherever the model feels like putting things this time" into "exactly what we locked in pre-production."
What actually goes in a world bible
A world bible is not a single pretty establishing shot. It's a working reference document with six specific components, and skipping any one of them reopens the exact drift problem it exists to close.
Master environment reference sheet. Multiple establishing shots of the location from different vantage points — not just the hero angle you'll open the scene on, but enough angles that when a later shot needs a view you didn't originally plan for, there's already a locked reference for what that side of the space looks like. One angle only answers "what does this location look like from here." Multiple angles answer "what does this location look like, period."
Top-down movement map. This is the piece that solves spatial continuity, and it's the one most productions skip because it doesn't look like a "cool" reference image. It's a literal bird's-eye-view floor plan of the space: where does a character enter and exit relative to fixed landmarks, which direction counts as "deeper into the location" versus "toward the exit." Without this, every shot's blocking is guesswork — a character who ran left in shot 3 can end up running right in shot 5 because nothing defined which direction was which. The movement map is the ground truth the storyboard's shot-to-shot geography gets checked against.
Key structures and landmarks, called out explicitly. Not "a warehouse" — the loading dock on the east wall, the catwalk above the north entrance, the specific stack of shipping containers that will read as visually distinct across multiple angles. Landmarks are what let a viewer (and the generation model) orient inside a space across cuts. A location with no named landmarks is a location the model has to reinvent every time it's asked to show a new angle of it.
Ground and damage-state references. This is the one production teams forget until it bites them. Does the ground stay pristine through an action sequence, or does it visibly accumulate damage — cracked concrete, scattered debris, scorch marks — as the sequence progresses? That's a creative decision, and it has to be made once, locked, and referenced consistently. The alternative is a floor that silently resets to pristine between shots for no narrative reason, which reads as a continuity error the instant a viewer notices it.
Sky and weather progression, if the scene spans time. A scene that opens at midday and closes in a storm needs a reference for what the sky looks like at each stage of that arc — not just the two endpoints, but enough intermediate states that shots generated independently (in different batches, on different days) land on the same trajectory instead of jumping erratically between weather states.
Material and texture callouts. Any surface that's going to be seen up close — a rusted railing, a specific concrete texture, a particular metal panel — gets an explicit reference. Wide shots can get away with generic materials. Close shots can't; the model needs to know exactly what that surface looks like at proximity or it will invent a different texture every time.
Why this is a pre-production investment, not a per-shot task
The character bible and the world bible share the same economic logic. Both are generated once, during pre-production, before a single scene is shot — and both are referenced, not regenerated, for the entire life of the production.
This matters because the temptation runs the other way. It's faster, in the moment, to just describe the location fresh in each shot's prompt and trust the model to keep it consistent. It never does — not because the model is bad, but because prose descriptions are lossy. "A dim warehouse with shipping containers" describes a category of location, not a specific one. Ask for that same prose ten times and you'll get ten different warehouses, each internally plausible and none of them the same building.
A locked world bible removes that ambiguity at the source. Every shot's prompt references the same reference sheet, the same movement map, the same damage-state decision — so instead of ten different warehouses, you get one warehouse, shot from ten different angles, in exactly the damage state the sequence called for at that point.
The three pillars of production continuity
The world bible doesn't work alone. It's one of three companion pre-production investments that together lock a production's entire visual continuity — and it's worth seeing how they divide the labor, because each one covers a gap the others don't.
The character bible locks who — identity, wardrobe, proportions, expressions, so the person in frame stays the same person across every shot.
The world bible locks where — architecture, landmarks, ground state, sky state, so the location stays the same location across every shot.
Continuity Law — the shot-to-shot chaining discipline where each new shot's start frame is anchored to the prior shot's final frame — locks the transition between shots, so the cut from one frame to the next doesn't introduce a jump nothing explains.
Each pillar covers a different axis of drift, and none of them substitutes for the others. A perfect character bible with no world bible still produces a warehouse that redecorates itself. A perfect world bible with no character bible still produces an actor whose face drifts scene to scene. And even with both bibles locked, without Continuity Law chaining shots together, you get technically-consistent shots that still cut together like unrelated stills instead of one continuous sequence.
Building the habit
If you're used to thinking about pre-production as "lock the character, then start shooting," the practical shift here is small but important: locking the character is half the job. Before any shot gets generated, ask whether the sequence needs a movement map (does anyone move through the space across multiple shots?), a damage-state decision (does anything in the environment change state across the sequence?), and a weather progression (does the scene span enough time for the sky to plausibly change?).
Not every scene needs all six components at full depth — a single static dialogue shot in one location doesn't need a sky progression if there's no time jump. But the moment a sequence has multiple shots, multiple angles, or spans any meaningful time, the world bible stops being optional pre-production polish and becomes the artifact that decides whether the sequence reads as one coherent place or as a collection of similar-looking but subtly different sets stitched together after the fact.