What the system is hiding
A three-minute scene might be built from twenty shots recorded across two days from six camera positions. Continuity editing is the accumulated craft that stops the audience noticing any of that. It is not one rule but a bundle of them, and they all serve one goal: keep space, time, and attention coherent across the join.
The convention is so absorbed that it only becomes visible when broken. Audiences with no vocabulary for any of it will still feel that something is wrong when a character's coffee cup switches hands, or when two people talking appear to have swapped sides of the table.
The core rules
- The 180 degree rule. An imaginary axis runs between the subjects; the camera stays on one side. This is what keeps screen direction stable across a conversation.
- Eyeline match. A look and the thing looked at must agree in angle and height. A character looking down should be followed by a shot angled downward.
- Screen direction. Someone exiting frame right should enter the next shot from frame left. The same applies to vehicles, chases, and anything travelling.
- Cutting on action. Placing the cut inside a movement so the motion carries across it. The most reliable way to make a cut disappear.
- The 30 degree rule. Change the angle by at least 30 degrees between shots of the same subject, otherwise the cut reads as a jump rather than a new view.
- Match the details. Wardrobe, props, light, weather, and hair have to agree across takes. This is what a continuity supervisor tracks on set.
When to break them
Every rule here is a default, not a law. Crossing the axis in the middle of an argument makes an audience uneasy before they can name why. Deliberately mismatched eyelines can suggest that two people are not really in the same conversation. Wong Kar-wai and Kubrick both violate the system regularly, and the effect works because the audience has been trained by the convention it is departing from.
The one condition is legibility of intent. A break reads as style when something else changes with it: the light, the sound, the pace. A break with nothing supporting it reads as an error.
Continuity across generated shots
This is the hardest problem in AI filmmaking, and it is worth being blunt about why. Continuity editing assumes a physical set that persists between shots. A generative model has no set. Each run re-imagines the room, the wardrobe, and the light from whatever the prompt implies, so the very things the system depends on are exactly the things that drift.
What actually helps:
- Anchor, do not describe. Generate one image of the space and one of each character, then use those as reference or start frames for every shot. Re-describing a character in words gives you a different person each time.
- Fix the camera side in words and keep it. Include the side of the axis in every prompt for the scene: camera on the left side of the table. This is the only handle you have on the 180 degree rule, since a model has no concept of the line.
- Generate long, cut short. A single longer clip that contains a movement gives you a real motion to cut on, which is far more convincing than joining two separate clips at a guessed frame.
- Use shared frames for adjacent shots. Where two shots must connect tightly, take the last frame of one and use it to condition the next. That is the closest available equivalent to shooting coverage on one set.
- Accept discontinuity where it is free. If continuity is fighting you, restructure the scene as a montage or a cross cut, where visible discontinuity is the convention rather than the failure.
Details will still slip: a shirt button, a glass, the direction of a shadow. Fix what a viewer's eye will land on, which is faces, hands, and anything moving, and let the rest go.