The 180-degree rule across generated clips
Independently generated clips flip screen direction and the cut reads as a jump. How to state the axis in every prompt, and the check to run before assembly.
You generate six clips for a sequence. Each one looks good on its own. You lay them on a timeline and shot four lands like a stumble — nothing in it is broken, the character is the same character, the room is the same room, but the cut feels wrong in a way nobody in the review can name. Nine times out of ten the subject was facing screen-right for three shots and is facing screen-left in the fourth. The camera crossed the line, and because each clip was sampled independently, nothing stopped it.
This is the oldest continuity rule in film and it has a new failure mode. In live action, crossing the line takes a deliberate act: someone has to physically carry the camera to the other side of the scene. In generation it takes nothing at all. It is the default outcome of asking a model twice.
What the line actually is
Draw an imaginary line through the two things a scene is about — two people talking, a runner and the finish line, a car and the road it is on. That line is the axis of action. Keep the camera on one side of it and screen geography stays stable: whoever is on the left stays on the left, whoever faces right keeps facing right, and anything moving left-to-right keeps moving left-to-right. Cross it and everything mirrors.
The reason the rule exists is that an audience builds a mental map of the space from the first shot and keeps using it. When the map inverts, the brain spends a beat re-orienting, and that beat is experienced as a bump. Viewers almost never diagnose it. They just report that the video "feels choppy" or "doesn't flow", which is exactly the feedback that sends people off optimising cut timing on a problem that has nothing to do with timing.
Three things live on the axis and all three flip together:
- Position. Which half of the frame each subject occupies.
- Facing. Which way a subject is looking or angled.
- Travel. Which way something moves across frame.
Travel is the one people forget. A car heading screen-right in the wide and screen-left in the next shot doesn't read as a new angle. It reads as the car turning around.
Why generation breaks it and editing doesn't
A text-to-video model has no memory of the clip you generated ten minutes ago. Each generation is a fresh sample, and the composition — including which side of frame the subject lands on — is decided inside that sample. Two identical prompts can return mirrored layouts, because "a woman standing beside a motorcycle" is satisfied equally well by either arrangement.
A seed doesn't rescue this. Fixing the seed reproduces a generation, so the same prompt and seed will give you the same clip back. It does not carry a spatial convention across a different prompt, and every shot in your sequence is a different prompt. Seeds are for reproducing a result, not for enforcing geography.
Nor does the model know the rule by name. Writing respecting the 180-degree rule or maintaining screen direction in a prompt does close to nothing, because these are names of a convention rather than descriptions of a frame. Models execute descriptions of frames. If you want the axis honoured, you have to describe the frame that honours it.
There is one clean way around the whole problem, worth knowing before you fight it: keep the whole exchange inside a single generation. A handful of frontier models take a multi-shot prompt and return the cuts already assembled, and because one generation holds one internal space, the axis is the model's problem rather than yours. That path and its limits are covered in the write-up on directing cuts inside one generation. Everywhere else, you state the axis yourself.
Stating the axis in a prompt
Pick a side once, write it into a reusable block, and paste that block into every shot in the sequence. The block should answer position, facing and travel in plain language with no film vocabulary in it at all.
Weak, and effectively ignored:
Two colleagues talking in an office. Respect the 180-degree rule
and keep continuity with the previous shot.
Strong, and repeatable:
[AXIS] The woman stands at frame left, angled toward frame right.
The man stands at frame right, angled toward frame left.
The window is behind them on the far side of the room.
The camera stays on the near side of the desk throughout.
Three details make that block work harder than it looks:
- A background anchor. "The window is behind them on the far side" gives the model a fixed element that can only appear one way if the camera stays put. Anchors are more reliable than subject descriptions because they don't move.
- Absolute frame positions, not relative ones. "To her left" is ambiguous — her left or ours? "Frame left" is not.
- A stated camera side. Naming where the camera lives is what makes the constraint hold across a shot size change from wide to close.
For travel, add the direction explicitly and repeat it verbatim: the car moves left to right across frame in every shot the car is in. If the sequence intercuts two subjects moving toward each other, they need opposite travel directions in every shot — one left-to-right, the other right-to-left — otherwise they read as a convoy rather than a collision course.
If you're already writing structured camera language, the axis block slots in beside it; the camera movement prompt glossary covers the move vocabulary that sits alongside it.
The pre-assembly check
The check takes about two minutes and it works on still frames, not clips. Pull the first and last frame of every generation into one contact sheet and go across it answering the same four questions per frame.
| Question | Pass condition |
|---|---|
| Which half of frame is the subject in? | Same half in every shot in the sequence |
| Which way are they facing or looking? | Same screen direction throughout |
| Which way is anything moving? | Same travel direction throughout |
| Where is the background anchor? | Same side of frame in every shot |
Stills are the right unit here: far faster than scrubbing six clips, and a mirrored composition is instantly obvious in a grid of frames while being easy to miss when clips play one after another and you're busy watching performance.
Do this before you spend on finished renders. If you want to check it in motion instead, build the assembly in the AI video editor and run a preview: true pass — 480p, no credit cost, with a short per-user cooldown between requests, and the final export charged once regardless of how many clips are on the timeline. Screen direction survives a downscale perfectly, so 480p answers this question as well as 4K would. The previews and final export breakdown has the details of what gets charged when.
The reroll economics favour catching it early. An axis flip is not a fixable defect in post — you can't flip the clip without flipping hair partings, text and wardrobe — so a shot that crosses the line is a shot you regenerate. Budgeting for those is its own discipline, laid out in the reroll rate breakdown.
Crossing the line deliberately
The rule is a default, not a law, and sequences that never break it can feel airless. There are three ways to cross without a bump:
- Cut on a neutral shot. A head-on shot, a straight-down-the-barrel angle, or a symmetrical composition has no screen direction to violate. Put it between the two sides and the audience re-maps on the neutral frame.
- Cut on movement. If a subject physically walks around the other, the camera can follow across the line and the audience sees the geography change rather than inferring it.
- Cut away first. An insert — hands, an object, a detail with no spatial cues — resets the map. This is why detail shots are worth generating even when the script doesn't ask for one.
The deliberate version reads as a choice. The accidental version reads as an error. The only difference is whether the shot before it prepared the audience.
FAQ
Does writing "180-degree rule" in the prompt do anything?
Very little. It is the name of a convention, not a description of a frame, and models act on descriptions. Replace it with the frame you actually want: which half of frame each subject occupies, which way each one faces, and which side of the scene the camera sits on. That phrasing is what a model can execute.
Why do two identical prompts come back mirrored?
Because composition is decided inside each sample, and a prompt like "a woman beside a motorcycle" is satisfied equally by either arrangement. Nothing in the text prefers one. Fixing the seed reproduces one specific generation but does not carry a layout convention over to a different prompt, so a six-shot sequence written as six prompts has six independent chances to mirror.
How do I check for an axis flip quickly?
Export the first and last frame of every clip into a single contact sheet and read across it. You are looking for one thing: does the subject stay in the same half of frame, facing the same way, in every shot. A mirrored composition is obvious in a grid of stills and easy to miss when clips play sequentially. It takes about two minutes and catches the problem before any finished render is paid for.
Can first-and-last-frame generation help hold the axis?
Yes, and it is the strongest available lock. Supplying both endpoints as images fixes composition at the start and the end of the clip, so the model interpolates between two layouts you chose rather than inventing one. First and last frame control is available on a small set of models, including Flux 3's first-last-frame-to-video mode. It costs you the freedom to be surprised, which on a continuity-critical sequence is the trade you want.