The three-shot pattern real stunt coverage runs on, adapted for a pipeline where every clip is generated blind.
Three weeks ago I finally got a duel scene in my AI-generated series past an AI video generator’s safety filter. Runway had flagged the first cut as graphic violence, a stylized blade fight with no blood or injury language anywhere in the prompt, and getting it approved meant rewriting everything around impact instead of injury. I was proud of that fix. Then I cut the approved shots together and showed the sequence to three people who hadn’t seen the source material. None of them could tell me who won.
Getting a fight scene past a content filter and getting a fight scene an audience can follow are two different problems, and generating the shots only accidentally solves the first one. Covering a fight scene for AI-generated video means deciding, before you generate anything, whether you’re shooting the geography of the fight or the impact of it, then building each exchange out of a three-part beat, intent, action, reaction, because nothing in the pipeline will make that decision for you. That’s what actually fixed the duel. Not a better prompt. A plan.
What does “coverage” mean, and why doesn’t AI video generate it automatically?
Coverage is the set of shots a scene needs so it can be cut into something coherent: wide enough to show where people are, close enough to show what they feel, matched so the cuts don’t disorient anyone. On a real set, a stunt coordinator and a second unit plan that coverage before a camera rolls, because a fight is the part of a film where bad planning shows up fastest. A confusing dialogue scene reads as flat. A confusing fight reads as noise, an audience that can tell something happened but not what, or to whom.
An AI video generator has none of that planning built in. Ask a model for “an intense sword duel” and it gives you motion, sparks, sound, energy, everything except an opinion about which shot should come next. Each generation is an isolated response to one prompt, with no memory of the clip before it and no stake in whether the finished sequence makes sense. A few things worth naming before you write a single fight prompt:
- No coordinator behind the model. Nobody is deciding blocking, spacing, or who should be on which side of frame. That call is entirely yours now.
- No persistent camera. Every clip is a fresh generation, so nothing enforces where the “camera” was standing for the last shot.
- No default coverage philosophy. The model won’t pick between showing you the room or showing you the hit. It just answers the prompt in front of it.
- No sense of escalation. A fight is an arc. Left alone, a model will happily generate beat one and beat eight at the same intensity.
Bad action isn’t filmed badly. It’s covered indecisively.
Ian Lynch Smith, Previs Pro
What are the two philosophies of covering a fight scene, and which one fits an AI pipeline?
Almost every screen fight leans toward one of two poles, and mixing them at random is what produces the muddled sequence nobody can describe afterward. The first is geography: the audience always knows where everyone is, using wider lenses and longer-held masters, so the fight reads like a chess problem you can follow. The second is impact: geography gets sacrificed for sensation, using tight singles and fast cuts so the audience feels every hit even if they lose track of exactly where it landed.
For a stateless generation pipeline, geography coverage is usually the more forgiving choice to start with. It asks for fewer precise continuity matches across independent clips, mostly wide, well-lit masters and consistent character positioning, which is exactly the kind of shot current models hold together best. Impact coverage asks for the opposite: tight, fast singles where a hand, a weapon, or a face has to land in almost the same place shot to shot, and complex physical interactions and fast-paced motion remain one of the areas where generative video models are still least predictable, even as individual models improve. ByteDance’s Seedance 2.0, launched in February 2026, was built specifically to handle harder motion and physical interaction more smoothly than earlier models, which is a sign of where the gap is, not proof it’s closed.
That’s the trade to make on purpose, not by accident:
- Geography-first if your pipeline still struggles with fast, close motion, or if the scene needs the audience to track more than two combatants.
- Impact-first if your model handles close motion reliably and the beat is a single, simple exchange you can afford to shoot with fewer, punchier singles.
What is the “triplet,” and how do you shoot it across generations with no shared memory?
The smallest readable unit of action coverage isn’t one shot, it’s three: intent, action, reaction. Intent is the moment a fighter sees the opening. Action is the hit itself, usually shot on the longest lens of the three, since compression is what sells the impact. Reaction is someone receiving the hit, or a third party watching, which is what tells the audience the moment actually mattered. A two-minute fight scene runs somewhere around twelve to twenty of these triplets, not twelve to twenty individual shots.
The hard part in an AI pipeline isn’t understanding the triplet. It’s that each of the three shots is a separate, memoryless generation call, with no shared context telling it what happened one prompt ago. Left unmanaged, the intent shot and the reaction shot can end up describing two different rooms. The fix is writing each prompt in the triplet as if the model has amnesia, because it does:
- Decide the philosophy for the whole scene first. Geography or impact, not a mix, chosen before you write a single prompt.
- Break the fight into beats, then each beat into intent, action, reaction, before generating anything. This is a planning step, not a generation step.
- Write the lens into every prompt. Wide or medium for intent, the longest lens you’re using for action, medium-to-close for reaction. Lens language is one of the few continuity cues models consistently respect.
- Feed the intent shot’s reference image into the reaction shot’s generation, when your tool supports it, so the room and the characters don’t quietly drift mid-beat.
- Log which triplet each generated clip belongs to, before moving to the next beat. A fight with nine triplets and no log is nine separate memory tests you will fail by shot six.
How do you keep the axis of action straight when the fight itself keeps moving?
In a dialogue scene the line sits between two faces and mostly stays put. In a fight, the axis of action moves with the combatants, and it pivots the moment they do. Cross it without a bridge and the cut reads as the two fighters swapping places, a disorientation the audience feels without being able to name.
A few practical rules carry over directly from real fight choreography, and matter more in AI video because nothing enforces them automatically:
- Track the line as the bodies move, and restate it in the next prompt every time it shifts, not just at the start of the scene.
- Cross deliberately, on a motivated move. A push-in, a whip-pan, a body passing the lens gives the audience a transition to read the new geometry against.
- Use a neutral, axial “cheat shot” as a bridge between two opposite-angle generations when you need to reset the geometry without a full regeneration.
- If you’ve crossed by accident, check whether a flipped version of the take fixes it. Mirroring only works cleanly when there’s no readable text or asymmetric prop in frame, which is one more reason to keep those out of a fight shot’s background in the first place.
Redoing the duel scene with this discipline meant generating nine shots instead of one continuous take: three triplets, geography-first, with a locked wide reference image carried into every reaction shot and the axis restated in each prompt. The same three test viewers who couldn’t follow the first cut could tell me, unprompted, who landed the last hit. I keep the philosophy decision, the beat breakdown, and the axis notes for that scene next to the shot list, so re-cutting the fight later doesn’t mean rebuilding the whole plan from memory.
FAQ
Does this only apply to fight scenes?
No. Any sequence where the audience has to track who is doing what to whom, a chase, a rescue, a struggle over an object, benefits from the same philosophy decision and the same intent-action-reaction unit. Fights just make the failure easiest to spot.
Do any AI video tools plan coverage for me?
Not yet. Newer models improve the physical plausibility of one generated shot, Seedance 2.0 being a recent example built for smoother motion in complex interactions, but none of them decide which shots a scene needs or in what order. That judgment still belongs to the person writing the prompts.
How many shots does a short AI-generated fight actually need?
Fewer than it feels like it should. If a full two-minute fight runs roughly twelve to twenty triplets, a fifteen-to-twenty-second AI-generated duel usually needs two or three, six to nine shots total, not one long continuous generation asked to do everything at once.