How to choreograph a fight scene with AI in 7 steps, from a real 30-second duel

To choreograph a fight scene with AI, you write the choreography into the prompt, because the video model performs only what you describe. Lock each fighter and the arena with reference images, give each fighter a way of fighting, and break every move into contact, force, reaction and the next movement. Then write the fight as numbered shots with a time window each, give the camera its own rules, and ask for "No slow motion."

Film: The duel by Jaime Yao (Flick team)

How it was made: One Seedance 2.5 prompt with 3 reference images: 30 seconds, 21:9 widescreen, 720p, sound generated with the picture. Every example below comes from it unless it says otherwise.

How to choreograph a fight scene with AI, step by step

Seven steps, each with real lines from The duel's prompt. "…" marks lines we left out. In the prompt, A is the spearman, B is the swordsman, and @ names an image on the Flick canvas.

1. Lock the fighters, their weapons and the arena

What it is

Reference sheet: A picture of one character from the front, side and back. The video model copies the face, costume and weapon from it in every shot.

Arena plate: A still of the location with no people in it. It fixes the ground, the light and anything the fighters can use.

Why it matters: The model redraws every frame. Without locked references, a face drifts, a spear changes length, or a scabbard disappears between cuts.

How to get it

Put the weapon on the sheet: Make each sheet with the weapon in hand, in the state it will be used. Jaime's sheet prompt for B says: "Make this explicit: the Han sword has already been drawn, the full blade is visible, and the scabbard contains no blade."

Keep the arena empty: The plate's prompt opens with "Pure environment, live-action realistic cinematic still frame. No people, no animals, no buildings." Give it features the fight can use: The duel's low trunks become a springboard in shot 9.

Open the prompt with one line per reference: Each fighter gets an appearance line pointing at their sheet, then the arena.

Model: Seedance 2.5 with references · Note: Opening lines of the video prompt

A's appearance match @Character Turnaround with Red-Tasseled Spear 100%.
B's appearance match @Character Sheet, Three-View with Han Sword 100%.
…
Use the provided environment reference: an open clearing inside an arid autumn poplar forest, with massive twisted old poplar trees, exposed roots, hard sandy earth, scattered gravel and dry fallen leaves, a narrow shallow stream nearby, and thick low-growing poplar trunks and branches creating natural changes in elevation.

2. Give each fighter a way of fighting

What it is: Before any single move, write how each fighter fights: which hand holds what, how they keep distance, and which parts of the body drive the movement. Jaime's prompt calls this the global fighting logic.

Why it matters: You number the big moves; the model fills in everything between them. If it knows A fights at the spear's length and B defends with both hands, the in-between moves stay in character.

How to get it

Give every hand a job: "B fights with the sword in the right hand and the scabbard in the left hand throughout the entire fight." Then close the gaps: "The scabbard never disappears, is never dropped, and is never passively left hanging."

Name body parts, not techniques: A model can show "wrist rotation" and "deep knee compression". It has to guess what a named technique looks like.

Model: Seedance 2.5 with references · Note: Global fighting logic, excerpt

A controls distance with the spear.

A’s spear movement is driven by:

both hands sliding and repositioning along the shaft;
wrist rotation;
shoulder rotation;
strong waist and hip turns;
pivots through the feet;
forward stepping;
deep knee compression;
explosive jumping.

A rarely brings the spear back to a neutral guard.

3. Break every move into beats: contact, force, reaction, next movement

How to Make an AI Film That Doesn't Look Like AI Slop, with Jaime Yao (Flick on YouTube). The action method starts at 8:54.

What it is: A beat is the smallest piece of a fight the viewer has to see. Jaime Yao's rule: "For action scenes, I break each move into contact, force, reaction, and the next movement." Then: "I also give those beats a time window."

Why it matters: "He attacks" leaves every detail to the model. Sam Wasserman hit this on The Miracle (Flick Residency): "My first few versions of the fight sequence, it was really plain."

How to get it

Write each beat as a chain of the body: Join the physical events with arrows, in order. Shot 9 lasts 0.72 seconds and is one foot landing on a trunk:

Model: Seedance 2.5 with references · Note: Shot 9, without its header

Physical sequence:

foot contact → ankle compresses → knee bends → body weight loads → leg immediately extends.

The branch acts as a springboard.

B does not simply touch it.

B visibly transfers weight into it.

Write the reaction as its own beat: The force has to go somewhere. "B visibly transfers weight into it" is what makes the landing read as heavy.

Let the agent write the detail: In Jaime's words, "you don't have to write all of that yourself." Tell the agent something like "his attacks are extremely fast. He delivers three consecutive strikes in 1 second, so make sure the timing reflects that." Give it Flick's Seedance Action Guide as its rules, and it breaks each action into stages for you.

4. Write the fight as numbered shots with time windows

What it is

Shot: The footage between two cuts.

Time window: When a shot starts and ends inside the clip, in seconds. Seedance 2.5 can make one 30-second clip that cuts between several shots, so one prompt can hold a whole fight.

Why it matters: The windows set the rhythm: which moments flash by and which ones the viewer gets to watch. Jaime wrote In ’twixt Vows (Flick Residency) the same way: 130 of its 263 kept video prompts number their shots (Flick's public breakdown of the film, October 2026).

How to get it

Plan the timing table before the prompt: Here is The duel's. Its 13 windows add up to exactly 30 seconds.

Shot	Window (seconds)	Length (seconds)	What it shows
1	0.00 to 0.20	0.20	A flash of A before the fight
2	0.20 to 1.80	1.60	B in his sword-and-scabbard stance, a short pause
3	1.80 to 3.48	1.68	Extreme close-up: A's spear comes alive
4	3.48 to 6.28	2.80	Wide, full body: A's opening spear combination
5	6.28 to 8.48	2.20	Close, over the shoulder: the first exchange
6	8.48 to 11.56	3.08	Tight on B: five near misses he dodges with his body
7	11.56 to 14.48	2.92	Medium two-shot: B turns the spear aside with sword and scabbard in turn
8	14.48 to 20.12	5.64	Wide, tilted angle: B's long leap through the air as A gives chase
9	20.12 to 20.84	0.72	Extreme close-up insert: B's foot lands on a poplar trunk
10	20.84 to 24.32	3.48	B rebounds into the fight
11	24.32 to 27.52	3.20	Wide, full body: the fighters trade high and low, jumping in turn
12	27.52 to 28.44	0.92	Tight on B: one last near miss
13	28.44 to 30.00	1.56	The ending: A kicks his own spear at B, B blocks and retreats

The duel's 13 shots and their time windows, read from its public Flick workflow, October 2026. Descriptions are ours.

Mix long and short shots: Give a full combination a wide shot of a few seconds (shot 4). Use inserts, tight shots of one detail, under a second as punctuation (shot 9).

Number the attacks inside a long shot: Shot 4 packs seven numbered attacks into 2.8 seconds. Each one starts from where the last one ended.

Model: Seedance 2.5 with references · Note: Shot 4, excerpt

Attack 1 … broad low-to-mid sweep

A steps in and rotates the hips strongly.

The spear makes a broad horizontal sweep across the space between them.
…
Attack 2 … immediate forward extension

A does not reset.

Ask for real speed and cut mid-move: The duel asks for "No slow motion." near the top, and near the end: "Cuts frequently occur mid-action. Do not finish a move, pose, and then cut." Cutting while a move is still going carries the speed across the cut.

Watch for: Treating the windows as exact. Flick's Seedance Action Guide says they guide pacing, not exact editing points. In The duel's finished clip, the wide leap starts at about 15.9 seconds, not 14.48. Check where the cuts land.

5. Give the camera its own rules and keep the fighting axis

What it is

Fighting axis: The imaginary line between the two fighters. Keep the camera on one side of it, and A stays on one side of the frame and B on the other in every shot.

Coverage: The set of angles you plan for the fight: wide, medium, close-up and insert, so you can cut between them.

Why it matters: Fast cuts only work if the viewer always knows who is where. Alan Xiao, director of Tusk Hunters (Flick Residency): "You need the blocking to remain consistent, and you don't want to cross the axis without a reason."

How to get it

Start every shot with a camera line: One framing per shot, before the action. Shot 3 opens "Extreme close-up on the spearhead and front section of the shaft. Background strongly defocused."

Plan the coverage as a list, then the rules: The duel ends with its angles in order and what the camera must not do. Wide shots show where everyone is; close shots and inserts show the contact.

Model: Seedance 2.5 with references · Note: Camera language, excerpt

spearhead insert;
wide opening spear combination;
close-range coverage;
extremely tight near-miss sequence;
medium two-shot showing coordinated weapon defense;
wide Dutch-angle airborne choreography;
foot-on-poplar-trunk insert;
…
Hard cuts.

Maintain fighting-axis continuity.

Subtle physical tracking is acceptable.

No unnecessary orbit.

No dramatic zoom.

No chaotic camera shake.

No excessive motion blur that hides hand or weapon choreography.

Shaky or steady, keep it readable: A rain duel in the In ’twixt Vows workflow asks for heavy handheld shake, then adds: "Despite the heavy shake, the choreography must remain readable enough to understand who attacks, who dodges, and where the weapons travel." For a first test, the Seedance Action Guide suggests a locked medium-wide shot (the camera stays still). More moves: AI camera movements.

6. Block the fight in 3D, or borrow the movement from a clip

What it is: Blocking is placing the fighters and the camera before you shoot: who stands where, who faces whom, and how they move. In Flick's 3D Stage you block with colored stand-ins in a rough 3D copy of your set, then record the camera.

Why it matters: Words stop working once fighters circle each other. Alan Xiao uses 3D Stage for this: "its main function is blocking where the characters stand and how they face each other. Once you have more than two characters, that relationship becomes much harder to keep consistent."

How to get it

Block with stand-ins, then hand the recording to Seedance: On the same canvas as The duel, Jaime blocked a one-against-three rooftop fight with one colored stand-in per fighter (the prompt calls them proxies) and grey boxes for buildings. The Seedance 2.5 prompt takes the camera, positions and cut points from that recording, and the look from the reference images.

Model: Seedance 2.5 with references (a rooftop fight on The duel's canvas) · Note: @ names the 3D Stage recording

Use @3d blocking as the exact spatial and camera reference. Keep its camera angles, heights, framing, perspective, camera movements, cut points, building positions, character positions, routes, timing, foreground cropping and occlusions 100% unchanged. … Translate the proxies into anatomically proportioned characters while preserving their positions and visible body portions.

Act out the moves like a game: 3D Stage's Play Mode lets you steer a figure with the keyboard and record it. Steps: 3D Stage animation.

Or borrow movement from a clip: For The Miracle, Sam Wasserman made depth maps, black-and-white videos that keep only shape and distance, from other footage "to help make the choreography." Use footage you shot or have the rights to, as a video reference in Omni Reference.

Watch for: A start frame (the clip's first image) that already shows the end of the move. Alan Xiao: "If the starting frame already shows the final pose, that process disappears."

7. Build the fight's sound in layers

What it is: Fight sound is the swishes, the impacts, the footsteps and the place around them. Seedance 2.5 can generate sound with the picture, and you add more layers in the edit.

Why it matters: Sam Wasserman, on The Miracle's rooftop fight in the rain: "you can't just have a thud." A hit that doesn't match the place pulls the viewer out.

How to get it

List the sounds in the prompt: Name every sound the fight makes, and turn off what you'll add later. The duel turns off dialogue, subtitles and music, which leaves room for a score in the edit.

Model: Seedance 2.5 with references · Note: Sound, excerpt

No dialogue throughout the entire video.

No subtitles.

No background music.

Only synchronized natural sound:

fast spear swishes,
sword swishes,
scabbard swishes,
…
rapid footsteps on hard sandy earth, …

Layer the hits for the place: Sam's method: "I would make booms and hits that were dry and clean." Then he made versions with a small and a bigger splash and layered them in for the rain. Flick Academy's Build the sound lesson covers cues and layers.

Case study: how The duel's 30-second prompt is built

The duel's 27,695-character prompt runs in eight parts, in this order:

1. References: one appearance line per fighter, then the arena.
2. Hands and weapons: which hand holds what, for the whole fight.
3. Speed: "This is an extremely fast, precisely choreographed wuxia fight between A and B." (Wuxia is Chinese martial-arts fiction.) Then "No slow motion."
4. Fighting logic for each fighter.
5. 13 timed shots, each with a camera line, then the action as beats.
6. Speed rules: each fighter's pattern of moves "without neutral resets", and cuts mid-action.
7. Camera rules: the angles in order, the fighting axis, and what the camera must not do.
8. Sound: no dialogue, no subtitles, no music, then a list of natural sounds.

Who writes it: Jaime's advice is "don't write the prompt yourself. Have the agent write them, then review and correct them yourself."

The ending: a block, not a blow. In the last shot, A kicks his own spear at B, and B gives way instead of stopping it dead. Only weapons touch in the whole fight, and it ends with the fighters apart.

Model: Seedance 2.5 with references · Note: Shot 13, excerpt

A explosively extends the leg and kicks the shaft of A’s own spear.

The kick is not directed at B.
…
The sequence must read clearly:

stable spear → knee chambers → foot strikes shaft → shaft suddenly accelerates → spear shoots violently toward B.
…
B does not try to stop all the force rigidly.

B deliberately yields with it.

Words that do the work: "The sequence must read clearly" tells the model which beats must be visible. "B deliberately yields with it" makes the block read as skill, not luck.

Fight choreography for AI: what changes from stage combat

Fight choreography is the planned sequence of attacks, defenses and reactions in a fight, designed so the audience can follow it. With AI video no one rehearses, so every job moves into the prompt and the references.

In live action	With AI video
Performers rehearse the moves	You write the moves as beats, and the agent turns them into timed shots
A stunt double matches the actor	A reference sheet keeps each fighter and weapon the same in every shot
Blocking happens on the set	Blocking happens with stand-ins in 3D Stage
The partner sells the hit by reacting	You write the reaction as its own beat
Near misses keep performers safe	Weapon-on-weapon contact and near misses keep the fight easy to read

Five jobs in a live-action fight, and where each one goes with AI video.

Which AI model to use for fight scenes, and what a take costs

Use Seedance 2.5 for AI fight scenes. Every fight prompt we found in Flick team and Residency films ran on it, most with reference images: The duel, two duels in the In ’twixt Vows workflow, and 13 takes of the rooftop fight on The duel's canvas. It makes 4 to 30 second clips that cut between shots, with sound.

Resolution	Credits per 5 seconds	Credits for a 30-second fight
480p	105	630
720p	240	1,440
1080p	540	3,240

Seedance 2.5 credits by resolution, from Flick's Seedance 2.5 page, October 2026. The 480p and 1080p totals are six times the 5-second price.

• 720p is the usual choice for a full-length take: of 421 successful Seedance 2.5 reference takes at a 30-second price, 200 cost exactly 1,440 credits, against 177 at 630 (480p) and 44 at 3,240 (1080p) (Flick data, last 90 days, October 2026).
• Test the choreography cheaply. The rooftop fight's 13 takes all ran at 480p, less than half the 720p price.
• More on settings: our Seedance 2.5 guide. If the motion still isn't what you asked for: The motion is not what I asked for.

Plan your first fight scene in Flick

Sign up, open a canvas and block the fight in 3D before you generate a take.

Every account gets 100 free credits when you sign up, and 100 more every Monday.

Create your free Flick account →

Keep going

• How to storyboard: plan a film in panels, including action scenes.
• How to make realistic AI videos: Jaime Yao's fixes for footage that looks real.
• How to make AI movies: the full film process, with blocking in 3D Stage.
• Behind In ’twixt Vows: Jaime Yao on the film's references, models and choreography.
• 3D Stage export (docs): record one camera or all of them, then use the take as a reference.

About the author

Bilguun Bayarkhuu is a growth engineer at Flick. This guide is built from the public workflow of The duel by Jaime Yao, Flick's public breakdown of In ’twixt Vows, and Flick YouTube videos by Jaime Yao, Sam Wasserman and Alan Xiao.

Frequently Asked Questions

How do you create a fight scene with AI?

Plan it before you generate. Lock the fighters and the arena with reference images, then write the fight as numbered shots with a time window each, every move broken into contact, force, reaction and the next movement. Run it on a model that can cut inside one clip, such as Seedance 2.5.

Which AI is best for fight scenes?

Seedance 2.5: every fight prompt we found in films made in Flick ran on it. It makes up to 30 seconds with several shots in one clip, with sound. A 30-second take at 720p costs 1,440 credits.

How do you script a fight scene?

Write it as numbered shots. Each shot gets a time window, one camera line, and the action as beats, for example "Shot 3, 1.80 to 3.48 seconds: extreme close-up as the spear comes alive." Put each fighter's style at the top and the camera and sound rules at the end.

What makes good fight choreography?

Good fight choreography lets the viewer always tell who attacks, who defends and where the weapons are. Each fighter has a style, moves flow into each other without resets, and every attack gets a visible reaction.

How do actors do fight scenes without hurting each other?

They rehearse, keep a safe distance and use near misses that the camera angle makes look like contact. With AI video no one can get hurt, but the same habits keep a fight easy to read. The duel, a 30-second AI fight made in Flick, uses weapon-on-weapon contact only.

What is fight choreography called?

Fight choreography is also called stage combat: the craft of creating the illusion of physical violence safely, for theater and film. With AI video, that planning goes into the prompt as beats inside numbered, timed shots.