What we made
A vertical short-form piece in the tone of 1990s Hong Kong noir. Not a single frame was shot with a camera. We generated 15 clips, cut them together, and attached a separate outro at the end.
0.00 ~ 73.80 Main film 15 generated clips
73.80 ~ 82.47 Black screen a line of text being typed
82.47 ~ 91.10 Outro 5 cuts re-trimmed from the main film
────────────────────────────────────────────
91.1 seconds
Every figure below was read directly from the CapCut project file. None of it is estimated.
How we made it
The work ran in four stages. Claude handled the first two, Higgsfield the third, and CapCut did the final finishing.
1. Character sheet Claude lock one person to the same face across 15 cuts
2. Shot prompts Claude design each of the 15 shots in writing
3. Clip generation Higgsfield generate the video from the design
4. Finishing CapCut color, captions, music, length
Stage 1 — Build the character sheet first
The thing that breaks most often in AI video is the face. If a different person shows up in every cut, the piece never holds together. So before generating any video, we first produced a sheet that locks one person in three views: front, back and close-up.
That sheet is what keeps the same person across the 15 shots that follow.
Things to watch when you use it
- Attach exactly one photo of yourself
- Pick a photo where the face is clearly visible
- The prompt references the face repeatedly, so this matters
Full character sheet prompt
Using the person in the attached image, generate a single
16:9 character reference sheet divided into three panels on
one continuous seamless light-grey backdrop (#9a9a9a).
WARDROBE (identical in all three panels — replace whatever
clothing appears in the source image):
1990s Hong Kong tailoring. Olive-green pinstripe two-piece
suit, boxy 90s cut with wide notch lapels, slightly oversized
shoulders, single-breasted, jacket worn open. Deep burgundy
silk dress shirt underneath, top two buttons undone, collar
spread over the lapels, no tie. Thin gold chain at the neck,
resting on the shirt. Pleated suit trousers with a soft break
over black leather derby shoes, worn and slightly scuffed.
Hair slicked back and away from the forehead, dry finish,
not wet-look.
LEFT PANEL — full-body FRONT view. Standing straight, arms
hanging naturally at the sides, palms facing thighs, feet
shoulder-width apart, weight even on both legs. Entire body
from the top of the head to the shoe soles inside the frame,
with margin above and below.
CENTER PANEL — full-body BACK view of the same person, turned
180 degrees away from camera. The back of the head and the
nape are visible; the face is not visible at all. The jacket's
back panel, shoulder line and vent are clearly readable.
Identical camera height, distance and scale as the left panel:
both figures' heads at the same height, both figures' shoes on
the same ground line.
RIGHT PANEL — face CLOSE-UP filling the panel edge to edge.
Head and shoulders, face turned 15 degrees to camera left,
eyes to lens, relaxed neutral expression. The burgundy collar
and gold chain visible at the bottom of the frame. Photoreal
skin with visible pores, subsurface scattering, fine flyaway
hairs at the temple. 85mm lens look, shallow depth of field.
RULES
- Same identity, same face, same hair, same wardrobe, same
shoes in all three panels. Do not restyle or beautify.
- Flat even studio softbox lighting, no dramatic shadows, no
color grading, no film grain, no neon. Faint contact shadow
under the shoes only.
- The olive suit must stay olive-green, not brown and not
khaki. The shirt must stay deep burgundy, not red.
- Seamless background across all three panels, no borders,
no dividers, no text, no labels.
- 50mm lens for the full-body panels, no perspective
distortion.
Why it is this long — write it short and the AI quietly “improves” things. The suit drifts to khaki, the face gets retouched, the lighting turns dramatic. Lines like “olive is not brown and not khaki” and “do not beautify” are what stop that.
Stage 2 — Design all 15 shots in writing first
We did not generate cuts and then pick. All 15 shots were locked in writing before anything was generated.
Every shot starts with the same common block below. This is what holds the 15 cuts to one tone.
Style: 1990s Hong Kong neon noir, shot the way films were shot in the '90s.
Arriflex 35 III, Zeiss Super Speed primes. Fuji Eterna 500T processed cool —
coarse organic grain, halation on practical lights, shallow focus.
Low-light night, crushed and moody.
The dominant color must be cold teal-green. Yellow exists only inside the lamp
bulb and a palm-sized halo beneath it. If the frame turns yellow, the frame is wrong.
FORMAT: vertical 9:16. The composition is built for a tall narrow frame — the main
action runs along the VERTICAL axis, not across it. The subject occupies the central
column. Nothing important sits at the far left or far right edge. Do not letterbox,
do not add black bars, do not render a horizontal image inside a vertical canvas.
Two sentences here did most of the work.
“If the frame turns yellow, the frame is wrong” — if you only instruct the color, the AI drifts warmer bit by bit. You have to declare it wrong for all 15 cuts to come out the same color.
“Do not render a horizontal image inside a vertical canvas” — without this line, even when you ask for vertical, it draws a horizontal picture and adds black bars top and bottom.
Below is the unique part of each shot. The common block above was prepended to every one of them.
SHOT 01 — Cold open (3 s)
PHYSICS RULES
- Walking: heel lands first, strict left-right alternation, one foot always on the ground. No gliding, never both feet airborne, never the same foot twice.
- Hands: there are only two hands in the frame, both belonging to the same person, entering from the same sleeve.
- Real-time speed. Slow motion is banned. Do not ramp, do not ease, do not stylize the tempo.
- Doors: the tiger is painted ONLY on the outer face of the door. The door is hinged on the left as seen from the corridor and swings INWARD into the flat.
ASSETS
@char_MEI — Chinese woman, early 30s, dark red striped mandarin-collar silk top, low-rise faded bootcut jeans, brown belt with silver conchos, black ankle boots. Black hair in a low bun. Deep red nail polish on both hands.
@door_tiger — heavy wooden flat door, deep green paint, a hand-painted tiger covering the OUTER face only. Brass lock cylinder at hand height. Inner face plain unpainted wood.
SHOT 01 — 3 seconds.
Extreme close-up, vertical: the brass lock cylinder of @door_tiger fills the center of the tall frame, painted tiger fur filling the space above and below the lock.
Two hands enter from the bottom of the frame, both belonging to @char_MEI, both from the same sleeve. Deep red nail polish catches the light. One hand steadies the door, the other works a thin pick into the cylinder.
The face is never visible. Camera locked off, no movement. Rain audible off-screen. Cold teal-green key light from frame-left, wet sheen on the brass. The red nails are the only warm color and they are small in frame.
SHOT 02 — The alley (5 s)
PHYSICS RULES
- Walking: heel lands first, strict left-right alternation, one foot always on the ground. No gliding, never both feet airborne, never the same foot twice.
- Hands: there are only two hands in the frame, both belonging to the same person, entering from the same sleeve.
- Real-time speed. Slow motion is banned. Do not ramp, do not ease, do not stylize the tempo.
- Scale: the corridor railing is 110 cm; on a 185 cm man the top rail lands just above his belt.
ASSETS
@char_LI — Korean man, late 30s, olive-green pinstripe 1990s two-piece suit with wide notch lapels and slightly oversized shoulders, jacket worn open, deep burgundy silk shirt with the top two buttons undone and the collar spread over the lapels, thin gold chain at the neck, pleated trousers, black leather derby shoes, scuffed. Hair slicked back and away from the forehead, dry finish. Silhouette from behind: square shoulder line, slicked-back hair with a clean nape line, the jacket's single back vent.
SHOT 02 — 5 seconds.
@char_LI walks away from camera down a narrow rainy Hong Kong alley at night, seen from behind, full body, centered in the tall frame. Vertical signage stacked overhead on both walls compresses the frame into a tall corridor of neon.
Camera tracks behind him at a steady walking pace, held at chest height, moving forward only — no pan, no tilt.
He carries four grocery bags, two in each hand. His shoulders are relaxed. He is an ordinary man carrying groceries home.
Cold teal-green throughout. Puddles reflect the vertical signs.
SHOT 03 — The stairwell (5 s) · failed once, rewritten
In the first version the grocery bags vanished. The four bags he carried in shot 2 were gone in shot 3, so the two shots didn’t connect. Below is the corrected version.
PHYSICS RULES
- Walking: heel lands first, strict left-right alternation, one foot always on the ground. No gliding, never both feet airborne, never the same foot twice.
- Real-time speed. Slow motion is banned. Do not ramp, do not ease, do not stylize the tempo.
- Hands: there are only two hands in the frame, both belonging to the same person, entering from the same sleeve. Both hands are occupied — nothing is ever set down or picked up in this shot.
- Carried objects obey gravity: the bags hang straight down from the fists, handles taut and vertical, and they rise at exactly the same rate as the body carrying them.
ASSETS
@char_LI — Korean man, late 30s, olive-green pinstripe 1990s two-piece suit with wide notch lapels and slightly oversized shoulders, jacket worn open, deep burgundy silk shirt with the top two buttons undone and the collar spread over the lapels, thin gold chain at the neck, pleated trousers, black leather derby shoes, scuffed. Hair slicked back and away from the forehead, dry finish.
@bags_grocery — four brown paper grocery bags, two held in each hand, carried by twisted paper handles gripped in the fists. Creased and slightly damp from the rain, one bag darkened along the bottom edge. A stalk of green onion sticks out of the top of one bag. Identical four bags in every shot.
@stairwell — narrow concrete tenement stairwell, wet handrail, one flickering fluorescent tube, a dark rectangular stair opening in the floor of the landing.
SHOT 03 — 5 seconds.
@stairwell landing. The camera is locked off at floor level, framing the dark rectangular stair opening in the lower third of the vertical frame.
@char_LI rises INTO frame out of the stair opening, carrying @bags_grocery — two paper bags in his left hand, two in his right, arms hanging straight down at his sides, both hands full for the entire shot.
He is lifted one invisible step at a time over three seconds: first the top of his head, then his face, then his shoulders, then his waist — and the bags clear the edge of the stair opening at waist height, still hanging from his fists — then his legs. The stair opening is a mandatory visible object and must remain an open hole in the floor — it is never a solid wall.
He does not jump, does not teleport, does not appear all at once. His head stays at a constant lateral position while rising vertically. The bags never float, never swing outward, never change count.
Flickering fluorescent tube overhead, cold teal-green. Wet paper catches one point of light.
The fix — register the bags as a named object called @bags_grocery and put
“the bags never change count” and “both hands stay full” into the rules.
Props need names just like characters do, or they disappear between cuts.
SHOT 04 — The tiger door (4 s)
PHYSICS RULES
- Walking: heel lands first, strict left-right alternation, one foot always on the ground.
- Hands: there are only two hands in the frame, both belonging to the same person, entering from the same sleeve.
- Real-time speed. Slow motion is banned.
- Doors: the tiger is painted ONLY on the outer face of the door. The door is hinged on the left as seen from the corridor and swings INWARD into the flat.
ASSETS
@char_LI — Korean man, late 30s, olive-green pinstripe 1990s two-piece suit with wide notch lapels and slightly oversized shoulders, jacket worn open, deep burgundy silk shirt with the top two buttons undone and the collar spread over the lapels, thin gold chain at the neck, pleated trousers, black leather derby shoes, scuffed. Hair slicked back and away from the forehead, dry finish.
@door_tiger — heavy wooden flat door, deep green paint, a hand-painted tiger covering the OUTER face only. Brass lock cylinder at hand height. Inner face plain unpainted wood.
@key_tiger — small brass key on a ring with a tiny carved tiger charm.
SHOT 04 — 4 seconds.
@char_LI stands in front of @door_tiger in the corridor. The door fills the vertical frame top to bottom, the painted tiger's face at his eye level.
Composition is a flat frontal two-shot of man and door, camera at his shoulder height, slight low angle.
He sets down the grocery bags, takes @key_tiger from his trouser pocket, and unlocks the door. The door swings INWARD, away from camera. The tiger is on the outer face only.
SHOT 05 — Taking off the jacket (5 s)
PHYSICS RULES
- Hands: there are only two hands in the frame, both belonging to the same person, entering from the same sleeve.
- Real-time speed. Slow motion is banned.
ASSETS
@char_LI — Korean man, late 30s, olive-green pinstripe 1990s two-piece suit, jacket worn open, deep burgundy silk shirt, thin gold chain at the neck, pleated trousers. Hair slicked back, dry finish.
@char_LI_indoors — the same man, jacket removed. Deep burgundy silk shirt untucked at one side, sleeves rolled to the forearm, thin gold chain visible, olive pleated trousers, barefoot on the tiles. Hair still slicked back but one strand loose at the temple.
@window_open — the flat's single window, sash pushed fully UP and staying open, rain visible outside.
SHOT 05 — 5 seconds.
Interior of the flat, night. @char_LI takes off the olive pinstripe jacket in one unhurried motion and hangs it on a hook by the door, becoming @char_LI_indoors — burgundy silk shirt, sleeves rolled to the forearm, barefoot.
He then pushes the window sash fully upward. Use @window_open: after his hands leave it, the sash STAYS fully up and open for the rest of the shot and the rest of the film. Rain and the night skyline are visible through the open frame.
Vertical composition: the tall open window occupies the upper two thirds of the frame, he stands in the lower third, back to camera. Camera static.
Cold teal-green. No yellow anywhere in this frame.
SHOT 06 — Midnight, undecided (5 s)
PHYSICS RULES
- Hands: there are only two hands in the frame, both belonging to the same person, entering from the same sleeve.
- Real-time speed. Slow motion is banned.
ASSETS
@char_LI_indoors — Korean man, late 30s, deep burgundy silk shirt untucked at one side, sleeves rolled to the forearm, thin gold chain visible, olive pleated trousers, barefoot on the tiles. Hair slicked back with one strand loose at the temple.
@lamp_kitchen — single bare warm bulb over the kitchen table, the only yellow light source in the film.
SHOT 06 — 5 seconds.
@char_LI_indoors sits alone at the kitchen table under @lamp_kitchen. Vertical composition: the bare bulb hangs in the top of the frame, he sits directly beneath it in the lower half, the rest of the room crushed to near black.
Yellow exists only inside the bulb and in a palm-sized halo on the tabletop beneath it. Everything else is cold teal-green.
He looks at a telephone on the table without reaching for it. He does not move except to breathe and blink once.
Camera static, slight low angle, 40mm.
SHOT 07 — A shadow in the corridor (5 s)
PHYSICS RULES
- Walking: heel lands first, strict left-right alternation, one foot always on the ground. No gliding, never both feet airborne, never the same foot twice.
- Real-time speed. Slow motion is banned.
- Scale: the corridor railing is 110 cm.
ASSETS
@char_MEI — Chinese woman, early 30s, dark red striped mandarin-collar silk top, low-rise faded bootcut jeans, brown belt with silver conchos, black ankle boots. Black hair in a low bun. Deep red nail polish on both hands.
@door_tiger — heavy wooden flat door, deep green paint, a hand-painted tiger covering the OUTER face only.
@bouquet_peony — half-open peony bouquet, deep red, wrapped in cheap cellophane and newspaper, slightly crushed on one side.
SHOT 07 — 5 seconds.
Corridor, after midnight. @char_MEI walks away from camera toward @door_tiger at the far end, seen entirely from behind, full body, centered in the tall vertical frame.
Her face is never visible in this shot. She is identified only by her silhouette: the low bun, the narrow shoulder line, the concho belt catching one point of light.
She carries @bouquet_peony down at her right side, held low so it reads as a shapeless dark mass, not as flowers.
Camera tracks behind her at a steady walking pace, chest height, moving forward only. The camera speed of SHOT 07 equals the camera speed of SHOT 02, exactly.
SHOT 08 — The lock (4 s)
PHYSICS RULES
- Hands: there are only two hands in the frame, both belonging to the same person, entering from the same sleeve.
- Real-time speed. Slow motion is banned.
- Doors: the tiger is painted ONLY on the outer face of the door. The door is hinged on the left as seen from the corridor and swings INWARD into the flat.
ASSETS
@char_MEI — Chinese woman, early 30s, dark red striped mandarin-collar silk top, deep red nail polish on both hands.
@door_tiger — heavy wooden flat door, deep green paint, a hand-painted tiger covering the OUTER face only. Brass lock cylinder at hand height. Inner face plain unpainted wood.
SHOT 08 — 4 seconds.
The same extreme close-up geometry as SHOT 01: the brass lock cylinder of @door_tiger centered in the tall frame, tiger fur above and below.
Two hands of @char_MEI, both from the same sleeve, red nails, working the pick. The cylinder turns. The door begins to open INWARD, away from camera, and a wedge of even darker interior opens at the edge of frame.
Camera locked off, identical to SHOT 01. No face.
SHOT 09 — He wakes (4 s)
PHYSICS RULES
- Real-time speed. Slow motion is banned.
ASSETS
@char_LI_indoors — Korean man, late 30s, deep burgundy silk shirt, sleeves rolled to the forearm, thin gold chain visible, olive pleated trousers, barefoot. Hair slicked back with one strand loose at the temple.
@bead_curtain — red plastic bead curtain in the kitchen doorway, strands reaching the floor.
SHOT 09 — 4 seconds.
Interior, dark. The face of @char_LI_indoors in close-up, lower half of the vertical frame, lit only by a thin cold teal-green spill from the corridor. @bead_curtain hangs in the foreground of the upper frame, red strands crossing vertically in front of the lens, out of focus.
His eyes open. He does not sit up, does not gasp, does not turn his head. Only the eyes move, tracking something off-screen to camera left. His breathing does not change.
Camera static, 85mm, very shallow focus.
SHOT 10 — The intrusion (5 s)
PHYSICS RULES
- Walking: heel lands first, strict left-right alternation, one foot always on the ground.
- Real-time speed. Slow motion is banned.
- Doors: the door swings INWARD into the flat.
ASSETS
@char_MEI — Chinese woman, early 30s, dark red striped mandarin-collar silk top, low-rise faded bootcut jeans, brown belt with silver conchos, black ankle boots. Black hair in a low bun.
@door_tiger — heavy wooden flat door, deep green paint, tiger painted on the OUTER face only.
@bouquet_peony — half-open peony bouquet, deep red, wrapped in cheap cellophane and newspaper.
SHOT 10 — 5 seconds.
Low angle from floor level, his point of view. The open doorway of @door_tiger fills the top of the tall vertical frame.
@char_MEI steps through as a full-height BACKLIT SILHOUETTE — no facial detail, no color detail, only the outline: the low bun, the narrow shoulders, the shape of @bouquet_peony hanging at her side reading as an unidentifiable lump.
She takes two steps into the room, heel first, strict left-right alternation.
Camera static, wide lens, extreme low angle so she towers into the top of the frame.
SHOT 11 — Fight 1 / hand insert (4 s)
PHYSICS RULES
- Real-time speed. Slow motion is banned. Do not ramp, do not ease, do not stylize the tempo. Video models slow fights down on their own — this must not happen.
- No third hand ever enters the frame. Exactly four hands exist in this shot, two per person, each pair entering from that person's own sleeves.
ASSETS
@char_LI_indoors — Korean man, late 30s, deep burgundy silk shirt with sleeves rolled to the forearm, thin gold chain, olive pleated trousers, barefoot.
@char_MEI — Chinese woman, early 30s, dark red striped mandarin-collar silk top, deep red nail polish on both hands.
SHOT 11 — 4 seconds.
Insert, tight on hands and forearms only, filling the tall frame.
@char_MEI's right hand comes forward fast and straight, red nails visible. The left forearm of @char_LI_indoors meets it from below and rolls it over in a soft circular chin-na redirect — his wrist rotates outward, her wrist folds inward, her elbow is turned up toward the ceiling. Minimum force, no strike, no impact.
His movement is soft, circular and economical — a champion spending as little force as possible. Hers is straight, fast and committed.
Real-time speed. Slow motion is banned.
Camera handheld, tight, following the hands vertically as they drop from chest height to waist height.
SHOT 12 — Fight 2 / rising from the feet (5 s)
PHYSICS RULES
- Real-time speed. Slow motion is banned. Do not ramp, do not ease, do not stylize the tempo.
- One continuous take. No cuts, no camera stops, no speed changes.
ASSETS
@char_LI_indoors — Korean man, late 30s, deep burgundy silk shirt with sleeves rolled to the forearm, thin gold chain, olive pleated trousers, barefoot on the tiles. Hair slicked back with one strand loose at the temple.
@char_MEI — Chinese woman, early 30s, dark red striped mandarin-collar silk top, low-rise faded bootcut jeans, black ankle boots.
SHOT 12 — 5 seconds, one continuous shot.
The camera begins at floor level on the bare feet of @char_LI_indoors, filling the bottom of the tall frame, and rises vertically along his body over five seconds — feet, knees, hips, shoulders, face.
Mid-rise, at the moment the camera passes his hips, his right foot sweeps @char_MEI's forward ankle out from under her and she begins to drop out of the bottom of the frame.
The camera does not stop, does not tilt down to follow her, and does not change speed. It keeps rising to his face.
Real-time speed. Slow motion is banned.
SHOT 13 — Fight 3 / the takedown (6 s, one take)
PHYSICS RULES
- Real-time speed throughout. Slow motion is banned by name. Do not ramp, do not ease.
- One continuous take, no cuts.
- Every move is vectored. Nobody travels in a direction not written below.
ASSETS
@char_LI_indoors — Korean man, late 30s, deep burgundy silk shirt with sleeves rolled to the forearm, thin gold chain, olive pleated trousers, barefoot.
@char_MEI — Chinese woman, early 30s, dark red striped mandarin-collar silk top, low-rise faded bootcut jeans, black ankle boots, low bun.
@bead_curtain — red plastic bead curtain in the kitchen doorway, strands reaching the floor.
@lamp_kitchen — single bare warm bulb over the kitchen table, the only yellow light source in the film.
SHOT 13 — 6 seconds, one continuous take, no cuts.
ANCHORS — these are absolute:
- The camera looks ONLY at the kitchen. The kitchen is the only room that exists in this frame.
- @bead_curtain hangs in the foreground, red strands crossing the frame vertically, close to the lens.
- The action travels DOWNWARD toward the kitchen floor. Nobody flies. Nobody travels toward the hallway, the window, or the camera.
TIMELINE:
0.0-1.5s — @char_LI_indoors takes @char_MEI's right wrist and steps in behind her shoulder, staying on her outside line.
1.5-3.0s — he guides her forward and down along the circle her own momentum is already travelling on, one hand on the wrist, one on the elbow. He never lifts her, never strikes.
3.0-4.5s — she goes to the kitchen tiles chest-first, head toward @lamp_kitchen, feet toward camera, passing through @bead_curtain so the red strands scatter outward. He follows her down onto one knee, wrist still controlled.
4.5-6.0s — everything stops. Both are motionless. The bead strands keep swinging.
Handheld, chest height, the camera stays behind and slightly left of him and never crosses to her side.
SHOT 14 — The bouquet (4 s)
PHYSICS RULES
- Real-time speed. Slow motion is banned. The fall is fast and ugly, not graceful.
- Gravity is normal. The bouquet falls straight down, it does not float or drift.
ASSETS
@bouquet_peony — half-open peony bouquet, deep red, wrapped in cheap cellophane and newspaper, slightly crushed on one side.
@lamp_kitchen — single bare warm bulb over the kitchen table, the only yellow light source in the film.
SHOT 14 — 4 seconds.
@bouquet_peony falls into frame from the top and lands on the kitchen floor tiles beside the palm-sized halo of @lamp_kitchen. The cellophane crumples, three deep red petals break loose and skid.
Vertical composition: the bouquet lands in the lower third, the empty crushed-black upper two thirds of the frame hold nothing but the edge of the halo.
Camera static, floor level, macro.
SHOT 15 — The face opens (6 s)
PHYSICS RULES
- Real-time speed. Slow motion is banned.
- Photoreal skin: visible pores, subsurface scattering, natural oil sheen on the nose and forehead, fine flyaway hairs at the temple.
ASSETS
@char_LI_indoors — Korean man, late 30s, deep burgundy silk shirt with sleeves rolled to the forearm, thin gold chain at the neck. Hair slicked back with one strand loose at the temple.
@lamp_kitchen — single bare warm bulb over the kitchen table, the only yellow light source in the film.
SHOT 15 — 6 seconds.
Close-up of the face of @char_LI_indoors, centered in the tall frame, kneeling — the first time in the film his face is fully readable. He is breathing normally, not out of breath.
He looks down, off-screen. Recognition arrives: the eyes widen for a fraction of a second, then settle.
The expression does NOT become warm. It settles into something flat and finished. He has just made up his mind.
The bare bulb of @lamp_kitchen swings gently behind his head, sliding its palm-sized yellow halo across his cheek once and then off again — the only yellow in the frame. Everything else is cold teal-green.
He blinks once. Cut to black on the blink.
Camera static, 85mm, slow almost imperceptible push in.
The devices that repeat across these 15
So you don’t have to read every prompt, here is only what actually worked.
Name the props — register them as @door_tiger or @bags_grocery and the same object shows up when the cut changes.
Leave them unnamed and the tiger painted on the door disappears, or the number of grocery bags changes.
Write down what not to do — “slow motion is banned”, “there are only two hands”, “the door swings inward”. Video models slow fights down on their own and draw an extra hand. You have to block it in advance.
Pin the connection between cuts with numbers — write “the camera speed of shot 7 equals shot 2, exactly” and the two cuts join into one piece.
Stage 3 — Generate with Higgsfield
We fed the 15 designed shots in as written. The picture we wanted rarely came out on the first try, so the same scene was regenerated several times. Shot 3 was rewritten as described above and fed in again.
Stage 4 — Finish in CapCut
Generated clips alone don’t make a film. All of the finishing was done by hand in CapCut.
[Banner] Insomnia / Chungking Express style two tiers, placed in 3 spots (first appears 0.00~5.27s)
[Guide] safe-area guide box laid over the full length to check for clipped text
[Filter] Wong Kar-wai 0.00 ~ 73.80s, main film only
[Video] 21 segments 15 main cuts + black screen + 5 reused cuts
[Music] Noir (Acoustic) 0~82.47 / 82.47~91.10, split in two
[SFX] typewriter, electric hum 4 sounds in the black-screen section
[Narration] CTA_hongkong.wav 4 pieces from 82.47s
[Captions] Korean 11 · English 10 side by side in the outro
What we did here
Applied the color once. Instead of grading every clip, one Wong Kar-wai filter was applied only to the main film, 0~73.80 s. The outro has no filter, because it is cut from main-film clips that already carry the color and doesn’t need it again.
Trimmed the cut lengths by hand. Generated clips come out at a uniform 5.07 s, and only the first three were shortened to 4.70 / 4.80 / 3.50 s. If the opening drags, people scroll past.
Inserted 8.7 seconds of black. A line of text is typed over it with a typewriter sound underneath. Jumping straight from the film into the call to action feels rushed.
Did not generate a new outro. Five cuts from the main film (6, 7, 10, 13, 16) were re-trimmed and reused, cut short between 0.87 s and 3.33 s.
Laid captions in Korean and English together. 21 lines go into the outro only.
In short — what the AI made is raw material; turning it into a film was handwork in CapCut. 21 segments, 21 caption lines, 4 sound effects and 2 pieces of music were placed by hand. Generation is not the end.
Where we got stuck
Re-applying the filter throws the color off.
When attaching the outro we first tried importing the original clips again and matching the color, but the Wong Kar-wai filter lives inside CapCut and the same values can’t be reproduced outside it. In the end we switched to cutting from the main film that already had the filter applied. The color matches 100%.
To avoid the same problem next time, the order is now fixed: color grading happens once, at the very end.
Open the project file and the decision is still there. The Wong Kar-wai filter track covers one range only, 0.00~73.80 s, and the outro after it has no filter at all.
What we threw away
We generated far more than the 15 cuts we used. Fingers came out wrong, the camera moved differently from the intent, or the lighting direction didn’t match the cuts before and after.
Making it with AI doesn’t mean it comes out in one pass. Choosing the usable cuts takes up most of the actual working time.
