THE SIGN OVER THE CHECKOUT SAYS PLEASE PAY HERE. THE STORE DOESN'T EXIST
A store near the self-checkout area, with a sign above it that reads Please pay here and other shoppers queuing in the background. A young blonde with tattoos on her right arm stands in the middle of the frame in a very short light-blue off-the-shoulder dress and white sneakers with white socks, a big beige bag on her right shoulder and a clear bottle in her left hand. She laughs and looks into the camera.
No store, no queue, no checkout to pay at.
- Readable text in the background is a trap
A sign with real words has to stay spelled right in every frame, and a queue of shoppers has to stay a queue while the camera holds on one face.
The room behind her has plenty to get wrong. The article below shows face.json, the file that keeps the face in front of it right.
SHE SAYS SHE'S 62, FROM TEXAS, AND DUMPED FOR BEING TOO OLD. NONE OF IT HAPPENED
A field behind fences in the countryside. A blonde in a green lace-up top and unbuttoned short denim shorts leans on a round hay bale and talks straight into the camera: she is from Texas, she is 62, and her ex cheated and called her too old.
No field, no hay bale, no ex, no Texas. The woman, her age and her heartbreak are all generated.
- A confession needs a voice that holds
Telling a story straight into the lens means lip movement, breath and tone staying consistent through every sentence, and a face that carries the hurt without drifting into someone else's. The longer she talks, the more chances to slip.
The article below shows how her voice is cut from a single minute of audio, so a monologue this long still sounds like one person.
THIS WHOLE CLIP IS AI. DID YOU CATCH THE RED FABRIC WHEN SHE BENT DOWN?
A woman in a black leather jacket, black shorts and tall boots stands by a door in a modern hallway, then turns and walks toward a round table. She bends, pulls a red thong down her legs, steps out of it, and drops it into a black bag on the table. She straightens, walks toward the kitchen, and smiles.She isn't real, and neither is the hallway.
-The bend is a clothing and hand test
As she folds at the waist the red fabric has to leave the body and stay in her hands while the jacket, shorts and boots keep their shape. One seam that vanishes or a hand that loses the string and the viewer rewinds before he can say why.The bend and the drop have to be the same motion every time. The article below opens motion.md, the file that holds it.
THE DEFENDER GETS ONE GESTURE IN THE AIR. THE CAMERA GETS ANOTHER ON THE GROUND
A basketball court at night. A girl in a black top and skirt takes off for a very high jump and dunks from above, gesturing at the defender while she is still in the air. She lands, smiles and makes a heart for the camera.
No court, no hoop, no defender to beat.
- Mid-air is where a body can't hide
There is no ground to hide a bad pose behind. The takeoff, the reach to the rim and the gesture have to belong to one body, and the face that smiles after the landing has to be the same one that was in the air.
The article below shows motion.md, the file that holds a girl's way of moving, so the body that lands is the one that took off.
THIS WHOLE CLIP IS AI. DID YOU CATCH THE DRESS WHEN SHE GOT OFF THE BIKE?
A woman in a white off-shoulder dress sits on a bicycle outside a 7-Eleven. She swings her leg over, stands, and walks toward the door. An older man in a polo shirt watches her, then looks down at the bike and leans in to examine the seat post. She disappears inside while he stays bent over the frame.She isn't real, and neither is the storefront.
-Getting off the bike is a dress and balance test
As she dismounts, the dress has to keep its shape and the hem has to stay at the same length. The bike has to stay upright until she is clear of it, and the man has to stay in place. One fold that changes or a wheel that moves on its own and the viewer rewinds before he can say why.Getting off the bike has to be the same motion every time. The article below opens motion.md, the file that holds it.
TIGER STRIPES, A WAIST CHAIN AND A BRACELET HAVE TO SURVIVE A WALK DOWN THE STAIRS
An indoor staircase faced with red tile and white grout, pale walls around it. A girl with long dark hair, in a long fitted halter dress with a tiger-stripe print, a decorative chain at the waist and a bracelet, walks down the steps toward the camera and smiles slightly at the end.
The staircase, the tiles and the dress are all generated, stripes included.
- Stairs break a walk into repeated steps
Each step is a small drop and a small landing, and the dress, the chain at her waist and the bracelet all have to answer every one of them. The stripes have to wrap around a moving body instead of sliding across it, and the tile joints behind her have to stay straight.
A flight of stairs shouldn't turn her into a different walker. The article below shows motion.md, the file that prevents it.
THE CAMERA SEES HER BACK FIRST. THE FACE COMES LAST
A theme-park plaza in bright daylight. A girl in a white and mint-green bodysuit with gold chains and tall boots poses with her back to the camera, then turns slowly over her right shoulder and smiles into the lens.
Nothing here was visited. The park is generated, and so is the turn.
- A turn is where a generated girl can become someone else
From behind she is only hair, shoulders and a costume. Over the turn the face has to arrive and match everything the viewer imagined from her back, and the chains and boots have to travel with her the whole way.
The article below opens face.json, the reference that decides what her face looks like before the camera ever turns round.
WATCH HER FACE WHEN THE FOLD GOES ALL THE WAY DOWN
A gym corner: a purple wall, a panel of small silver tiles, a piece of equipment at the edge of the frame. A blonde with long hair, in a light blue top, taupe shorts and white sneakers, stands with her feet wide apart and folds forward until her head hangs upside down toward the lens, her hands reaching for the floor. The camera sits low, behind her.
No machines, no mat, no wall. The flexibility was written, not trained.
- Upside down, every feature changes direction
Hair falls toward the floor, the eyes sit below the mouth, the jaw points at the ceiling. A face that is only half learned can turn into a stranger from this angle. Here she has to stay the same girl the viewer would recognize in any other clip.
The article below shows face.json, the file that keeps one girl's face hers even when the camera catches it from below and upside down.
ONE ORDER. ONE CAR WINDOW. ONE LINE OF DIALOGUE. NONE OF IT EXISTS
A sunny drive-thru lot: palms, a bed of purple flowers, a menu board, parked cars and a tall restaurant sign behind. A brunette with bangs, big black glasses and gold hoops, in a black knotted tee, a very short light-gray tie skirt and white knee socks, stands with her back to us by a palm and bends to pick up a takeout bag, a large cup with a straw in her other hand. She turns and walks straight at the camera, smiling. The camera sits inside a parked car and shoots through the window. At the glass she says, "Hi sir, I have your tacos and your lemonade. Have a good one!" and hands the bag and the cup into the car.
No lot, no car, no order. The voice is generated too.
- A car window is a hard frame
A face walking straight at the lens, glare on the glass, a hand reaching into the shot. She has to look the same from the palm to the window, and her voice has to sound close, like someone leaning in.
The article below shows how one girl's voice is made from a minute of audio, so a line like "have a good one" comes out as hers.
HER VOICE HAS TO HOLD UP WHILE SHE IS OUT OF BREATH
A white motorboat on open sea, an overcast sky, a green hilly island in the distance, a white cooler on deck. A small dog peeks out from behind the legs of a short-haired brunette with tattooed arms and neck, in a gray crop top, a brown tie skirt and a blue cap. She braces against the side with a black rod and heaves it back and up, breathing hard, smiling and answering a man's voice from off-screen. Then she clamps the rod under her arm and keeps cranking the reel.
No boat, no sea, no island. Both voices are generated too.
- Talking is easy. Talking while hauling is not
Short breaths between words, a smile on top, a rod pulling the other way. The sound has to follow the effort in her body and still be recognizably her.
The article below shows how one girl's voice is made from a minute of audio, the same voice at rest and out of breath.
SHE DISAPPEARS BEHIND A GAME BOARD AND WALKS OUT THE OTHER SIDE
A sunlit pedestrian avenue paved in light tile, leaf shadows on the ground, green trees, ornamental street lamps and a white building with a tall tower in the distance. In front, a giant outdoor four-in-a-row game with a dark blue frame and holes ringed in yellow and green plastic. A blonde with long hair, in a short white short-sleeved dress and beige high-heeled sandals, stands behind it, smiles at the camera and raises both hands to fix her hair. Then she turns right and walks around the back of the game, hidden for a moment. She comes out on the left side and strides off down the avenue, still smiling.
No avenue, no tower, no game. She never walked past any of it.
- The face has to survive the hiding spot
For a moment the board covers her. When she steps out the other side, the dress, the hair and the face have to match the girl who stepped in. That is where a generated character can easily turn into someone else.
The article below shows face.json, the file that keeps one girl's face hers from one shot to the next.
SHE SENDS THE BALL BACK WITHOUT TAKING OFF HER HEELS
A beach at sunset, a packed-sand path in front, goals, the sea and a group of guys playing behind. A brunette with long dark hair in a beige crop top, a very short white skirt and heels with clear straps, a small black bag under her right arm, walks toward the camera looking at the phone in her left hand. A ball rolls up to her feet. She stops and turns, one of the guys shouts "Sorry, can we have it back?", and she sends it to them in a high arc with her right foot. Then she turns away, smiles a little, goes back to her phone, and they shout after her, "Thank you, have a good day!"
No beach, no match, no ball. The shouts are generated too.
- The walk before the kick and the walk after it
Heels on packed sand, one leg swinging, a ball leaving the foot in a high arc. Then she turns and goes on as if nothing happened. The body has to land back in the same walk it left, with the same pace and the same hold on the phone.
The article below shows motion.md, the file where one girl's way of walking and moving is written down once, so the walk is hers in every clip.
SHE ASKS ONE QUESTION AND THE HOST LEAVES THE STAGE
A late-night studio, a painted night ocean with a huge full moon behind a wooden desk. A host in a dark suit sits at it. A brunette in a white corset top, silver shorts and tall white fur-trimmed platform boots asks, "Can I show you something?" He says yes. She walks to the middle of the stage, goes down on her hands, spins, then drops into a headspin with her legs in the air. The host watches with his mouth open. She lands on her knees and smiles, and he jumps up and runs backstage.
No studio, no host, no moon. The voices are generated too.
- Her spin and his face have to be timed together
The joke only works if the host reacts at the right second: his jaw drops while she is still turning, not after, and the run offstage waits until she has finished. Two characters, one line of dialogue, one timeline.
The article below covers how a girl's voice is built from a minute of audio, so a line like the first one in this clip sounds like her every time.
HER REACTION TO THE FALLING PLATE TOOK ZERO TAKES
A gym camera view marked "Camera 01", other people and racks in the background. A blonde in a fitted white long-sleeve crop top, white shorts and white sneakers stands at a purple bench holding a small barbell. She turns sharply and a plate slides off one end. She looks surprised, bends down and sets the bar with the remaining plate on the bench. At the end she straightens up, spreads her hands and looks at the equipment, embarrassed.
No gym, no camera, no plate. It only looks like security footage.
- A falling plate has to fall like a plate
A sudden turn, a plate leaving the bar, then a human reaction: surprise, a careful bend, a shrug. The object, the hands and the face all have to agree with each other, from the first frame to the last.
The article below shows face.json, which keeps one face through surprise, embarrassment and everything else a script throws at it.
THIS GOLFER HAS NEVER BEEN ON A COURSE
A green fairway ringed by trees. A blonde in a fitted brown long-sleeve crop top, a short black skirt, white sneakers, a black visor and a white glove on her left hand lines up a driver over a white ball. She swings, finishes balanced on the toe of her right foot and follows the ball with her eyes. Then she lowers the club into her right hand and walks across the grass toward the camera.
No course, no ball, no club. She has never swung one.
- The walk toward the camera is the real test
A swing is fast and her face is turned away from us. Then she walks straight at the lens and the face gets bigger frame by frame. That is where a face either holds or turns into someone else.
The article below shows face.json, the file that keeps her recognizable from a full swing to a close-up.
FOLLOW HER FEET FROM THE PLATFORM TO THE HORSE
A circus ring with a sand floor, stands full of spectators. A blonde acrobat in a lilac crop top and mini skirt, white socks, sneakers and a headband stands at the edge of a high platform. She jumps, tucks into a forward flip and lands on a small trampoline, bounces onto a pad on an elephant's back, walks forward and steps onto its raised trunk. From there she flips backward onto the same trampoline, bounces again, turns over once more and lands standing on a black horse galloping past, right arm up.
No ring, no animals, no crowd. Nothing in this clip ever left the ground.
- Six landings, six different surfaces
Platform, trampoline, elephant, trunk, trampoline again, horse. Every landing is on something else, and her body has to arrive on each one the way it left the last: same height, same balance, same rhythm.
The article below explains how one girl's way of moving is written down once, so every clip she appears in can reuse it.
NO CAMERA. NO ACTORS. NO FARM
Sunny field, a sunflower-shaped sign with a round hole in the middle, pumpkin leaves behind it, tall corn at the back, a scarecrow in overalls and a green tractor in the distance. A blonde in a white top and an olive-green skirt walks past the sign with her back to us, bends down and picks a long striped squash off the ground. She turns sideways to look at it, then to the camera, fixes her hair with her left hand and walks back smiling.
Even the squash in her hand was never grown.
- The sign in front is a test of its own
She passes behind the sign, shows up on the other side, picks up something solid and turns. The sign, the girl, the squash and the field each have to stay on their own layer.
The article below has a section on what kills a clip like this, and the check that catches it before it gets posted.
WOULD YOU HAVE SCROLLED PAST THIS? WATCH WHERE THE ORANGE BALL GOES
Bowling alley, blue neon over the lanes, a hockey game on the screens above. A blonde in a light blue top and a white skirt takes an orange ball from the return, glances back over her shoulder with a smile, then walks to the foul line and sends it down the middle. She stands with her back to us and watches. Strike. At the end she turns to the camera, laughing, and walks back.
No lane, no pins, no one who ever threw that ball.
- Her face has to show up twice and match
She glances back at the start and turns to the camera again at the end. Between those two moments she has her back to us, a ball rolling down the lane and ten pins falling. The face that comes back has to be the same one that left.
The article below covers face.json - the file that holds who she is - and what sits next to it.
THIS WHOLE CLIP IS AI. DID YOU CATCH THE UCLA LOGO WHEN SHE SAT UP?
A girl with dark hair lies on her stomach on a dark blue blanket in a sunny park, kicking her legs. Around the five-second mark she flips over, sits facing the camera as it zooms in, and holds an iced coffee and a yellow marker. She looks around, then breaks into a wide smile and waves. In the last seconds she shyly fixes her hair and the two-tone UCLA Bruins cap, still laughing. Trees, a brick building, and other people fill the background.
She isn't real, and neither is the park.
- The flip-to-sit is a face, scale, and logo test
As she turns and sits upright the camera closes in. The cap logo has to stay sharp and correctly shaped on the moving head, the skirt and long-sleeve have to hold their folds, and the smile has to lock without warping. One stretched letter on the Bruins mark or a face that drifts during the zoom and the viewer feels it before he can say why.
The article below shows how a character's motion is locked per girl, and what a locked character earns.


