WorkAI film

Innocence

Childhood, seen through a child’s own eyes. Thirty-five seconds, twenty shots, every frame generated and then cut like a film.

Released
An excerpt: the memories, without the opening and closing shots. There is no speech.
My role
Director: concept, shot list, prompts, selection and edit
Length
34.9 seconds, 20 shots
Picture
4:3 for the memories, inside a widescreen opening and closing shot
Models
Kling 3.0 and Seedance 2.0 for motion, image models for the start frames

What it is

A man closes his eyes. What follows is what he remembers: eighteen moments of childhood, each seen from exactly where the child stood. The aisle of a school bus. A crayon dinosaur. The top of a slide. A teacher crouching to say hello on the first day. Birthday candles. A lost tooth in the mirror. The view from dad’s shoulders. Then he opens his eyes.

I directed it: the idea, the shot list, every prompt, every choice of take, and the cut. The pictures were generated with Kling 3.0 and Seedance 2.0. The whole film was shot, if that is the word, in one day.

It is short on purpose. The average shot is under two seconds, the way a memory arrives.

In numbers

  • 20shots in the cut, 1.74 seconds each on average
  • 49video takes generated: 43 with Kling 3.0, 6 with Seedance 2.0
  • 221start frames generated to get 20 right
  • 199 sof footage generated for a 35-second film
  • 1 dayJune 11, 2026, from 9:36 in the morning to 9:54 at night
  • 7 takesfor the crayon shot, which runs half a second
  • 45start frames for the school bus alone
  • 18memories between one man closing his eyes and opening them

Every figure on this page is counted from the project itself: its code, its history and its files.

How it was made

One rule, two steps per shot, and a lot of takes on the floor.

  1. One rule: the camera is the child

    Every memory is strictly first person. The camera is the eyes of a young boy. We see what he sees, at his height, and the only part of him ever in frame is his own small hands.

    Models do not want to do this. Ask for a child’s birthday and you get a child, seen from across the table. So the rule sits at the top of every prompt, and it gets louder as the day goes on. It started as one line, “camera at 4 foot 8 height, point of view”, and ended as a paragraph that opens in capitals. 111 of the image prompts use that final template.

  2. Frames first, then motion

    Each shot is made in two steps. First a still: the first frame of the shot, generated and regenerated until the height, the hands, the light and the room are right. Then that still is handed to a video model with a short instruction for what happens next: walk down the aisle and the bus does not move, blow out the candles one by one.

    Almost all the work is in the stills. 221 were generated for 20 shots. 48 of the 49 video takes started from one.

  3. Getting one shot right

    The school bus opens the memories, so it had to be right. It took 45 start frames. The first bike ride took 29, across four rewrites of the prompt, before the handlebars, the path and the evening light agreed with each other.

    The crayon dinosaur was the stubborn one. The model kept adding marks to the drawing while the hand moved, so the instruction became “he draws nothing on the image”, and two of the takes pinned the last frame as well as the first. Seven takes. The shot runs half a second in the film.

  4. The cut

    199 seconds of footage were generated. 35 are in the film. The memories are in a 4:3 frame, like a family camcorder, and the two shots of the man are widescreen, so the picture itself tells you when you are inside his head.

    There is no title card and no words on screen. The order does the work: bus, crayons, LEGO, a butterfly, the slide, the first day of school, Halloween, candles, the high dive, the barber’s mirror, a torch under the blanket, the bike, a smiley drawn on a rainy window, a party watched through the banister, the tooth, dad’s shoulders, grandma reading, the swing.

The hard parts

Holding the point of view

What happened
The models kept putting a visible child in the frame, or drifting to a view from across the room.
What fixed it
One fixed opening paragraph on every prompt, and instructions that say what must not appear: no other face, no third-person view, only his hands.

A hand that draws when it should not

What happened
In the crayon shot the model kept adding lines to the drawing.
What fixed it
Telling it what does not happen, and fixing the last frame as well as the first. Seven takes for half a second.

The same place in two frames

What happened
A shot that starts in one bus and ends in a slightly different one breaks the memory.
What fixed it
The approved first frame goes back in as a reference, with the instruction to match it.

Somebody else blows out the candles

What happened
At a birthday table, the model lets everyone join in.
What fixed it
The prompt says who does it, and says it again as what nobody else does.

The shot list

In film order, with the number of video takes each one needed.

Every shot in Innocence, when it starts and how many takes it took
ShotStarts atVideo takes
A man closes his eyes0:001
School bus aisle0:023
Crayon dinosaur0:037
LEGO castle0:041
Butterfly in cupped hands0:062
Playground slide0:076
First day of school0:093
Halloween doorstep0:114
Birthday candles0:133
High dive0:142
Barber’s mirror0:161
Torch under the blanket0:171
First bike ride0:182
Smiley on a rainy window0:202
A party, through the banister0:222
Lost tooth in the mirror0:234
On dad’s shoulders0:251
Grandma reading0:281
The swing0:313
He opens his eyes0:32–

How I worked

This is directing with a different kind of crew. I decide what the shot is, where the camera stands and what is in the frame. The model offers takes. Most are wrong in a way worth learning from, and the prompt for the next one says exactly what the last one got wrong.

What makes it a film and not a pile of clips is the selection and the cut: 49 takes down to 20, and about one second in every six generated making it to the screen.

Made with

Motion
Kling 3.0 and Seedance 2.0, through Higgsfield.
Start frames
Image models, one still per shot, regenerated until right.
Edit
Twenty shots, cut to 34.9 seconds at 60 frames per second.