UES PENELOPE // CANTEEN B-14 // CAM 215 // SLEEPER AI, EPISODE FOUR
Sixty seconds of film in two days: a chase, a canteen, a goddess who sees through cameras, and a girl who blinds them. Fourteen video generations for nine scenes, and most of the lessons came from the ones that did not ship.
The finished film, sixty-one seconds. Watch it first: everything below is what it took to get here, and most of it did not make the cut.
Open on YouTube · episode four of Sleeper AI · episode three's diary
Act 1 · the run
Episode four opens where episode three ended: after the "Boo", she runs. The start frame already existed, the running plate from the ep3 look tests. The first clip moved beautifully and had one problem Alex spotted at once: the low camera behind her pushed her hips wider than any reference.

Never animated in episode three. A motion reference that sat unused for a week and became the first second of the next film.

Slimmer, then a higher camera. The slim edit went too far ("now hips are too narrow") and the raised camera read like surveillance but lost the drama. "I liked the old one better."

About ten percent, original camera. The edit asked for a subtle change and attached a body reference with the right proportions. The animation prompt said "natural athletic build", not "narrow hips".
A second run toward the camera in the mushroom corridor was planned as a straight line. Alex changed it: she bursts out of a side door, almost falls, looks both ways, then runs at us. The lamps above her light up one at a time as she passes, the first sign that the ship is following her.

The door landed closer than asked. Better: her face is bigger. Near lamps dark, far lamps lit, so the animation only had to switch them on.
Act 2 · the canteen
The old crew canteen from episode two, a tree grown through the long table. Alex's staging: she runs in, hides under the table, and only then does the lamp above her switch on. Then he added a camera flying around the table while she dives under it.

Perfect staging, the source plate's look. The ep2 room was rendered sepia, and the edit followed it instead of our cel register. No green anywhere.

Restyled, composition untouched. Flat cel, green mushrooms back, the lamp the only orange.

A different angle on purpose. High and wide from the entrance, so the model has to fly the camera down and around to reach the last frame.
Measured, not assumed: the fear was that a lit last frame would make the model fade the light up across the whole clip. Brightness per frame said otherwise: on, off, on, off, on, in five frames, then it holds. Then a new trap: converting to twos (every frame held twice, the hand-drawn look) would have eaten that flicker, so the last second of this clip runs at full frame rate.
Act 3 · the goddess
After the lamp, cut to the one who switched it on. Alex asked for a zoom out, and then for a real camera move instead of a digital zoom on a still. To stop the model inventing her body, both ends were pinned: a new close-up of the lens as the first frame, the canonical episode two plate as the last.

One lens, no face. The series rule held. The red light in the centre is her colour.

No red in this lens. So the model faded the light out on the way back. Accepted.
Rule written into the series bible: twos are for figure motion only. A camera move on twos judders, because the whole frame steps at twelve frames a second.
Act 4 · eyes everywhere
The plan for the second half was to calm down. Alex refused: "there is no reason to calm yet. How can the Goddess see her? She can't, only through the cameras." So the voice explains itself with a slideshow of cameras, and the last one zooms into a lens.

Brutal, as asked: concrete, metal, unknown machines. "A little too much of cameras" was the note. At a second on screen, density reads as "everywhere".

The Evangelion shot. Rings of eyes, built to watch.

Cut down to three cameras after the note. They grew into the bark like knots. The lowest one is the camera she will blind.
Alex asked for one more beat: a close frame of her under the table, hand over her mouth, and then something changes. She lowers the hand, breathes, and stares into the lens for a few seconds. In the cut it answers the zoom: the camera looks at her, she looks back.

The strongest face frame of the episode. "No 3D render look" said twice in the prompt.
Act 5 · the extinguisher
The first plan was a smash. Alex had a better one: a fire extinguisher, foam on the lenses, and one clip from the camera's point of view turned into three feeds with different crops and offset timing, so they do not show the same moment. Then he split the clip in two, because crawling out and spraying is too much for one render.

The extinguisher planted in the plate. So the grab has somewhere to go.

One frame shared by two renders. The join between them is invisible by construction.
Monochrome, scanlines, the red REC dot and the labels are all post. The clips were rendered in colour so her face stays on model; a grey start frame would have drifted her.
Act 6 · the fourth camera
Over the three white feeds a fourth one flickers on, a camera behind her. She turns and destroys it. The idea took one plate. The execution took three videos and a correction on how people actually hit things.

"There were three cameras, right?" Only one was foamed.

"She sprayed it a lot." Three cameras, not enough mess.

Foam on the table, drops on the floor. It reads a little like snow. It passed.

The face was fixed and then rejected anyway. "Right now, none of the images we have are her." The image model's close-up face had drifted from the girl in the clips.

Her face, pulled from the video we already had. Alex picked two: the steady stare and the determined look up at the camera.

The impact itself as the last frame. Left hand under, right hand pushing, cold and in control. The model then had to animate a real hit.
She raised her arms too high. Who smashes cameras like that? If I needed to punch something hard with a fire extinguisher, I'd use a thrust. Left arm holds it from below, the right one does the pushing. She calmed down and got the situation under control.Alex, directing the third take
Act 7 · the ending
She walks out of the canteen and onto a catwalk above the ship's spine, a hollow megastructure with forest on its bridges and a tower far away. Two clips, two defects, and one of them taught the most useful rule of the episode.

Alex's angle. Foam and mushrooms huge in the foreground, her walking away small.

A good image that disagreed with the start frame. Made on its own, then used as the last frame for an empty-catwalk first frame. The railing did not match.

An edit of the empty catwalk that only adds her. The railing is the same pixels in both frames; checked along both rails before paying for the clip.
Rule for first+last renders: if the two frames disagree about the geometry, the model will find a path through it. Build the end frame as an edit of the start frame, place the character where nothing has to cross her, and check the overlapping parts before the render. A patch can hide the crossing but not remove it.
Act 8 · her voice
The same actress as episodes one to three, sad and whispered. Alex wrote most of the words. The new idea was that she speaks two ways: close and clean, and through the ship's speakers, which had to sound like speakers. The first speaker version "sounds almost the same"; the second is narrow, clipped and crushed.
| At | Line | Voice |
|---|---|---|
| 18.4s | You cannot hide from me, honey... I can see you. | close |
| 24.5s | I have eyes... everywhere. | close |
| 30.9s | So... how about you stop running, and we talk? | speaker |
| 38.6s | What are you... | close |
| 40.9s | Oh, great... | close |
| 43.6s | You missed one. | speaker |
| 49.3s | Humans... | close |
| 53.6s | Go on, honey... I will not stop you. | close |
| 58.4s | ...Not yet. | close |
The transcriber was the problem, not the takes. The small speech model heard "airbait", "start running" and "I can see her" in whispered takes. The next size up read all fifteen correctly. Whispered lines are now checked with the larger model.
Act 9 · sound and subtitles
Most of the film was silent until the end. The sound bed is procedural, twenty-two stems from one script: the ship's hum, cicadas and drips, lamp clicks and mains buzz, the iris servo, camera relays, the spray, the crack of the lens, a sub hit on the black, wet steps and steps on metal grating. Then subtitles, because most people meet a short with the sound off.

"I like yellow." The speaker voice became amber Courier capitals in brackets, over a soft dark band so it survives the pale feeds. Her close voice stayed thin white with a faint red glow.
| Item | Detail | Cost |
|---|---|---|
| Fourteen video generations | Seedance 2.5 via evolink, 720p; five kept as they came, three killed, three fixed in post | $25.12 |
| Twenty-three image plates | Every one an edit with references | $2.93 |
| Thirty-one voice takes | Nine lines, one actress | $0.29 |
| Slides, collage, patches, subtitles | ffmpeg, Pillow, a few masks | $0.00 |
| Twenty-two stems of sound | numpy and scipy | $0.00 |
| Total | Episode three $22.38 | $28.34 |
The overage is the smash and the catwalk: two killed smashes and the railing clip are $5.61 of the total. Every free fix in this diary replaced a paid re-render; two of them (the lamp, the rails) were attempts before paying anyway.
Go on, honey... I will not stop you. Not yet.Where episode five starts