The Amber Square

The method · 02

One face, one building

The hardest problem in AI video is not making a good shot. It is making the fiftieth shot look like the same person in the same building as the first.

One face

Lex's face was designed once, as a set of approved reference portraits, and every shot of Lex is generated from them. The eyes have three approved states (dark, ember, full amber) because the story needs them to change on cue.

The three eye states on one face sheet.
The eye-state sheet the shots are checked against.

One building

Image and video models invent whatever is out of frame. That is fine for a single shot and fatal for a camera move through a building: at the end of the film the camera drops through the floor into the room below, and the opening in the floor kept changing size, or doubling, from one take to the next.

The fix was to stop asking the model to imagine the building and give it one. The rooms were built as simple 3D sets in Blender. The first and last frame of a move are rendered from the same set, then restyled, and only the motion between them is left to the video model.

Blender render of the lower hall from above, a production line of yellow and orange robots.
The lower hall set, camera above the line.
Blender render of the lower hall facing the lift doors and a counter.
Same set, reverse angle.
The lower hall in the finished film.
The same room in the finished film.