The People Are Awake

Step Printing

Most of what I like in an image is a flaw. Grain, smear, misregistration, the stutter of a frame held too long. Wong Kar-Wai and Christopher Doyle built a whole visual language out of one of them: shoot undercranked, print each frame several times, and motion tears while the frames themselves visibly hold.
Generative tools are built to remove exactly this. They’re optimised toward the clean, the resolved, the average. So the question I was interested in wasn’t whether a model could imitate a look, but whether it could be taught an imperfection the mechanism rather than the result.

That meant training a custom LoRA for MiniMax rather than using anything off the shelf and building the generation pipeline around it in ComfyUI — so the technique is part of the model itself, and every shot can be directed rather than filtered.

Video

Sound on | Fullscreen
⬑ Process / The People Are Awake
[21. September 2026]

Helpful Resources