Skip to content
Oedipus King of AI

The making

Cinema Code

La fabrication, code et image

Once we questioned the Gods. Today, we code the Oracle.

Today, the artificial intelligence models that generate images and voices are distributed across a few providers: Runway ML, ByteDance, Kling, Google, OpenAI, ElevenLabs and others. Each excels in a domain, none covers the whole. Without orchestration, you spend your day juggling and tidying files. Not prompting.

So I coded a tool: Squiddle. A homemade production platform that creates the rushes, talks to DaVinci Resolve, retrieves the content, and orchestrates post-production and visual effects.

Squiddle — Canvas view, generation graph for one scene
Squiddle — Canvas view. Generation graph for a scene (characters, prompts, derivatives).
Squiddle — Timeline view, post-production batch
Squiddle — Timeline view. 633 shots, post-production batch synced with DaVinci Resolve.

A 65-minute film, created by one person, would not have held without this. The processing chains are long: prompt → image → image edited to fix characters → image turned into a start frame for a video model → video dubbed with a voice → voice lip-synced → SFX. Multiplied by the hundreds of shots in the film. Without code, impossible.

Co-creator, not co-director

The director's central task is staging: shot breakdown, angles, the camera's position and movement. And that is precisely what AI does worst. If I let it propose the shots, it decides in my place — it co-directs. But I have no co-director: I have a co-creator, building each shot with no memory of the others. The director is me.

So I had to code my own tools, with Claude Code, to move the camera and cut my edits the way I want — and, rather than hide the inconsistencies, to make a style of them.

Ideally, there would be a unified tool — a blend of Blender, DaVinci Resolve and image and sound generation gathered in one place. That complete solution does not exist: we only have bricks. Until a universal tool arrives — and that will take a while — better to build your own. That is the whole point of agentic coding.

For the real challenge of a feature film is continuity — of dialogue, of dramaturgy. An ensemble film, where characters cross paths without answering one another, would be simpler; a true dramaturgy demands these tools, and demands building them yourself.

AI time

January 2025 → June 2026. During the making, the models mutated several times. February's technique is not November's. Instead of fighting it, you embrace it: you redo, you regenerate.

AI regenerations are non-destructive. You can create a shot in draft, recreate it clean, reject it, retry — as many times as you want, without ever degrading quality. Classic shooting, by its sheer weight, forbids this — or at least narrows the options. Writing allows it. Making an AI film resembles writing: you spend most of your time rewriting, right up to publication.