Our stack for AI filmmaking: what actually works in mid-2025

Everyone asks which model we use. It is the least interesting question. Models change monthly; what survives is the plumbing around them. Here is the shape of our stack, minus the hype.

One queue to rule them all

Every generation — image, video, voice, music, SFX — goes through a single job queue with priorities, retries and budgets. Artists never wait on a spinner; they brief, move on, and review when the work comes back. Idle GPUs at night are a bug, so the queue renders around the clock.

Providers are adapters, not religions

Each capability sits behind an interface. When a better video model ships, we swap an adapter, not a workflow. This sounds obvious; it took us a year of discipline to make it true.

QC is a model too

Vision models review renders against the written brief — framing, subject, palette, artefacts — and score them before a human ever looks. Humans arbitrate the interesting failures instead of triaging the obvious ones.

The boring parts win

Asset naming, brief templates, review rituals. None of it demos well on a keynote stage, and all of it is why the work ships. The stack is not the moat. The habits are.