The AI filmmaker's field guide.

Every few weeks a director asks us how to start making films with models. The honest answer is short: stop thinking about the model. The model is the camera. Nobody ever made a film about the camera.

What follows is the working document. It assumes you can already direct. It exists because the craft questions are the same as they have always been, and the tooling answers are what changed. It is also, not coincidentally, a description of how AIM PREVIZ, the studio's previsualization platform, is built. Every rule below exists because the platform enforces it.

Start with the shot, not the prompt

A prompt is a wish. A shot is a commitment: subject, lens, light, movement, duration. The directors who ship are the ones who arrive with a shot list, not a moodboard. Write the film in setups before you write a single prompt. Wide, medium, close. What the audience learns in each. What each costs the story if it fails.

The reason is practical. Models return their best work when the ask is cinematographic. "Slow push past a frozen ridgeline, 32mm, blue hour" beats "epic arctic landscape" every time, across every current video model. Speak camera. The machine is fluent.

Determinism before dazzle

The first thing to establish on any AI-native production is repeatability. If a frame cannot be regenerated on demand, it is not a plate, it is a miracle, and miracles do not survive a conform. Capture the seed, pin the checkpoint, fingerprint the references. This is not exotic. Diffusion sampling has been deterministic-on-seed since DDIM [1]. The discipline of recording it is the whole job.

A shot you can rerun is an asset. A shot you cannot is a screenshot.

Character is a system

One good frame of your hero means nothing. The film needs the same face at dawn, in rain, screaming, three scenes apart. Treat the character like a product line: a locked reference set, a written description of non-negotiables (jaw, eyes, hands), and a test battery you rerun every time anything in the pipeline changes. When a new model version lands, the battery tells you in an hour whether your character survived.

Recast nothing mid-picture. Lock the character the way you would lock an actor's contract.

Sound is half the film

Audiences forgive a soft frame. They never forgive thin sound. Budget sound design from the first animatic, and cut the picture to a real temp track, not silence. Every trailer we ship is mixed before it is graded. The emotional math of a scene lives in the low end and the room tone, and no video model will hand it to you.

The legal frame

Keep a paper trail the way a laboratory keeps a notebook. What was generated, with which model, from which references, under which license. Enterprise clients now ask for this at bid stage [2], and the studios that can answer in one email are the ones getting the second job. Clean inputs, logged outputs, no scraped likenesses. Boring, and decisive.

The takeaway

The model is not the film. The film is intent, repeatability, character, sound, and paperwork, executed in that order. Master those five and any model that ships next quarter is just a better camera.