First AI Short Film: A Step-by-Step Workflow

A first AI short film usually fails for reasons that have nothing to do with the model. The premise is too large for the runtime, the script asks for a crowd and a handshake in the same shot, the shot list is really a mood board, and the edit begins before anyone has watched a full sequence from beginning to end. The tools are the least mysterious part. What follows is one workable route through the whole pipeline, written for a film of two or three minutes with one or two characters and a handful of locations. Treat the numbers as defaults to adjust, not rules.

Decide what the film is before you open a tool

Write a logline of no more than three sentences, then write the film's single turn: the moment where something changes. A short film can survive on one turn; it rarely survives on three. It helps to name the constraint out loud before production — a single room, a single night, one actor's face — because constraints are what make a generated film look deliberate rather than assembled.

Keep a one-page brief with four lines: the premise, the turn, the look, the runtime. Everything after this step is judged against that page.

Write a script that respects the tools

Write the script as if you were producing it with a small animation team, because that is closer to the truth. Some current video models return clips of five or ten seconds, and the Runway Gen-4 video guide frames each generation as a single scene rather than a sequence, so scenes that need three actions in four seconds will be harder to direct than scenes that need one.

In practice, write:

• dialogue scenes with people sitting, standing or walking, not embracing;
• reactions rather than stunts;
• inserts that carry information — a phone screen, a door, a letter — because you can cut to them;
• sound that will carry the scene, because audio is easier to control than faces.

Then cut the script by a fifth before you generate a frame. First films are almost always too long, and the cheapest place to shorten them is the page.

Build a shot list and a lookbook

A shot list is a plan for the edit, not a wish list for the trailer. Use one line per shot: number, size (wide, medium, close), subject, action, camera move, duration. StudioBinder's walkthrough of the process is a useful reminder that the list exists to be shot, checked and revised, not to be admired. If you want to go deeper into sizes and movement, the chapter on how AI video models work is a better starting point than a glossary.

The lookbook is separate: six to ten still images that establish palette, contrast, lens character and texture. These are not the film; they are the reference set that keeps separate generations in the same world. Choose images you would be happy to match, because the model will.

Test one shot all the way through

Before committing to a look, take the shot at the centre of the film and carry it through the entire pipeline: generate it, select a take, upscale it, cut it against a second shot, add a sound effect, export it. This is a test of the workflow, not of the model. It will show you where the delays are — usually in selection and re-generation, not in the first render.

Decide two things during this test: which model you will use for which kind of shot, and how you will name files. Both decisions are cheap now and expensive later. A shot that cannot survive the pipeline is a shot to rewrite, not to generate eleven more times.

Generate in passes

Generate in three passes rather than shot by shot.

1. A blocking pass: the geometry, the camera and the light, with performance good enough to judge whether the scene reads.
2. A performance pass: faces, timing, gestures, the small things that make a take actable.
3. A texture pass: inserts, cutaways, atmosphere, anything whose job is to cover a cut.

Keep a prompt log with the model, the settings, the reference images and the exact prompt for every approved take. When a shot is locked, stop touching it. The most common way to lose a week is to improve one shot while three others wait.

A shot the model cannot finish is a shot to rewrite, not a shot to re-roll.

Assemble, then cut for performance

Import everything into the editor and assemble a rough cut before judging any individual shot. DaVinci Resolve's free training material covers editing, colour and sound in one place, which is convenient when the same person is doing all three. Cut for the performance you have, not the coverage you wanted. If a face breaks on the fourth second, cut at three and let the next shot carry it. Generated footage rewards cutting early; it punishes holding a shot to show off.

Build a temporary soundtrack as you go. Silence makes every artifact louder.

Finish sound, grade and titles

Sound is the cheapest realism available. Lay ambience under every scene, put effects on specific actions, and let music carry transitions. Then grade for consistency rather than style: compare each shot against its neighbours and match contrast, black level and colour temperature before adding a look. Titles are the one place where generated lettering still disappoints, so set type in the editor instead of asking for it in a prompt.

Disclose, deliver, and learn from the first one

Assume you will need to describe your methods. The Runway AI Festival's film rules, for example, ask entrants for a short written description of the AI techniques used, alongside a three-to-fifteen-minute runtime and a linear narrative; platforms such as YouTube also maintain disclosure guidance for realistic generative content. Put the tools in the credits, keep the disclosure short and plain, and do not make the audience guess.

Export in a standard codec at a sensible bitrate, keep a master and a compressed copy, and file the project so it can be rebuilt. Then do the only review that matters: watch it with someone who does not care about tools and see where they stop watching. That timestamp is the beginning of the next film. If you want a longer runway, the 90-day learning plan breaks the same skills into weekly practice.