what sign up
nixm nixm

← learn

Short film with AI — how to hold a story together across shots

November 3, 2026

A short film is shots that agree with each other. Generated clips are produced independently, so continuity comes from what you hold constant — the same character described the same way, one named light, short takes cut together. nixm composes each still with you before animating, which is where continuity is actually won.

Can you actually make a short film with AI video?

Yes, with one adjustment to expectations: you are not generating a film, you are generating footage. People searching for an AI movie maker or an AI film generator are almost always looking for this exact workflow — a sequence of directed shots, not one long render that arrives finished. Each clip arrives independently, knowing nothing about the last one — so the film exists in the edit, the way it always has. What AI changes is the cost of coverage, not the grammar of assembling it.

Where does continuity actually come from?

From discipline, not a setting. Nothing carries state between clips, so anything you want repeated has to be re-specified every single time — and specified identically. Write the character once, as a fixed block of words, and paste that exact block into every frame you compose. Same features, same wardrobe, same order of description. Vary the wording and you have cast a different actor.

How do you keep a character consistent?

Lock the face as a still before you shoot anything. Compose one frame you would genuinely cast, approve it, and treat it as your reference for the whole shoot — every subsequent frame described against that image rather than against your memory of it. Because nixm builds the still first and animates only the approved frame, the character is settled while it is still cheap to change. Cast once, then shoot.

How do you hold the light and the grade?

Name them explicitly, every shot, in the same words — one source, one direction, one palette. "Sodium streetlight from camera left, cyan fill, deep shadow" repeated across six frames will hold a look far better than six different attempts at "moody night." Continuity of light reads as continuity of place, which is what makes separate clips feel like one scene.

What makes a shot read as real?

Realism is not a resolution setting. It is an absence of tells, and the eye audits physics, not pixels. It runs three checks, in this order. Light: one source, consistent shadows, believable falloff. Name the light in the prompt ("late sun through blinds") and half the realism arrives with it. Physics: weight in motion, momentum that carries, one continuous action rather than several competing ones. Sound: a silent clip is pre-declared fake, while room tone and footsteps sell a shot before the visuals finish loading.

Route around the known hard modes: hands doing precise work, dense crowds, fast dialogue in close-up, and text in the scene. Wide-shot the crowd, cut before the handshake, let the sign blur with depth of field. Real cinematographers dodge their equipment's weaknesses too. Every tell in the still gets animated, so judge realism at the frame, where iteration is cheap. And label it: realistic synthetic media is exactly what platform disclosure toggles exist for, serious models embed provenance metadata regardless, and the label costs one tap.

How long should each shot be, and how do you plan them?

Short. One held image with one continuous motion, eight seconds at most — long takes expose drift, cuts hide it. Plan in beats rather than minutes: an establishing wide, the action in a medium, a close for the turn, then whatever the moment costs the character. Shoot coverage of the beats that matter, so the edit has options instead of obligations. Native ambience renders with each clip and can be mixed or muted per shot; how the sound layer works covers that in detail.

What does a short film cost in clips?

Countable before you start. A two-minute short is roughly fifteen to twenty-five kept clips, plus the takes you discard as stills — which cost no video credit at all. At one non-expiring credit per HD clip the whole thing is arithmetic, and the project can pause for a month without a subscription running through the gap. Cutting to an existing track instead? The music video workflow is the same shot logic, paced to a song.

Direct the first shot free

Your first HD clip with sound is free on a new nixm account. Open the cinematic studio and cast the character before you shoot a frame.

Case study: The Seam — two takes, one room

The Seam is a short we are making on nixm. An astronomer, Mara, finds her own observatory at the edge of the universe; forty years later, as an old woman, she raises a lantern whose light becomes the signal her younger self finds. The key image is a crossing: young Mara and old Mara in the same control room at the same moment, without seeing each other. Only the viewer sees both. That is two separate AI takes, superimposed in the edit, and it only works if the two takes line up almost perfectly. Here is how it went, including what failed.

1. Make the empty room first

Before any character appeared, we made a still of each location with nobody in it: a control room and a catwalk around a dome. That empty still, called a plate, became the start frame for every take, so both versions of Mara walk into exactly the same picture.

An empty observatory control room: a wall of blue monitors, a desk with a takeaway coffee cup, two chairs, and windows onto white domes under stars.
The control-room plate: the empty room every take started from.

One side effect we had not planned for: the output takes the shape of the start frame, not 16:9. Our plates came out at roughly 4:3, so every take rendered at 1676×1236. If you need widescreen, make a widescreen plate.

2. "Locked-off camera" was not enough

The first take of young Mara failed. The prompt asked for a locked-off camera, and the video engine pushed in anyway: over the first three seconds the table and chairs slid out of frame. A take that moves cannot be superimposed on one that does not.

A strip of frames from the first seven seconds of a take, showing the camera pushing in and the table and chairs leaving the frame.
The failed take: the camera creeps in over the first seconds, despite the instruction not to move.

What fixed it was naming what had to stay in frame: the table, the chairs, the cup. The retry held perfectly still, and so did every take after it that used the same kind of line. If you want a static shot, describe the frame you want to keep, not just the camera you want to stop.

3. Check the overlay before you edit

Before cutting anything, we blended the matching frames of each pair at 50% and looked for doubled edges. In the control room, the window, domes, monitors, table, chairs and cup registered exactly; the two figures read as one crossing the other.

Two takes blended at fifty percent: the control room is perfectly sharp, with a translucent older woman raising a lantern on the left and a translucent young woman at the central monitors.
The control-room pair blended at 50%: the room registers exactly, and both women are there at once.

On the catwalk, the start frame matched the plate and the overlay lined up too, but both women end up small and dark at the dome door, so at full frame only the lantern's glow really reads. The fix there is a closer plate, not a better take.

Two catwalk takes blended at fifty percent: the walkway and dome register cleanly, with two faint figures at the dome door and a lantern glow above.
The catwalk pair: aligned, but the crossing is faint because the figures are so small.

4. What did not go to plan

  • The path. In the second control-room take, the older Mara entered late and stopped at the front left instead of the central screen where her younger self stops. The background still aligned, so we kept it, but the two never quite meet at the same spot.
  • Ageing. Asking for Mara at 74 kept her face recognisably hers, but she read younger than asked, closer to 65.
  • Small figures. At a distance, faces cannot be judged at all. Plan the closer shots you will need for the audience to recognise who is who.

5. What it cost

Five takes, including the failed one, at 3 video credits each: 15 credits. Each take rendered in roughly 9½ to 15½ minutes. One retry in five is a reasonable budget for shots that have to match.

The takeaway: superimposing AI takes is very doable if you fix the background before you shoot, using one plate for both takes and naming what must stay in frame, and check the overlay before you edit. Why references and details slip between takes is in why AI video drops your references, and how a start frame works is in image to video.

frequently asked

Can you make a short film entirely with AI video?

You can generate all the footage, but the film is made in the edit — clips are produced independently and assembled in an editor, the same as any shoot.

Is there an AI movie maker that does the whole film?

No — no single render makes a film. Every AI movie maker and film generator produces shots, and the film is built from shots and cuts, which is the workflow this guide covers.

How do you keep a character consistent across AI clips?

Describe them identically every time, in a fixed block of words, and work from one approved still as the reference. Nothing carries between clips automatically.

Why do my AI shots not look like the same scene?

Usually the light changed. Naming one source, one direction and one palette in the same words across every shot is what makes separate clips read as one place.

How long should each AI clip be?

Around eight seconds, holding one image with one continuous motion. Short takes cut together cleanly; long takes expose drift.

How many clips does a short film need?

A two-minute short is roughly fifteen to twenty-five kept clips, plus discarded takes — which cost nothing if you reject them as stills before animating.

Can I try it free?

Yes — nixm includes one free HD video with sound on a new account, no card required.

What makes AI video look realistic?

Coherent single-source light, motion with real weight and momentum, and native audio that matches the scene, in that order. Avoid precise hand work, dense crowds, close-up dialogue and in-scene text, and judge the still frame before rendering.

cast one character. put it anywhere.

try it now · cinematic studio

cast one character. put it anywhere.

cinematic ai video, with sound. nixm proposes every scene before a credit is spent.

start directing → first video free · no subscription

Enter your email and receive a link to sign in. No passwords, just a secure signal to connect.

why connect

  • 20 free credits to start
  • your signals, remembered
  • buy more credits anytime
  • image & video generation features
or