How Do You Make an AI Short Film From Script to Final Cut?

A two-minute AI short film is roughly 20 to 26 shots, and every one of them is generated, checked and cut separately. The pipeline is script as shots, a locked character sheet, storyboard stills, then animation with first and last frame control, then sound and an edit. Seedance 2.5 runs up to 30 seconds per shot and Kling 3.0 holds the planned endpoint, so continuity is a production decision rather than luck.

Updated October 2026.

Write the script as shots, not scenes

A scene is a paragraph. A shot is a unit the model can generate. Convert every scene into shots before you touch a tool.

• One action per shot. “She enters the room” is a shot. “She enters, argues, and leaves” is three.
• A duration you can pay for. Most shots land at 4 to 8 seconds. Seedance 2.5 can run to 30 when a take genuinely needs it.
• A frame size and a camera move. Wide, medium or close, plus one move. That is the director’s decision, and it is cheap to make here.
• A sound note. Dialogue, ambience or a music cue, so the audio pass has a plan.

Number the shots and keep the list in one document. A 2-minute film at an average of 5 seconds is about 24 shots, which is a useful number to hold in your head when the script starts growing.

Lock the characters and locations with a character sheet

Before generating a single clip, build a reference sheet for every recurring character and location. For each character, generate three views on a plain background: front, three-quarter and profile, in the exact wardrobe the film uses. For each location, generate one wide establishing plate and one detail plate.

Store those images and use the same set for every shot they appear in. That is the only reliable way to stop a face changing between shot four and shot nineteen. The character consistency guide covers the reference discipline in detail, and reference-based models such as HappyHorse 1.1 are built for the same job across separate generations.

Consent applies here too. A character based on a real person needs that person’s permission, and a synthetic cast should not be presented as real people.

Storyboard stills before you spend on video

Generate one still for every shot at the frame size the film will use. This is the cheapest place to discover that a sequence does not cut together, and it gives you the first frame for the animation step.

Work through the stills in order and ask three questions: does the eye follow the cut, does the light match the neighbouring shot, and is the character recognisable at this size? Fixing a composition here costs one image; fixing it after animation costs the whole clip. The storyboard tab is the place to lay the sequence out, and our AI storyboard generator guide covers the still-first workflow.

Keep a contact sheet of the whole film on one screen. Problems that are invisible shot by shot show up immediately when the frames sit side by side.

Animate shot by shot with first and last frame for continuity

Animate each storyboard still as the first frame. When a shot has to land on a specific composition, generate the closing still as well and use Kling 3.0 first and last frame control, so the model builds the motion between two approved frames instead of improvising an ending.

Pick the model by the job. Seedance 2.5 for a take that needs 10 to 30 seconds, Kling 3.0 for a planned move with a fixed endpoint, HappyHorse 1.1 for a 4 to 15 second shot at 720p or 1080p where a lower per-second cost helps. Keep the prompt to motion, camera and sound: the still already carries the subject, wardrobe, framing and light.

Render at a draft resolution first and watch every shot in sequence before moving on. A shot that looks fine alone can break the rhythm when it sits between two others. The image-to-video tool keeps the same model picker across the whole film, so switching between models does not mean switching accounts.

Shot need	Model	Why
10 to 30 second take	Seedance 2.5	Long single generation
Planned ending	Kling 3.0	First and last frame control
4 to 15 second shot	HappyHorse 1.1	720p or 1080p at a lower per-second cost

Add voice, music and sound effects

Sound is what makes a sequence feel like a film rather than a set of clips. Build it in three layers.

1. Voice. Generate narration or dialogue in text to speech. Write for the voice: short sentences, clear pauses, no tongue-twisters.
2. Music. One bed for the whole film beats a different track per shot. Keep it under dialogue and let it drop at the emotional turn.
3. Effects and tone. Footsteps, doors, wind and room tone glue the cuts together. A little continuous ambience under every shot hides the joins between generations.

Check the mix on a phone speaker. If a line disappears under the music there, it will disappear for most of the audience. Voice cloning and synthetic dialogue need consent and disclosure where the rules require it.

Assemble in the Timeline Editor

Bring the shots into the timeline editor in story order, then trim the heads and tails. Generated clips often start with a beat of stillness and end with drift; cutting both tightens the film immediately.

• Cut on motion. A move in the outgoing shot hides the join better than a static frame.
• Match the light. If two neighbouring shots disagree, a short colour adjustment is cheaper than a re-render.
• Hold the wide shots. A close-up can be short; an establishing shot needs a moment to be read.
• Add captions if the film will run on a muted feed, using the same word-timed captions workflow as social clips.

Watch the assembled film end to end without stopping. Notes taken during a continuous watch are more useful than notes taken shot by shot.

Budget and drafts: how many takes a 2-minute film really needs

The planning formula is simple: shots multiplied by takes multiplied by the credit cost of one clip. A 24-shot film at three takes is 72 renders, which is why drafts matter. Render the first two takes at 480p, pick the best, and re-render only that take at delivery resolution.

Three takes per shot is a realistic working number for a shot with people or motion; a static insert often needs one. Build the budget from the prices shown in the app rather than an estimate, because the cost depends on model, duration and resolution. A 5-second 720p Seedance 2.5 clip is 126 credits, for example, while a 5-second 480p draft is 56.

Failed generations are refunded, and the price is shown before every render, so the budget is controllable shot by shot. For the wider cost picture behind a trailer, our AI movie guide covers a 60-second version and what each tier costs. If you are buying credits, one-time packs never expire and no subscription is required.

Frequently asked questions

How many shots does a 2-minute AI short film need?

About 20 to 26 at an average of 5 seconds a shot. More shots mean a faster cut and more chances to hide a weak generation.

How do I keep the same character across every shot?

Build a three-view character sheet in fixed wardrobe, reuse that reference set for every shot, and use a reference-based model when a shot does not start from a storyboard still.

Which model should animate the shots?

Seedance 2.5 for long takes, Kling 3.0 for a planned ending with first and last frame, HappyHorse 1.1 for shorter 720p or 1080p shots. All take the storyboard still as the first frame.

How much does a 2-minute film cost?

Shots multiplied by takes multiplied by the per-clip credit cost shown in the app. A 24-shot film at three takes is 72 renders, so draft at 480p and finish only the keepers.

Do I need to disclose that the film is AI-generated?

Where a platform or jurisdiction requires it, yes. Keep the cast synthetic and honest, and get consent for any real person whose likeness or voice is used.

Open the storyboard tab to lay out the shot list, then animate in image to video and cut in the timeline editor. New accounts start with 10 free credits and no card, so you can build the first few frames before committing to a full film; one-time credits never expire, there is no subscription needed, and pack options are on the pricing page.