How to Make a Long AI Video: A Multi-Scene Guide

There is no model right now — ours or anyone else's — that takes a prompt and hands back a coherent three-minute film. Every frontier video model in this generation produces clips measured in seconds. A "long AI video" is therefore not a longer generation. It is a sequence of short generations that were planned to fit together.

That reframing is the whole craft. Below is the working method we use, and the mistakes that cost the most time.

Why long AI videos are hard

Three problems compound with length.

Subjects drift. The face in scene four is not quite the face in scene one. Small differences in jawline, hair and skin tone are invisible in isolation and glaring in sequence.

Lighting wanders. Each generation makes its own lighting decisions. Two clips that look good separately can read as two different productions when cut together.

Attention is unforgiving. On Reels, Shorts and TikTok, a weak second scene kills every scene after it. Longer content raises the bar for each individual beat rather than lowering it.

None of these gets solved by asking for a longer clip. They are solved before you generate anything.

Plan your scenes

Write the beats down before you open the app. Six to eight lines, one per shot, is the right size for a 30–60 second piece:

Hook — your strongest image, first. No setup.
Setup — who this is and where they are.
Turn — something changes.
Consequence — what the change does to the subject.
Resolution — the payoff, and make it short.

For each line, write three things: what is in frame, what moves, and how it is lit. "Woman in an olive jacket on a rain-slick street at night, neon reflections, she turns toward camera" is a scene. "She feels conflicted about the city" is not — you cannot generate an abstraction.

If a beat needs a paragraph to describe, it is two scenes. Splitting early is cheaper than regenerating later.

Build a multi-scene story in Rigel

Rigel generates one clip at a time, 5–30 seconds depending on the mode, and saves it to your camera roll. The multi-scene workflow is simply doing that in order, deliberately:

Pick the source photo. A sharp, well-lit image with a clearly visible subject does more for the result than any prompt trick. This is the same first step described on the photo to video page.
Describe the scene. One line for the environment, one for the action, one for the light. Add a script if the beat has dialogue, and Rigel generates the clip with lip-sync.
Generate and check. Watch it once at full attention. If the subject drifts or the motion fights physics, regenerate that scene now rather than at the end — every later scene inherits its look.
Save and move on. Keep the clips numbered in the order of the beat sheet so assembly is mechanical.
Assemble. Any editor you already have — the built-in Photos editor is enough for straight cuts. Rigel produces the scenes; the cut is yours.

Each generation shows its credit cost before you confirm it, which matters here: a six-scene story costs roughly six generations. Plan the beat sheet first and you pay for six, not sixteen.

More on the structural side lives on the long AI video page.

Keep characters consistent

This is where most multi-scene pieces fall apart. Four habits fix the majority of it:

Reuse one source image for every scene featuring that person. Changing the source photo mid-story is the single fastest way to break continuity — a new photo means a new face.
Repeat the description verbatim. Copy and paste it. "Woman, early thirties, dark curly hair, olive green jacket" in scene one must be the same words in scene five. Paraphrasing invites the model to reinterpret.
Hold wardrobe and time of day fixed unless the story explicitly changes them. Change one variable per scene — location or action, not both plus the lighting.
Cut on motion. A cut that lands mid-movement hides the small inconsistencies that a static cut puts under a spotlight.

Consistency has improved markedly across the field over the past year, largely through multi-reference conditioning — the shift we wrote about in Pro Reality and in our Veo 3 write-up. It has not become automatic. The habits above still carry most of the weight.

Export for Reels and Shorts

Rigel outputs vertical and square clips saved straight to your camera roll, which is what every short-form platform wants. Three rules beyond that:

Strongest scene first. Platforms judge the piece on its first seconds; so do people.
Cut it shorter than feels comfortable. Most underperforming multi-scene AI stories are thirty seconds of idea stretched to ninety.
Check it with sound off and sound on. If a lip-synced line carries the story, put the same words on screen as text.

Audio is the cheapest upgrade available: a single music bed across all scenes papers over small tonal differences between generations better than any amount of colour matching.

FAQ

How long can one generation be? 5–30 seconds, depending on the mode. Anything longer is assembled from several generations — true of every model in this class right now, not a Rigel limitation specifically.

Do I need an editor? For straight cuts, no — your phone's built-in editor handles it. Anything with transitions, titles or a music bed is easier in a proper editor, but nothing about Rigel requires one.

How many scenes should a first attempt have? Three. Hook, turn, payoff. Learn how your source photo behaves across generations before you commit credits to eight scenes.

Why did my character change halfway through? Almost always a changed source image or a reworded description. Go back to the beat where it drifted, restore the original photo and wording, and regenerate from there.

What does a longer story cost? One generation per scene, and the credit cost of each is shown before you confirm. Rigel is free to download on iOS 16.6+ and Android with starter credits included; credit packs and a subscription are available in-app.

Start with three scenes and one photo you actually care about. The planning is the skill — the generating is the easy part.