AI Multi-Character Scene Generator

The hard part is who is in frame, not how many

Generated video has a specific failure with more than one person in a shot: it will put in a face that should not be there. Someone mentioned in the text — looked at, shouted to, thought about — arrives on screen, wearing a face invented on the spot, because nothing told the model the difference between “present” and “referred to”.

So the shot list states it per shot. Each shot carries an explicit list of which characters are in frame, and that list — not the prose, not the previous shot's ending state — is the only thing that decides whose reference art gets attached.

At a glance

Why naming an off-screen character costs you the shot

This is measurable, not theoretical. Shots that named an off-screen antagonist produced the wrong face or an extra person in five out of eight renders. The same shots rewritten to give a direction instead of a name produced zero such failures in twelve renders. So the pipeline rewrites them: in a single-person shot, a character who exists only as an eyeline or a direction of movement is converted to a spatial phrase and removed from the frame list.

The same logic covers the dead: a character referred to only as a body in this segment is written as a plain noun, not as a character reference, so nobody attaches their reference sheet and stands them up.

Two people facing each other

There are exactly two ways to shoot it and the list is required to pick one: a profile two-shot, or an over-the-shoulder. What it may not write is “A facing B, with B behind in the background, out of focus” — for B to be in A’s depth of field, A has to have their back to B, and that camera position does not exist in a normal medium shot. Asked for it anyway, the renderer puts one person face-on to camera and the other standing behind them. The axis is set in the first shot where both are actually in frame together.

Crowds are declared or absent

A room of three hundred guests is a visible state of the scene, so it is written as one — a line in the segment header saying who the background people are, where, and what they are doing. Consistent across every segment in that scene. Without that line the shot may not refer to a crowd at all, because “murmur of guests” in the ambient audio is not a visual instruction, and a renderer told only that will give you guests in one shot and an empty hall in the next.

When a crowd is declared, the frame instruction changes with it: the crowd stays in the depth of field, and the in-frame list still counts only named characters.

What it will not do

• Put eight named characters in one shot reliably. Past four, identity degrades. Ensemble scenes are covered in singles and over-the-shoulders — the way they are shot on real sets.
• Keep a character whose in-frame flag is missing. If the model forgot to list someone who should be visible, they are left out rather than guessed in. Missing a person is recoverable; a stranger's face in your drama is not.
• Choreograph continuous group action. Crowds moving in formation, sustained multi-party fights — those are outside what current models hold together.
• Read a crowd out of sound design. Declared or absent; nothing in between.

What it costs

Nothing extra — these are properties of the shot list, which is billed by script length at 1 credit per hundred characters of the episode's share. Reference sheets are 7 credits each beyond the 15 free ones in your first project, and rendering starts at 10 credits per second of output. See character consistency for how faces stay the same across shots, and the pricing page for rates.

Frequently asked questions

Pricing·Novel to Video·Guides·Samples·Support·Privacy·Terms·Legal·Contact·© 2026 SceneMixer