Guides

A Character Bible for AI Filmmaking: Keeping a Cast Distinct Across Shots

7 October 2026 · RenderBob team

Keeping one generated character consistent is hard. Keeping three distinct in the same frame is harder. A character bible has to solve both before the scenes get built.

Most consistency advice is written for one hero. Films have ensembles. Published production guides report the same failure: when two or more characters share a frame, generation models blend their features. A bible for this work has to hold distinctiveness as well as consistency.

Design the contrast

One vendor guide recommends deliberately different colouring, builds, and silhouettes across a cast. That is ordinary character design, and it matters more here because a model that confuses two similar faces will do it on the next shot as well. A distinct shape and palette give the model more to hold onto.

Build the sheets in neutral light

Capture each character from several angles under consistent, neutral lighting, so the reference carries identity rather than a mood. One image-model guide describes a model that accepts up to fourteen references at once, which is why a proper sheet is worth making. That fourteen-reference figure is the vendor's. Store the sheets with the project so every scene pulls the same files.

Composite the two-shots that matter

The same guidance suggests generating each character on a neutral or green background and compositing them for a critical two-shot. It is slower, and it protects identity. It also matches how a VFX supervisor already thinks about plates and elements.

Test a style change before you build on it

Moving a character from photoreal to illustrated is reported as one of the harder consistency problems. If the film changes style between sequences, test the transition on the reference sheet before scenes depend on it.

Record consent next to the pictures

If a character is based on a real person, or a voice is cloned from a real performer, the bible should hold the consent and the licence terms beside the visuals. Netflix's partner guidance treats generated voice, face, and performance as a separate approval track that requires consent. For animatable digital doubles, pipelines such as Unreal Engine's MetaHuman and Character Creator 4 are commonly cited alongside LoRA identity training.

Use the sheets in casting conversations

Reference art generated from script descriptions gives directors and casting directors a shared visual target to send to agents. Pre-production guides list that use explicitly. The film still depends on knowing who these people are. The tool helps once that is written down, and once the sheets, the consent notes, and the licence terms live in the project store the production controls.