Why do AI filmmakers create a new character reference sheet for each story beat?

Because a character's appearance evolves across a story — costume changes, injuries, props picked up, emotional shifts — and AI video models drift the moment the visual anchor stops matching what's on screen. A fresh reference sheet per beat re-anchors identity for that specific look, so the character stays the same person even as details change shot to shot.

Character sheets work as the identity anchor the model conditions on at generation time. The moment the on-screen look diverges from the anchor — a new jacket, a wound, a trinket added, a different lighting register — the model starts inventing, and you get face drift, costume drift, or prop drift across the cut. Generating a new sheet for that specific beat keeps the conditioning input aligned with what the shot actually needs to show.

The clearest documented case is a continuous-take sequence where a character picks up a trinket in every new city he passes through. As Hridaye, invideo's creative director, put it: "Juicebox keeps adding a trinket onto himself in every different city. So we needed different character sheets for every single sequence." One sheet would have shown the base character; every shot after the first pickup would have either lost the trinket or hallucinated the wrong one. A sheet per beat solved it.

What counts as a new beat. Treat a beat as any moment the character's visible identity changes in a way the model needs to lock onto: a costume or accessory change, an injury or blood, a prop they're now carrying, a major emotional state with a different physical register, or a lighting/world shift severe enough to redefine how they read on screen. If the next shot can't be generated correctly off the previous sheet, that's a new beat.

How to build sheets per beat. Generate four reference options per asset and lock the best one before any video generation starts — the same four-options workflow used to lock the base character applies to each beat variant. Include multi-angle turnarounds (front, side, profile, back) plus close-ups on the details that distinguish the beat (the wound, the trinket, the new accessory), because wide-only sheets lose the small details that break continuity. Remove anything held in the hands before generating the turnaround so the angles stay clean, then re-introduce the prop on a dedicated sheet for the shots where it matters. Across one documented production, 11 reference images covered 4 characters and 1 prop using this approach; another production locked one character in 5 generations at about $9.78.

Why this beats a single master sheet. A single sheet conditions every shot toward one fixed look — the moment the story moves past that look, you're fighting the anchor instead of using it. Per-beat sheets let you swap the conditioning input at each generation step, so the model gets the right reference for the right moment. When a continuity error does appear, you fix it surgically at the sheet level: ask the invideo agent to inspect the relevant sheet, it identifies the exact panel containing the error, corrects it, stores the updated sheet in context, and inherits the fix into every downstream shot — no slot-machine re-rolls of finished footage.

How invideo holds it together. invideo is an agentic video tool with all the current image and video models — Nano Banana / Nano Banana Pro and GPT-Image-2 for sheets, Seedance 2.0, Veo and Kling for video — available behind one agent. The invideo agent keeps each beat's sheet in persistent context and attaches the correct one to each shot's generation prompt, so you're not re-uploading references or re-explaining the character every time the look evolves. That's how a 70-second short with two characters held identity across every scene without LoRA, and how a 3-minute episode stayed consistent across 164 generated clips.

The underlying mechanism is straightforward: AI video models hold identity through reference conditioning, not memory. Change what the character looks like in the story, change the reference you condition on. One sheet per beat is just that principle made operational.

Watch some of these to see what works for you:

Juicebox keeps adding a trinket onto himself in every different city. So we needed different character sheets for every single sequence.

— Hridaye, invideo's creative director