Why does changing a character's costume or adding a prop require a new AI reference sheet?

AI video models have no memory between generations — every clip is built fresh from whatever reference you attach right now. Add a hat, a necklace, or a torn sleeve, and if your sheet doesn't show that state, the model invents one. A new sheet per appearance state is what keeps the character recognizable.

Treat each costume or prop change as a new identity the model has to learn. The reference sheet is the only thing anchoring the character in any given shot — front, side, profile, back, plus face and mid-angle close-ups — and a video model reading it will reproduce only what's visible on the page. If the sleeve is clean on the sheet but torn in the script, the model either erases the tear or hallucinates a different one each clip. Hridaye, invideo's creative director, frames the failure mode directly: "the AI always needs to see what the character is exactly, right? Or else it'll kind of hallucinate and imagine something that's under the cap. So, we don't want to do that. We always want the character to be seen as we see it on the character sheet."

This is why a per-beat sheet is the working unit. In a 5-day production where a vampire character collected a new trinket in every city he passed through, the team built a separate character sheet for every sequence — "Juicebox keeps adding a trinket onto himself in every different city. So we needed different character sheets for every single sequence." The same rule applies to costume changes, prop swaps, injuries, dirt, wet hair, anything visibly different from the last beat. One sheet per visual state, locked before the video generation for that beat begins.

Build each sheet to the same spec: multi-angle turnaround at 4K in Nano Banana Pro (stronger character fidelity than Nano Banana 2), with close-up panels included — not just wides. Close-ups carry the small details (scars, accessories, fabric texture) that wide angles smear, and without them those details drop out across shots. Generate four options per sheet, pick the best, and lock it — the invideo agent stores the locked sheet in context so every subsequent generation in that beat inherits the same identity. Across one short film, eleven images covered four characters and one prop this way; locking the sheets up front is what kept the cast consistent across 164 generated clips.

Before generating any new sheet, remove props from the character's hands — a held object confuses the multi-angle turnaround and breaks consistency across the four views. Generate the prop on its own sheet, then reintroduce it in the video prompt alongside the character sheet for that beat. Treat the prop as its own asset with its own creative pass: if it reads lifeless, regenerate it against story logic, not just looks.

When a continuity error does slip through — wrong earring in one shot, jacket color drifting — fix the sheet, not the shot. Ask the invideo agent to inspect the sheet for the mistake; it identifies the exact panel, corrects it, stores the updated sheet, and regenerates only what's needed. "It traces the source in your character sheet, fixes it there, leaves the rest of the film intact." That's faster and cheaper than re-rolling clips and hoping the next generation lands.

Platform-native anchoring features in current video models (visual memory, identity adherence across Runway, Veo, Kling, Seedance 2.0) reduce minor drift between adjacent shots, but they do not solve deliberate costume evolution — they're built to keep things the same, not to manage a planned change. The invideo agent routes each beat's generation to the right model and attaches the correct per-beat sheet, so you direct the change rather than fight the model.

Watch some of these to see what works for you:

the AI always needs to see what the character is exactly, right? Or else it'll kind of hallucinate and imagine something that's under the cap. So, we don't want to do that. We always want the character to be seen as we see it on the character sheet.

— Hridaye, invideo's creative director