How do you keep a character looking consistent across 20+ scenes in an AI-generated film?

Lock the character once, then attach the lock to every shot. Build a multi-angle character sheet (front, side, 3/4, back, plus a face close-up) at 4K, lock it before any video generation, and feed it — together with a fixed descriptor block — into every one of the 20+ scenes. No LoRA needed; consistency comes from the reference, not fine-tuning.

Start by generating the character's face portrait in Recraft for skin-level detail (pores, stubble, lines), then turn that portrait into a multi-angle character sheet in Nano Banana — four angles plus face and mid-angle close-ups, at 4K. Generate four options per sheet, pick the strongest, and lock it. Do this before a single video clip is generated; locking character sheets and environment references upfront is what prevents consistency problems across the rest of the film.

invideo is an agentic video creation platform with every current generation and image model available inside one agent, so the same locked sheet routes through whichever model each shot needs. Spin up a sub-agent and call it your casting agent: give it the portrait, have it build the sheet, and store both in the agent's context. From that point on, every shot prompt the invideo agent assembles attaches the sheet automatically — across 20, 50, or 169 scenes (one documented project's notebook shows scene numbering up to #169).

Write a fixed descriptor block for the character — hair, skin tone, wardrobe, distinctive features, age, build — and prepend it to every generation prompt alongside the sheet. Pair that with a 9-element prompt assembly order (camera spec, lens, lighting source, palette, composition, atmosphere, mood, film attribution, negative prompt) so the character lock sits inside a consistent stylistic frame on every shot. The negative prompt matters: explicitly prohibit drift ("not photorealistic" if you're animated, "no alternate hairstyles, no costume changes") on every prompt.

If the character evolves across the film — picks up a trinket, changes costume, ages a beat — generate a separate character sheet per beat rather than relying on one. In one documented production, a vampire character collected a different trinket in each city, so the team built distinct character sheets for every sequence and locked each before its scene generated. Same character, beat-specific reference.

For the video model itself, route to whichever holds character context best for the shot: Seedance 2.0 reference-to-video accepts the character sheet plus a location reference and carries identity across clips; Kling and Veo are options for shorter beats. The invideo agent does the routing — you don't pick the model per shot, the agent does, and the locked sheet rides along regardless.

Validate before you compound drift. After each scene generates, ask the invideo agent to inspect the new clip against the locked character sheet and flag any mismatch — wrong eye color, a missing accessory, a hairline shift. When an error appears, do not re-roll the shot. Trace it back: ask the agent to inspect the character sheet for the source mistake, correct that panel, store the corrected sheet, and only regenerate what's affected. One creator had the agent identify the exact panel containing a stray earpod and fix the sheet surgically — the rest of the film stayed intact. Surgical edits beat slot-machine re-rolls every time.

A few practical numbers: locking one character typically takes ~5 generations at roughly $9.78 per character, and across documented productions character consistency held for 70-second to 7-minute films covering 20+ scenes — without LoRA, without fine-tuning. Expect ~3 generations per usable shot and budget overgeneration as a line item, not a failure. Work act-by-act (finish ~25% of the film before starting the next chunk) so the agent's context stays sharp on the locked character through the whole 20+ scene stretch.

The failure mode to watch: drift compounds. If any scene's output silently becomes the next scene's input without validation against the locked sheet, small mismatches snowball. The validation pass after every scene is what makes 20+ scenes survivable.

Watch some of these to see what works for you:

Seventy seconds. Two characters. The same person across every scene. No LoRA needed.

— Hridaye, invideo's creative director