How do you maintain character consistency in AI video for commercial brand productions?

Lock the character before you generate a single second of video. Build a multi-angle character sheet (front, side, profile, back, plus a face close-up), generate four options per asset and pick one, then attach that locked sheet — alongside a brand-style block — to every clip prompt. A dedicated character consistency sub-agent inside the invideo agent guards drift across the campaign.

Start with a frames-first pipeline: approve the static character look before any motion. Generate a brand-accurate portrait with Recraft (for skin-level realism — pores, lines, stubble), then expand it into a 4K character sheet in Nano Banana with four angles plus face and mid-angle closeups. Generate four options per character, per costume, and per hero prop, then lock one. Locking sheets and environment refs before video generation is the single step that prevents consistency problems across the rest of the spot.

Write a character bible the invideo agent holds in context for the entire campaign — invideo is an agentic video tool where every current model and upscaler lives in one place, so the same context routes through every shot. Put these in it: character name, face description, wardrobe with brand colors and logo placement, hair, a unique anchor identifier (scar, accessory, brand-specific marker), prop spec, voice parameters, and motion register. Repeat the block verbatim on every prompt. For a brand with evolving looks across a campaign (the character picks up a new accessory each spot, for example), build a separate character sheet per beat rather than one master sheet — one documented production needed a fresh sheet for every city its character traveled through because a trinket kept getting added.

Stand up a typed-agent crew on the invideo agent: a creative producer agent holding the full script, brand guidelines, and shot breakdown; a casting agent that runs the same character prompt on two image models in parallel (Nano Banana and GPT-Image-2) so you pick the stronger aesthetic before locking; a DOP agent per scene; and a character consistency sub-agent whose only job is to audit every returned clip against the locked sheet and flag drift before approval. Name the consistency agent yourself — "Brand Continuity" or similar — and feed it the sheet plus the negative list (what the character must never become: wrong jaw, wrong jacket, missing logo).

For generation, route each shot to the right model through the invideo agent rather than picking one platform. Seedance 2.0 reference-to-video is the strongest option for brand work because it accepts a character reference and a location reference simultaneously and carries that context across clips — extend cannot do this. Kling handles native multi-shot sequences from a single setup. Veo gives you synced dialogue audio in the same generation, which solves voice-lock for talking-head brand spots. Runway is available for shots where its motion model wins. The invideo agent holds all of them; you direct, it routes.

Use always-ask mode so every prompt comes back for shot-by-shot approval before credits spend. Attach the locked character sheet, the brand style block, and any prior approved clip from the same scene to every prompt — context attachment, not re-prompting, is what holds the character. Expect roughly three generations per usable shot and budget overgeneration as a line item, not waste; in one documented production only 25% of clips made the cut (41 of 164) and on average five seconds of each 15-second clip were used.

If a continuity error appears — wrong earring, jacket flip, logo missing — do not re-roll the shot. Ask the consistency agent to trace the source in the character sheet, fix that panel, store the updated sheet in context, and the next generation inherits the fix while the rest of the spot stays intact. One creator caught a stray AirPod in a character grid this way; the agent identified the exact panel without being told where to look. Surgical edits, not slot-machine re-rolls.

For voice, lock parameters in the bible (timbre, pace, accent, brand-tone descriptors) the same way you lock face. Use Veo for shots where lip-sync and dialogue must match in one pass, and keep a single voice clone on file that every spot in the campaign references — voice drift kills brand recall as fast as face drift.

At campaign scale, work act by act or spot by spot in 25% increments rather than across the whole campaign at once — the agent holds context tighter on bounded chunks. Before assembly, send the rough cut back to the invideo agent with "what's working, what's not" — the maker-checker pass catches register and continuity errors a human editor misses. Across documented productions running this workflow, finished spots ranged 2–7 minutes at $315–$750 per finished minute, with a 2-minute brand promo coming in at ~$1,500 against a $100,000–$500,000 traditional equivalent.

These are the load-bearing pieces — the exact mix depends on your character, your brand system, and how many spots the campaign needs to ship.

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.

— invideo's creative team