What's the best pre-production workflow for keeping AI-generated characters consistent across scenes?

Lock characters BEFORE you generate a single second of video. The pre-production sequence is: load the full script as context, answer four foundational questions (character, antagonist, prop, deliverable), generate 4 options per character sheet at multiple angles, select and lock one, then store those locked sheets in your agent's context so every downstream shot inherits them.

invideo is an agentic video creation platform where you brief a creative producer agent once and it routes every shot to the right model (Recraft, Nano Banana Pro, Seedance 2.0, Veo, Kling) while holding your character locks across the entire film. Here is the pre-production sequence to run before any video generation:

1. Load the full script first. Upload the complete screenplay to a creative producer agent so it holds character arcs, beats, and prop logic before you ask for a single image. This is what prevents the agent from inventing details mid-production that contradict an earlier scene.

2. Answer the four pre-production questions. Before generating assets, force the agent to ask and resolve the four things that, in Hridaye's words, "will change every frame": what does the character look like, what is the antagonist/entity reference, what is the prop specification, and what is the deliverable format. Leaving any of these vague is the most common source of drift later.

3. Generate frames before video — and generate FOUR options per asset. Run portraits first (Recraft handles photoreal skin — pores, lines, stubble), then 360-degree character sheets in Nano Banana Pro with four angles plus a face close-up and a mid-angle close-up. Generate four variations of each sheet, pick one, and lock it. In one documented 70-second short, four options per character + environment, locked before any video, was the step that solved consistency for the entire film at $750 / 3,000 credits total.

4. Build a per-beat sheet when the character evolves. If your character changes costume, picks up a trinket, or transforms across scenes, generate a SEPARATE character sheet for each beat — not one master sheet. Also remove any objects from the character's hands before generating turnarounds, or the prop will drift across angles.

5. Run parallel model casting to pick the strongest base. Tell a casting sub-agent to run the same character prompt on two image models simultaneously (e.g. Recraft vs. Nano Banana Pro), compare, and pick the aesthetic that holds. The invideo agent routes both — you don't switch tools.

6. Store the locked sheets in agent context, then prompt like a director. Once sheets are approved, the agent saves them as the canonical reference and attaches them to every downstream generation (Seedance 2.0 reference-to-video carries character context across clips natively; Kling and Veo take the same sheets as image refs). From here you direct in plain language — "hold on him until he lunges" — and the character stays locked because the reference is already in context, not re-pasted per prompt.

7. When a continuity error appears, fix the sheet — not the shot. If an earlier-generated character has a stray AirPod or wrong scar, ask the agent to inspect the character sheet, identify the panel with the error, correct it in place, and re-store. Every subsequent shot inherits the fix. As Hridaye put it: "Surgical edits. Not slot-machine re-rolls."

One discipline rule across all of it: change only one variable at a time (lighting, lens, or wardrobe — never two together), and reject drift on the first frame instead of letting it propagate.

Which model for what, briefly: Recraft for photoreal character portraits, Nano Banana Pro for multi-angle character sheets, Seedance 2.0 reference-to-video for shots that need character context carried across the clip, Kling and Veo for stylistic range — all available inside invideo, all routed by the agent off the same locked sheets.

Watch some of these to see what works for you:

Before I build assets, four things will change every frame: The Girl: What does she look like? What era? The Entity: Closer to Bathsheba? The Toy: Doll, ball, something else? The Deliverable: The frames first, then video? These four answers unlock everything.

— Hridaye, invideo's creative director