How do you turn an AI portrait into a character reference sheet for consistent video generation?

Turn the portrait into a reference sheet in three steps: lock the portrait at photorealistic quality, expand it into a multi-angle sheet (front, 3/4, side, full-body, plus close-ups), then upload that sheet as the persistent character reference for every video generation. The invideo agent holds the sheet in context so every shot inherits the same face.

Start with the portrait itself, because everything downstream inherits its fidelity. A low-fidelity anchor — smooth plastic skin, vague features — drifts faster across shots than a high-fidelity one, regardless of how disciplined the rest of your workflow is. Generate the portrait in Recraft for photorealistic faces (it renders pores, lines, stubble — the imperfections that make a face actually read as a face), or in GPT-Image-2 or Nano Banana for stylized characters. Generate four options and pick the strongest before going further. invideo is an agentic video tool with every current image, video, and upscale model available inside one workspace, so the portrait, the sheet, and the video all live in the same context.

Next, expand the locked portrait into a multi-angle character sheet. The minimum viable pack is four angles — front, 3/4, side profile, and full-body head-to-toe — generated in Nano Banana (Pro outperforms 2 for character fidelity) at 4K, using the portrait as the reference input. Add a back view and a close-up of the face and mid-angle for a six-panel set if your shots include rear coverage or tight inserts. Remove any objects from the character's hands before generating the turnaround — props in hand cause inconsistency across angles. For one documented production, 11 images covered the full reference set for 4 characters and 1 prop; another locked each character in about 5 generations at roughly $9.78 per character.

Lock the descriptors in writing alongside the sheet. Write the character's identity as a fixed glossary — exact phrasing for hair colour, eye colour, skin tone, build, wardrobe, signature props — and reuse those exact words in every scene prompt. Add a negative-prompt line for the drift you want to prevent (wrong age, extra accessories, wrong hair length, live-action when you want stylized, or vice versa). The identity block and the scene block stay separate: identity stays verbatim, scene/action changes per shot.

Then upload the sheet into the invideo agent as the persistent character reference and let it route to the right video model per shot. Seedance 2.0's reference-to-video accepts the sheet plus location references and carries the character across clips; Kling 3.0 handles native multi-shot sequences; Veo and Runway each have their own strengths the agent picks between. You don't pick a platform per model — the agent attaches your sheet to whichever model the shot needs. Hridaye, invideo's creative director, frames the discipline this way: "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."

For sequences where the character's look evolves — a new trinket each city, a costume change, an injury added mid-film — generate a separate sheet per beat rather than trying to patch one master sheet. And when a continuity error slips through (an AirPod in one panel, a wrong button), ask the agent to inspect the sheet and identify the panel with the mistake; it fixes the source, stores the corrected sheet, and every downstream shot inherits the fix without re-rolling the film. Documented productions hit this benchmark without any LoRA fine-tuning: a 70-second short kept two characters consistent across every scene using sheets plus persistent agent context for ~$750; a 3-minute animated episode held character identity across 41 final clips for ~$315 per finished minute.

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