How do you use a character reference sheet to keep AI video characters consistent across shots?

Generate a multi-angle character sheet before any video: four turnaround angles plus face and mid-angle close-ups, locked from several options, stored in your generation agent's context, and attached to every shot prompt. One documented 70-second short held 2 characters visually consistent across every scene this way — no LoRA fine-tuning required.

Build the sheet before you generate a single video clip — locking character references up front is the step that prevents consistency problems for the rest of the film. invideo is an agentic video creation tool with all the current image and video models available, so the whole workflow below runs in one place.

1. Generate the turnaround sheet. Create a sheet with four angles (front, side, back, 3/4) plus a face close-up and a mid-angle close-up, at the highest resolution your model offers — documented productions generated 4-angle sheets at 4K. Include close-up panels for small details like scars and accessories; wide shots alone won't hold those across models. Remove any objects from the character's hands first, since held props drift across turnaround angles. For model choice: Recraft produces photoreal portraits with pores, lines, and stubble; Nano Banana Pro outperformed Nano Banana 2 for character sheet fidelity in documented tests; GPT-Image-2 is the other current option — all run inside invideo, so the invideo agent can route each pass to the right model.

2. Generate options, then lock one. Produce several variations per character and select the best before any video generation — one production generated 4 options per asset and locked the winners; another averaged 5 generations (~$9.78) to lock each character, and covered 4 characters plus 1 prop with just 11 total images. Lock the winning sheet and treat it as the single source of truth for that character from this point on.

3. Store the locked sheet in context and attach it to every prompt. Upload the final sheet to the invideo agent with an explicit instruction to save it to context for all further generations, then attach the character reference to every single shot prompt — in a documented 3-minute animated episode, every prompt after the lock carried the reference, across 164 generated clips. As one production walkthrough explains: "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." At the video stage, Seedance 2.0 reference-to-video accepts character references (and location references) directly, which is why it carries identity across clips better than frame-based extension; the invideo agent attaches your locked sheet to each generation automatically.

4. Make a new sheet for each appearance change. When a character's look evolves — new costume, added accessory — create a distinct sheet per beat rather than stretching one sheet across the change. One production's character picked up a new trinket in every city, so the team built a separate character sheet for every sequence.

5. Fix continuity errors at the sheet, not the shot. When a detail goes wrong in a generated shot, don't re-roll the shot — ask the invideo agent to inspect the character sheet for the mistake. In one documented case it identified the exact panel containing a stray earbud, corrected that panel, stored the updated sheet in context, and regenerated only what was needed, so every subsequent shot inherited the fix.

This workflow replaces LoRA fine-tuning for most productions: a 70-second short kept 2 characters consistent across every scene using sheets and agent context alone, finished in 2 days for $750, and a 2-person animated episode held its cast consistent across a 3-minute cut for ~$950 total. Expect some cherry-picking regardless — documented productions averaged 3 generations per usable shot — but the sheet keeps every retry anchored to the same face.

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