Yes — upload the reference video and AI can reproduce the choreography's movement vocabulary, rhythm, and energy, though frame-exact replication is less reliable than stylistic transfer. In one documented production, a filmmaker screen-recorded a film's fight choreography, uploaded it to the invideo agent, and the agent dissected and reinterpreted the movement for generation — no choreography expertise required.
Screen-record or upload the reference clip directly and ask the AI to study the choreography — you don't need to describe the movement in words. invideo is an agentic video creation tool with the current video models built in, and the invideo agent accepts video uploads as reference inputs, not just images. One filmmaker producing an AI series put it plainly: "I screen recorded that section, and I put it on Agent One, and I asked it to study the choreography of it. And it surprisingly dissected it ironically without even using a knife." The agent extracts the movement structure and reinterprets it for your characters and setting, which is how most productions actually use choreography references — the AI carries the rhythm, staging, and energy rather than cloning the dancer frame by frame.
For tighter motion fidelity, route the reference into a model that takes video as a direct input. Seedance 2.0 reference-to-video accepts a full video clip alongside character and location references, so the generated dancer follows the reference motion while keeping your character's identity; Kling offers reference-video-driven dance replication, and community tests show it reproducing choreography from a source clip. Both models run inside invideo, so you don't pick a platform per model — the invideo agent routes the shot to whichever handles the motion best.
If no reference footage exists, record one yourself: perform the movement on your phone and upload that clip as the motion reference — documented productions use phone-shot mock footage as a visual anchor when prompting alone can't produce the motion.
For complex or extended choreography, split the work across sub-agents. One documented production ran 5 parallel agents on a single fight sequence — one brainstorming choreography, one updating shared context, one generating images, one converting image to video, and one developing an alternate version of the same sequence — and the solo creator shipped a full 10-minute episode in under 10 days this way. The creator also credited roughly 50% of the drama in one action sequence to the agent's own creative additions, so expect the AI to contribute movement ideas, not just copy them.
Two limitations to plan around. Multi-character physical contact — partner work, lifts, bodies touching — breaks video models faster than almost any other scenario, so lock a separate character sheet per dancer before generating and expect more iterations on contact moments. And Seedance 2.0 generates frame by frame, so dynamic pans and whip moves can lose character identity mid-shot — build the dance as cut coverage rather than one long moving master. Finally, if the choreography you're copying is someone's published work, note that choreography can carry copyright protection and the legality of AI-replicated dance is an active debate — reinterpretation for style is safer ground than one-to-one reproduction.
Watch some of these to see what works for you:
I screen recorded that section, and I put it on Agent One, and I asked it to study the choreography of it. And it surprisingly dissected it ironically without even using a knife.
— a filmmaker documenting an AI series production workflow