Start with a song.
Add a reference, direct the scenes, compare takes, then assemble the film.
Source material & song structure
01 / Source material
Original generation script
Already have the lyrics and directions you used to make the song? Keep them here as the intended words and performance. The recording may differ; section and lyric timing need review.
Match the sections to the recording
Listen and mark each section's start in seconds, including any lead-in. Start the first section at 0; the last ends with the song. Brackets and parentheses are shown as performance cues. Check that the section breakdown matches your script. These are section markers, not word alignment.
The draft splits your marked sections into shots of up to four seconds, carrying their original lyrics and directions as context. Review each shot's visual direction before saving the scene plan.
Machine transcription (optional)
Use this as a comparison with the recording. It can mishear singing or repeat phrases; it never replaces your original script.
No transcript yet.
Word timings and raw result
Advanced scene planning
02 / Direct the scenes
Direction proposals
An agent can read this production and suggest a shot plan with its reasoning. Review the shots, leave feedback, then accept a proposal to make it the saved plan. Accepting does not start a render.
Edit the scene plan directly
Timeline positions use 24 frames per second. Scene planning can span any length within the song. Individual renders: Wan up to 5 seconds; H3 up to 15 seconds. Include local vocal timing and character directions in each prompt.
Use the live timeline above to split scenes, roll cuts, slide shots and select takes. Whole-plan replacement is disabled once video takes exist.
Take inspection & master playback
03 / Scene direction & cut
Accept a proposed plan or save a scene draft to start directing.
This preview may pause between clips as they load. Final assembly produces a continuous film.
Scene takes
H3 creates a new performance using the song audio. Your previous takes and selection stay saved. Wan shots are limited to 5 seconds. Queued requests need their matching GPU mode. Mode switching is manual; saving a request does not switch the hardware.
Clean film
Select one take per scene. Assembly keeps the original song and trims every scene to its exact place in the timeline. Gaps don't block export: a draft fills them with the scene's board, or black.
Agent access
This workstation and agents use the same saved state. Start with the agent collaboration guide or the API documentation. Use a stable Idempotency-Key for every render, transcription, and assembly.