Separate identity references from motion ownership. One driving performance does not establish the choreography of every visible person. If only the lead needs replacement, preserve the crowd from the base footage where the workflow permits. If several dancers must change, map each role to its own supported motion input and evaluate those passes independently before assembling the full scene.
Choose the actual replacement scope
Write down whether the job is lead only, lead plus one partner or the whole group. For each role, record the source performance interval and whether the original pixels should remain. A crowd identity image says how someone might look; it is not a time-coded recording of every crowd member’s movement.
Some performance systems operate on one visible character. Runway’s Act-Two guide describes its driving-performance input. Verify the current capabilities of the actual model you use; do not generalize that contract into native full-group motion capture.
Audit one unmistakably different beat
- Find a moment where the lead and background performer visibly differ.
- Record each role’s pose and movement at that moment.
- Generate or process only the supported role assignment.
- Check the same interval in the output before adding music or transitions.
- Reject a pass that assigns the lead’s action to protected performers.
A staggered entry or delayed clap makes a useful diagnostic. A scene where everyone intentionally moves alike cannot show whether the system understands independent role timing. Select the test for what it can disprove, rather than because it produces a pleasing thumbnail.
Protect the comparison timeline
Keep source and output starts aligned. If one pass is retimed, record the mapping before comparing the group. Otherwise, an apparent delay may come from an edit rather than model behavior. Review transitions into and out of the diagnostic beat, since a role can look correct at one frame and drift immediately afterward.
Can numbering people in the prompt help?
It can clarify the brief, but numbering is not proof that the model supports per-person control. Confirm the actual input fields and inspect the output’s role-specific movements.
What if replacing the whole crowd is too unstable?
Prefer a narrower approved scope, a simpler shot or a compositing workflow with explicit role passes. Describe any changed choreography honestly. A usable adaptation and an exact group reconstruction are different deliverables.
Sources and scope
This is an original Imagild Editorial workflow guide, assisted by AI and reviewed against the linked sources on October 12, 2026. External tool procedures do not imply that Imagild provides those tools or guarantees their results.