Start with the performance you want to keep
Runway Act-Two is a performance-driven character workflow. The useful starting point is a recording with the right words, pauses and expression. A polished portrait alone does not specify how the scene should be acted. Prepare the performance and the character reference as different production inputs, then review the returned take against both.
Our original exercise is a quiet two-person exchange about an unsent letter. A asks, “Did you send the letter?” B replies, “Not yet. I wanted you to read it first.” Both are fictional adults. The line is deliberately simple so the review can concentrate on speaker assignment, a listening face and a consistent envelope. These lines are an unperformed script, not a recording or a tested model result.
Choose the correct character input
Runway’s official Act-Two guide describes a driving-performance video plus a character image or video. Image inputs support gesture control and added environmental motion. With a character video, facial performance is applied while its existing scene and camera motion provide the foundation; gesture control is unavailable. The documented maximum is 30 seconds, with output at 24 fps. Check the current interface and generation quote before committing a take.
| Starting material | Production decision | Review priority |
|---|---|---|
| Accepted character still | Plan the pose and visible gesture | Does the actual body action match the intended turn |
| Accepted character video | Keep its scene timing explicit | Does the face align with the shared background movement |
| Two people in one scene | Assign a separate performance pass to each person | Does only the intended face speak in each pass |
Inspect the original dialogue timing reference
Watch this clip and view its details
The diagram assigns A a planned turn from zero to three seconds, leaves a pause, then assigns B four to seven seconds. It lasts eight seconds at 24 fps with 192 frames and no audio. These intervals are a rehearsal scaffold. They are not measured speech durations or proof that a generation will follow a particular schedule.
Download the original performance plan, record both parts and replace planned times with the actual recording times. Keep the listener present through the other person’s line. A listening shot needs an intentional expression and stillness, not an unexplained second speaking turn. Open the dedicated timing player to inspect the cues at a larger size.
Record each speaker with a shared time origin
For the letter exercise, keep the camera position, lighting and envelope state fixed while recording each role. Start both recordings with a clear common cue outside the intended edit. Record a short lead-in and tail so a cut does not land directly on the first consonant or final expression. Preserve the original files before trimming.
Name the files by role and revision, such as letter-A-take-02 and letter-B-take-03. Note which take is accepted. A filename alone does not prove that its spoken line matches the script, so listen to both recordings from beginning to end. Rehearse the exchange against the other part and adjust the pauses before transferring the performance.
Build a two-character scene from separately reviewed passes
Runway’s multi-character tutorial works around single-character inputs by isolating each person from a shared character-video base, applying the corresponding driving recording separately, then combining those outputs in an editor. Its workflow uses the common base to coordinate scene motion. For longer performances, it also addresses matching the base duration to avoid reversed looping.
For our exercise, save an accepted shared composition before producing either pass. Record where A and B sit, which person holds the envelope and what the room looks like. Preserve the full scene dimensions and time origin even if an individual pass uses a crop. Changing the reference halfway through creates a different alignment problem for the final composite.
Review A’s output before producing a replacement for B. If A already changes the envelope or background, record that defect instead of assuming a second pass will cover it. The blank dialogue-composite ledger keeps the base, both performance sources and the actual returned files together.
Check alignment at the face and the edge
Place the actual outputs over the accepted base in your editor and compare their positions at the beginning, during the overlap and at the ending. A useful composite must preserve the chosen scene relationship across time. Aligning one eye on the first frame is not enough if the head or camera moves later.
Watch the listener’s mouth during the other person’s turn. Inspect the edge around the jaw, hair, shoulder and envelope. Review the audio track separately so a duplicate performance recording or guide playback does not remain underneath the scene. Keep dialogue, room tone and music as separate editable layers while making the acceptance decision.
The AI rotoscoping and matte-review guide includes an original moving overlap study with a visible-subject alpha example. Use it to practice recognizing an incorrect foreground relationship. It does not automatically create mattes for your Act-Two outputs.
Preserve story continuity across the next angle
A shot change should preserve who knows what and what has happened to the letter. If the envelope is closed in the shared view, a close-up should not begin with it already open. Check the accepted last state before briefing another angle. The multi-shot cinema guide covers cut boundaries and screen direction; the short-drama guide supplies the broader story packet.
Use Clout AI Cinema to plan your original scene sequence and work with its available scene-generation controls. Create or select accepted character and room references before extending the story. This guide describes Runway’s documented performance workflow; it does not establish an Act-Two integration or a dialogue-compositing editor inside Clout.
Review the export as the deliverable
Reopen the exported scene and watch it at its delivery size. Confirm the actual words, speaker order, image dimensions, duration and sound. Inspect the cut into and out of the exchange. Save the accepted export alongside the source performances, reference revision and editor project so a later line change has a clear replacement boundary.
Acceptance should name the file and the observed result: A asks the approved question, B listens, B replies, and the envelope remains with its assigned holder. A successful generation status or an attractive still frame does not establish those facts. Complete that review before expanding the exercise into a longer episode.
