Other speaking routes

The routes for speaking shots that don't fit the default.

A native actor's read driving the face

A fallback only (decided). The method is [single strong source], with no Arabic case shown yet. Use it for a line the AI voice still can't carry after the fix ladder (see Voice and Arabic dialects → The voice route), or on a project shot with real actors. Book it only when you decide to, and write that choice in the plan.

• Film the actor's read on a phone or camera, framed and lit like the planned shot.

• Run Seedance video-to-video with the read as the video reference for performance, timing and lips; identity and set from the anchors; audio off. The prompt keeps identity, face, voice, the whole spoken performance, framing and lens exactly as they are, and tells the model to follow the source audio's lip sync without change.

• The actor's audio ships, or a voice-changed version of it. A Sync-3 touch-up on the same audio is optional.

• Frame it at the delivery framing from the start, and never crop in during the edit: lips drift when a 1080p video-to-video shot is cropped.

• An actor's fee runs about $25–80 (US dollars) per read. Whether video-to-video accepts Draft Mode is unconfirmed.

Wides, small faces, backs and turned faces

• Lay the line over the picture with no lip sync: the speaker's mouth naturally active if visible, or the face turned away.

• If the mouth reads at delivery size, use the default instead.

• Background walla (the murmur of a crowd, with no words picked out) in wides: real recordings or ElevenLabs crowd lines. Arabic walla is never Seedance's (decided): it comes from real CC0 crowd recordings (see Sound design and mix → Real recordings from Openverse (CC0)) or from voices recorded in ElevenLabs.

• Wordless laughs, gasps and breaths come from Seedance as effects. A laugh or a word that is a scripted line is recorded and locked like any line.

Songs and jingles

• The vocal comes first: the song is locked before the shots that sing it (see Music → The music-first route). A native singer for finals; Suno's guide vocal for drafts and timing.

• Sync-3 runs on the isolated vocal stem, cut into a plate-length file per shot.

• The plate sings: open, held vowels written into the beats; the camera moves with the music, under the same limits as speech. The prompt is Template F.

• When the singer's vocal replaces the guide, the shots already synced to the guide are synced again to the singer's stem, one pass each inside the shot's budget. The plates stay.

Two or more speakers

Keep each speaker separate [official guidance]:

• Record the exchange together, in one session or in ElevenLabs' dialogue mode (see Voice and Arabic dialects → ElevenLabs settings, tags and pace), then split it into one stem per speaker.

• Each generation and each Sync pass holds one speaker. Cut between singles, the listener quiet or off screen; overlapping words sit on the off-screen speaker.

• Give each speaker a single camera position and framing, held for every line they say, with the same gentle move each time.

• Avoid two-shots where both people talk: Sync can't be told which face to fit.

• For touch and shared looks, use the two-hander pattern (see Speaking shots and lip sync → Harder speaking shots).

A voiceover from off screen

A voiceover from off screen gets no lip sync. On-screen mouths stay closed, the shot's sound says no voices, and the voiceover is laid in the edit. A treatment may choose no on-camera speech at all, to avoid the AI look.

Sources

• Runway, multi-character dialogues with Act-Two (a recorded human performance drives the face); Flawless TrueSync on Watch the Skies (Decrypt) (the actors re-record and the face follows their delivery); adilinthewild's video-to-video guide for Seedance 2.5 (a real take's lips kept through video-to-video, in English).