/quickdesign

When a multi-segment UGC video needs voice continuity across segments (so all speakers sound like one person), use Seedance R2V's native audio_urls reference parameter — not a TTS+lipsync chain.

The right way

1. Segment 1: standard Seedance R2V with native audio. Script in prompt as quoted speech, --generate-audio on (default). This establishes "the voice".

2. Audio extraction:

   quickdesign video extract-audio seg1.mp4 -o seg1-audio.mp3
   # OR raw ffmpeg fallback:
   # ffmpeg -y -i seg1.mp4 -vn -acodec libmp3lame -q:a 2 seg1-audio.mp3

   Auto-uploads via CLI --reference-audio if local path.

3. Segments 2..N: pass the extracted audio as voice reference. Each segment's prompt has its own quoted speech; Seedance generates native audio for that segment but matches the voice character of the reference:

   quickdesign video generate \
     --provider seedance \
     --reference-image <same-or-different-image> \
     --reference-audio ~/path/to/seg1-audio.mp3 \
     -p "<scene + 'The person says: \"...\"' + no-music + no-subs + 9:16>" \
     --duration 12 --resolution 1080p --aspect-ratio 9:16 \
     --wait -o segN.mp4

4. Concat with quickdesign video concat (or ffmpeg -c copy) — same codec/resolution/audio params across segments because every segment is a fresh Seedance R2V output.

Why NOT TTS + lipsync

A typical wrong-pipeline run:

• Seg 1: native Seedance R2V (correct)
• Then orchestrated: extract audio → voice-clone → TTS for segs 2/3 → 2 silent Seedance R2V → 2 lipsync → concat

Problems:

• Result lengths drift (35s instead of target 45s) because lipsync's sync_mode=cut_off truncates the silent video to TTS audio length.
• Cost: ~2x the right way (clone + TTS + lipsync = ~330 wasted credits per video).
• Wall-time: ~12-15 min vs ~6-8 min for the audio_urls approach.
• Multi-step orchestration risk: each step (clone → TTS → silent gen → lipsync) is a separate API call; one transient failure stalls the whole pipeline. audio_urls is one parameter on one API call.

How to apply

• Whenever the user says "voice continuity" / "same voice" / "use this audio as reference" → think --reference-audio, not TTS.
• Plan template for multi-segment UGC:

  "Voice-locked multi-segment: Seg 1 native audio → extract → pass to Segs 2..N as --reference-audio (Seedance R2V native voice matching). Cost ≈ N× R2V (no extra orchestration)."

• Reserve TTS + voice-clone + lipsync for cases where the user explicitly wants a SCRIPTED voice (e.g. dub a custom voice over Seedance video), not for character voice continuity.
• Constraints: --reference-audio repeatable, max 3 audio refs per call, mp3/wav, 2-15s each, ≤15MB.

What audio_urls does and does not do

• Does: lock the voice character (timbre, pacing, accent feel) across segments.
• Does not: dictate WHAT is said in each segment. Each segment's prompt determines the words via quoted speech.
• Does not: replace the need for --generate-audio to be on. The reference is a character anchor; the segment still generates its own dialogue audio.