/quickdesign

When to use

Default for any spoken-script video — UGC, talking avatar, promo, explainer, multi-scene ad. Don't downgrade to seedance-2.0-i2v because the script is "short enough" — R2V handles 4–15s single-shot just as well as multi-segment, with strictly more capability. See ../SKILL.md cardinal rule #0.

Switch off R2V only on explicit user opt-in ("use Sora 2", "use Kling") or if R2V is no longer in the registry.

Hard facts (live)

quickdesign cost seedance-2.0-r2v -d 12 -r 1080p   # exact compute
quickdesign video models | jq '.data[] | select(.slug=="seedance-2.0-r2v")'

Known behavior:

• Duration grid: integer 4..15 seconds (plus auto). No 5/10/15 rounding — pick the tightest value that fits the segment script.
• Resolutions: 480p, 720p, 1080p. Default to 1080p for all UGC; 720p is meaningfully cheaper but soft on faces.
• Aspect ratios: 9:16, 16:9, 1:1, 4:3, 3:4, 21:9.
• Native audio: on by default. Voice character is generated unless --reference-audio locks it to a prior segment.

Reference grammar

Seedance reads the inputs as positional labels:

Don't re-describe what the references already show. Verbose verbatim re-descriptions dilute the reference anchor → cross-segment drift in face / hair / wardrobe / setting / smaller props.

❌ Wrong:

A young man with messy brown hair, a short beard, wearing a black t-shirt,
in a casual indoor home setting with a ceiling fan, warm interior lighting.
He says: "..."

✅ Right:

@Image1 in the same setting. Hands at his sides. He says: "..."
No music score. No subtitles or on-screen text.

What stays in the prompt vs what becomes a reference

Compact prompt skeleton

@Image1 in the same exact setting throughout.
<one-sentence action/state for this segment>.
He/She/The person says: "<verbatim quoted speech for this segment>".
No music score. No subtitles or on-screen text.
Vertical 9:16 format.

For multi-product UGC, the skeleton expands:

@Image1 in the same setting. The held product matches @Image2 / @Image3 exactly.
<action>. He/She says: "<script>".
No music score. No subtitles or on-screen text.
Vertical 9:16 format.

Gotchas / failure modes

1. Auto-layered music bed under voiceover. Seedance defaults to layering a music track when prompts contain quoted speech. Add No music score. to suppress. Don't enumerate ambient sounds you want to keep — over-prescribing makes audio feel scripted. See ../references/no-music-no-subtitles.md.

2. Burned hallucinated subtitles. When a prompt has quoted speech, Seedance burns captions into pixels. They hallucinate (partial sentences, paraphrased words, wrong timing). Always add No subtitles or on-screen text. Use quickdesign video subtitle post-generation if you actually want captions — the dedicated endpoint runs real ASR. See ../references/auto-subtitle.md.

3. Multi-segment voice drift. Without --reference-audio from Seg 1's extracted audio, every segment generates a fresh voice character. Always lock with audio_urls. See ../references/voice-continuity.md.

4. Camera-motion verbs cause defects. Don't write "slowly zooms in", "pans across", "static hold". Seedance produces natural micro-motion (breathing, gestures, head turns) on its own — explicit verbs override that with mechanical-feeling motion. See ../references/first-frame-not-camera-motion.md.

5. Fine print / engravings drop. When the product has visible text (brand wordmark, hallmarks like "925", model number, care label), Seedance has no obligation to preserve it pixel-faithfully across motion. For ads where label fidelity matters: ensure the source banana edit (or first-frame reference) preserves the text legibly first; mention the specific text by name in the prompt; check the rendered first/last frame.

6. BFF cost inconsistency. The registry says per_second × duration (e.g. 12s × 21cr/s = 252cr at 1080p). The current BFF runtime occasionally falls back to legacy (duration/5) × 200 = 480cr when the registry cache is cold. Compare your job's actual token_cost via quickdesign video status seedance <reqId> against the registry compute. This is a known BFF bug, not a CLI bug.

Cross-references

• Multi-product / multi-angle reference usage → ../references/multi-reference-pattern.md
• Voice continuity across segments → ../references/voice-continuity.md
• Music + subtitle suppression (minimal directive) → ../references/no-music-no-subtitles.md
• Why "first frame" matters more than camera-motion verbs → ../references/first-frame-not-camera-motion.md
• Confirmation gates before paid generation → ../references/confirmation-rules.md
• Multi-segment UGC method → ../pipelines/ugc-video.md