/quickdesign

When to use

• User explicitly says "Sora 2", "use Sora", "Sora kalitesi" — audio fidelity is the explicit driver
• Single-shot brief with one of the three durations (4 / 8 / 12s), no multi-segment cuts needed
• Voice continuity across segments is NOT required (Sora 2 doesn't accept audio_urls)

Don't use as a default substitute for Seedance R2V on multi-segment UGC — voice character will drift across cuts because Sora 2 generates a fresh voice per call. R2V's --reference-audio is the only way to lock voice character across multiple segments today.

Hard facts (live)

quickdesign cost sora2-i2v -d 8
quickdesign video models | jq '.data[] | select(.slug=="sora2-i2v")'

• Duration grid: discrete 4, 8, 12 seconds only. No 5s, 6s, 9s, etc. Plan accordingly.
• Cost is duration_lookup, not per-second: 4s=40cr, 8s=65cr, 12s=95cr. Linear-ish but cheaper per second than Seedance 2.0 R2V at 1080p.
• Aspect ratios + resolutions: see registry.
• Native audio: yes, generally cleaner mix than Seedance (less aggressive music bed).
• Reference grammar: NONE. No @Image1/@Image2 labels. Single --image only.

Compact prompt skeleton

<scene description with subject + setting + lighting + mood>.
The person says: "<verbatim quoted speech>".
No music score. No subtitles or on-screen text.
<aspect ratio> format.

Note: there's no @Image1 to lean on, so subject / setting description in the prompt does carry weight here (unlike R2V where verbose descriptions dilute the reference). Keep it descriptive but tight.

Gotchas / failure modes

1. Less aggressive music bed than Seedance — but it can still layer one. Keep the No music score. line.
2. Subtitle hallucination — same as Seedance, add No subtitles or on-screen text.
3. No reference labels means reference image is hint-only, not anchor — small details (shirt color, room props) sometimes reinterpret. Less reliable for product-fidelity work; for products use Seedance R2V.
4. Voice character is fresh per call — if you generate a 12s ad with Sora 2 and want a follow-up segment, you can't lock the voice. The follow-up must be either (a) standalone with no continuity expectation, (b) generated as a single longer Sora call, or (c) switched to Seedance R2V.

Cross-references

• Multi-segment voice continuity needs Seedance R2V → ./seedance-2.0-r2v.md
• Music + subtitle suppression rules → ../references/no-music-no-subtitles.md
• Confirmation gates before paid generation → ../references/confirmation-rules.md