ugc-video

The single paid generation. Turns the three references + script into a finished ad.

Inputs

• STORYBOARD_JOB_ID, CHARACTER_JOB_ID, PRODUCT_MEDIA_ID
• SEEDANCE_PROMPT (the block from script.md)
• DEST, SLUG, optional MODEL (default seedance_2_0), optional DURATION (default 15), optional RESOLUTION (default 720p — see Resolution note)

Steps

1. Confirm reference roles. Call models_explore(action="get", model_id="seedance_2_0"). Seedance accepts medias roles image, start_image, end_image, video, audio. Pass the three references as: [{ "value": "<STORYBOARD_JOB_ID>", "role": "image" }, { "value": "<CHARACTER_JOB_ID>", "role": "image" }, { "value": "<PRODUCT_MEDIA_ID>", "role": "image" }]
2. COST GATE (mandatory). Call generate_video with get_cost: true and the full params (model, prompt, duration, aspect_ratio, medias, generate_audio). Echo the returned credit cost to the user:

   "This Seedance generation will cost ~<N> credits. Generate? (yes/no)" Do NOT proceed without an explicit "yes". If the user declines, stop and report.

3. Generate (only after confirmation). Call generate_video:
   • model: <MODEL> (default seedance_2_0)
   • prompt: <SEEDANCE_PROMPT>
   • duration: <DURATION> (≤15)
   • aspect_ratio: "9:16"
   • resolution: <RESOLUTION> (default 720p)
   • generate_audio: true
   • medias: the three references above
   • Poll job_status with raw_data:true (keeps the prompt echo small). Real render time is ~3-8 min and is dominated by AUDIO synthesis, so 720p is NOT faster than 1080p — choose resolution on cost/quality, never speed. Capture job_id + URL.
4. Download: curl -fsSL <url> -o "$DEST/$SLUG.mp4".

Output

DEST/<SLUG>.mp4 + video job_id.

Model router (alternates — only on request)

• kling3_0 — when you need manual camera-path control or voice mapped from a reference video.
• cinematic_studio_video_v2 — when the concept needs heavy stylistic / premium-commercial themes rather than phone-selfie realism.

When switching MODEL away from seedance_2_0, re-confirm audio and media-role parameters via models_explore — generate_audio:true and the image reference roles are Seedance-specific and other models (e.g. kling3_0) handle audio and roles differently.

Constraints

• NEVER call generate_video (without get_cost) before echoing cost + getting a yes.
• 9:16, native audio on, ≤15s.

Resolution (validated 2026-06-22)

• Default to 720p. For this phone-selfie UGC format, 1080p (~135 credits) looks no different from 720p (~67 credits) and renders no faster (audio-bound). Ship 720p unless the caller explicitly needs 1080p (large-screen / paid placement).
• Audio is resolution-independent, so there is NO "720p draft then 1080p final" two-step — render 720p once and keep it.
• After download, the human must verify the VO (pronunciation, especially brand names). If wrong, fix the phonetic spelling in script.md and re-roll at 720p (cheap).