scenario-seedance-drafts

Scenario Seedance Drafts

Overview

A Seedance draft is a 480p preview of a take that can later be finalized to 1080p without regenerating it: the completion is the same take at full resolution. Seedance has no seed, so a draft is the only way to approve a take cheaply and then ship that exact take. Re-running the payload at 1080p performs it again differently, and a generic upscaler returns a softer copy, not the 1080p render.

The economics decide when to draft. A draft prices like a plain 480p render, and a draft plus its completion costs at least a direct 1080p render: at authoring time 2,040 against 2,024 CU for a 15 second edit (366 plus 1,674), and 544 against 456 for a 4 second reference shot (84 plus 460). Drafting pays only through the takes that get discarded: five 15 second takes and one finalist came to about 3,500 CU, against about 10,100 for five 1080p renders. A single take costs more drafted than rendered.

Seedance modes, conditioning and sound: scenario-seedance. Connection and the core loop: scenario. Music over the finals: scenario-video-assembly. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

Quick reference

Step	Call	What to know
Discover	search, target="models", query="seedance", public=true	model_schema_get the newest member: draft is a boolean there (2.5 at authoring time), and resolution is ignored while it is on
Draft	model_run with "draft": true, wait=false	One call per take: no seed and no batch-count field, so repeat the identical payload. Set duration to the shot (Auto priced the 30 s cap); edit mode takes -1
Wait	jobs_wait with every job id	15 s edit-mode drafts took 7 to 8 minutes; re-call with the returned pending_job_ids
Review	asset_download, then local contact sheets and close-ups	See Reviewing takes
Finisher	search, target="models", query="seedance draft complete", public=true	Its one input is draftVideo, a scalar file
Complete	model_run with draftVideo set to the draft job's asset id	dry_run first; took 1 to 3 minutes. asset_get reports an H.264 MP4, but asset_download returns the original MOV (HEVC in one run): compose by asset id
Verify	asset_get	1080 wide, the draft's frame rate and duration

A draft stays completable for 7 days from its creation. Record each take's asset id straight off its jobs_wait row. Review copies made from a draft (a music version, a trim) are fine, but each is a new asset: only that recorded id resolves to the draft task the finisher completes.

Reviewing takes

Takes of one payload diverge exactly where shots fail, so review every one, never just the first. Download each draft (inspecting a download is not editing it) and check, in this order:

1. Sweep: a contact sheet over the whole clip (ffmpeg -vf "fps=2,scale=150:-1,tile=10x3" fills it for a 15 s clip; raise fps on shorter ones), one per take, stacked per variant. It catches identity swaps, wrong wardrobe, broken backgrounds, extra or missing props.
2. Close-ups where generators fail: hands in fast moves (gloved fingertips turned glowing red in two of three takes of one batch), long hair through turns, the face at four moments, and every prop the brief fixes, counted against its reference (one take stacked a second jack-o'-lantern on a one-lantern plate).
3. Timing: when a source clip exists, frames at four matched timestamps side by side.

Sheets at this size cannot rule out a fault, only find one: crop before judging fingers or faces. Pixel counting is a weak substitute: a red-pixel test for the glove fault also fired on warm candle light in an empty plate, so it ranked nothing. Sampled frames also miss motion faults: a ponytail problem visible at speed passed every sheet. So rank per variant, give each defect a timestamp, say what each take was and was not checked for, and have the user watch the finalists at speed before paying for completion. Unattended, complete the top-ranked take that meets the brief's acceptance criteria and report the checks it did not get.

Worked example: five takes, one 1080p final with licensed music

1. Discover and read the schema as above. dry_run the exact payload twice: with "draft": true, and at the delivery resolution without it. Quote both with the arithmetic (N drafts plus one completion, against N finals).
2. The music is laid afterwards, so generateAudio: false. Launch five identical model_run calls with "draft": true, the shot's duration (in edit mode -1, which follows the source clip) and wait=false. A 429 parallel-custom-jobs creates no job: keep the launched ids, and launch the refused takes as earlier ones finish, within actionLimit per the scenario skill.
3. jobs_wait with every launched job id; record each row's asset id as that take's draft id.
4. Review as above. Present the ranked takes with their links and the defects found, then wait for the pick (unattended: the brief's acceptance criteria pick).
5. Find the finisher, model_schema_get it, then model_run with {"draftVideo": "<draft id>"} and dry_run=true. Run it and jobs_wait.
6. asset_get the completion: 1080 wide, frame rate and duration equal to the draft's. Sweep it once more: the completion matched its draft on every sampled frame at authoring time, and that match is the proof it finalized rather than regenerated.
7. Lay the track on the completion with model_scenario-compose-video (a fixed id: Scenario's single timeline compositor): a video layer and an audio layer, both at startTime: 0, durationMode: "custom" at the clip's length rounded down to the field's 0.1 s step (a 4.04 s clip loses its last frame at 4.0), canvasMode: "custom" with numeric canvasWidth and canvasHeight set to the completion's size, and fps at its frame rate (the field defaults to 30). Check the sync per scenario-video-assembly, then asset_download.

Common mistakes

• Rendering the variations at the delivery resolution: every discarded take costs its full 1080p price.
• Re-running the chosen payload at 1080p, or upscaling the 480p draft: the first is a new take, the second a soft copy of the right one.
• Passing a music version, trim or re-encode as draftVideo: only the draft job's own asset is the draft.
• Letting the 7-day window lapse: an expired draft can only be regenerated, as a new take.
• Laying music on the drafts and completing those: the completion carries only what the draft generated. Review copies with music are fine; lay the delivery audio once, on the completion.
• Ranking from one look at the first take, or from contact sheets alone.