Scenario Seedance Video

Overview

Seedance, ByteDance's video family on Scenario, folds every mode below into one model that infers which from your inputs, so the conditioning you pass decides more than prompt wording. Discover it with search and treat model_schema_get as the contract: members ship side by side, agreeing on the shape and disagreeing on every number.

Connection and the core loop: see the scenario skill in this repo; model-agnostic video work: the scenario-video skill; choreography held across cuts from a storyboard: the scenario-seedance-storyboard skill. 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

Mode follows from the inputs (names from the live schema):

Mode	Inputs	Behavior
Text	prompt	aspectRatio honored (21:9 through 9:16)
First frame	image (+ prompt)	opens on that frame; aspectRatio ignored, size follows resolution and may land only near the ratio
First + last	image + lastFrameImage	lastFrameImage is only valid alongside image
Reference	referenceImages	carries identity and world, not the opening state
Edit	referenceVideos + edit prompt	requires duration: -1; output follows the source
Extend	referenceVideos + extend prompt	inferred from wording: "extend @video1", what holds (character, wardrobe, light), the action that continues, no cut; geometry follows the source; duration is the footage appended and the output holds only that tail (a 10 s source with duration 15 returned 15 s of new footage at authoring time; Auto priced as the source length), so concatenate source and tail and read the delivered length off asset_get

image and the reference arrays are mutually exclusive. referenceVideos (motion conditioning) and referenceAudio (timing and energy, or a speaker's voice) combine with referenceImages; on the 2.0 line reference audio also requires one image or video reference, where 2.5 takes it alone. Reference parameters are arrays even for one asset. Prompt tags bind by array order: @image1 is referenceImages[0], likewise @video1 and @audio1, and the prompt says what each tag governs and what it leaves alone ("@image1 defines the bottle and label, nothing of the set"), for video and audio tags as much as for images. No seed, mask, or camera parameter exists: camera moves live in the prompt, one dominant move and one action per shot. Several actions go as timecoded beats ("0-2s ... 2-4s ..."), one per beat, the last ending inside duration, the shot named as one continuous take.

A reference track the prompt does not claim is easily ignored, a generated voice line most of all, so bind it to a role in words: "@audio1 is her own voice; she speaks it to camera, lips in sync" for speech, "the camera move and each action land on the beats of @audio1" for music.

duration defaults to Auto, which on 2.5 at authoring time priced a text-to-video job as the full 30 second cap (the same 480p prompt quoted about four times its 8 second price). Set the seconds the shot needs on every run; with referenceVideos, Auto follows the longest reference clip instead.

Caps are per member, so read them off model_schema_get. At authoring time 2.5 took 30 reference images, 10 videos, 10 audio, 4 to 30 seconds, up to 1080p; the 2.0, Fast, and Mini hits took 9, 3, 3, and 4 to 15, with 4k on 2.0 alone and no lastFrameImage on Mini. Price moves further than the caps do, six times across the family for one 4 second 480p job, so dry_run the same job on two members before a batch.

The conditioning rule

In reference mode, frame one anchors to the base state of the reference world, and no prompt wording overrides it. If the shot must open in a specific state, render a still of it and pass it as image; use referenceImages when only identity, world, and palette matter. Deciding this per shot before generating is worth more than any prompt tuning.

Sound from the shot, music from elsewhere

generateAudio (default true) is one switch over the whole track: it scores the shot as well as sounding it. A score written inside a shot survives no cut, since each shot invents its own key and tempo and the next one restarts them.

So run the lanes apart. Keep generateAudio: true for what the picture makes, footsteps, impacts, room tone, dialogue, each landing on frame, and name only those sounds ("diegetic sound only: heels on tile, room tone"). Never name what to leave out: "no music" is what the audio filter reads, and a track blocked with OutputAudioSensitiveContentDetected usually traces to such an exclusion, so re-run on the same model with the sound named positively before any provider switch: the filter read the wording, not the provider (see the scenario-moderation skill). Spoken lines go in quotes, sized to the clip's seconds, with the speaker named and the delivery in two or three words. Score the sequence once with a music model (see the scenario-audio skill) and lay that track over the assembled cut (see the scenario-video-assembly skill, or scenario-seedance-music-video when the song comes first): re-scoring then costs one audio run, not every shot again.

The prompt is the only lever, so listen to what comes back: when music leaks in, re-run that shot with generateAudio: false and add sound from a sound-effects or video-to-audio model.

When the finished video is blocked

A failed row whose error names OutputVideoSensitiveContentDetected (often with PolicyViolation, sometimes worded as possible copyright) is the provider screening the finished render. The filter belongs to the family, so a Seedance sibling refuses the same payload and the identical retry renders it again. Handle it the way the scenario-moderation skill teaches: switch to another provider first, prompt unchanged apart from moving the @image tags to the alternative's own reference field per its schema. For recommend, text-to-video is txt2video, first-frame and reference jobs are img2video, edit and extend are video2video. Two observations at authoring time decide what comes after. A character the screen takes for existing intellectual property stays blocked whatever the wording: a studio's own well-known character failed while a lesser-known one from the same studio, prompt unchanged, passed. Weapon and harm vocabulary ("axe", "blade", "eliminated") tripped it even in a cartoon, which is where that skill's one honest rewrite belongs. When the block persists past the switch and that rewrite, on the team's own characters for example, tell the user they can write to support@scenario.com with the failed job id. Unattended, launch the switch only within the spend the task authorized; otherwise stop and report the block with the failed job id.

Worked example: a product shot from references

1. search with target="models", query="seedance", public=true. Prefer the newest non-deprecated hit, e.g. model_bytedance-seedance-2-5 (a live hit at authoring time: re-discover each session).
2. model_schema_get with that id: fields, caps, and defaults before anything else.
3. upload_asset the product stills (see the scenario skill) to get asset ids.
4. model_run with that model_id, dry_run=true, and the exact parameters={"prompt": "@image1 defines the bottle and label. Slow dolly-in as condensation beads. Diegetic sound only: droplets on glass, soft room tone. No text, no captions.", "referenceImages": ["asset_a", "asset_b"], "duration": 8, "resolution": "720p", "generateAudio": true} for the cost estimate; re-estimate after any change to duration, resolution, or references.
5. Repeat model_run with wait=false, then jobs_wait with the returned job id, re-called with pending_job_ids on timeout, never a second model_run.
6. asset_display the output and watch it with sound before the music goes on.

Common mistakes

• Prompting an opening state in reference mode: it will not appear; pass it as image.
• A bare string where the schema says array: one reference still goes as ["asset_x"].
• Combining image with referenceImages or referenceVideos: mutually exclusive inputs.
• Leaving duration on Auto: on 2.5 a text-to-video job priced as the 30 second cap, not the shot's length.
• Passing a voice track beside a character reference without saying whose voice it is: the model may animate the face and ignore the track.
• Carrying one member's caps or price to another: 30 references, 30 seconds, and 1080p are each true of one and false of the next.
• Letting shots score themselves and then cutting them together: the music restarts at every cut.
• Writing "no music" into an audio-enabled prompt: the exclusion trips the audio filter; describe the sound positively, or turn generateAudio off.
• Rendering captions, prices, logos, or UI: reserve clean space and composite text in post.
• Retrying an output moderation block on Seedance or a Seedance sibling: the filter is the family's, so switch provider first (the scenario-moderation skill).