scenario-wan

Scenario Wan Video

Overview

Wan, Alibaba's video family on Scenario, ships each job as its own member, and members a generation apart disagree on parameter names, not just numbers: pick the member with search, then treat model_schema_get as the contract before every run. Wan 2.7 Image is an image model, covered by the scenario-image skill.

Connection and the core loop: see the scenario skill in this repo; model-agnostic video work: the scenario-video 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

Members by job (at authoring time):

Job	Member	Required inputs
Text to video	2.7, 2.6, 2.5 T2V	prompt
Animate a still	2.7, 2.6, 2.5 I2V	image + motion prompt
Extend a clip	2.7 I2V	video + prompt (excludes image)
Restyle footage	2.7 Video Edit	video + edit prompt
Character performs a video's motion	2.2 Animate (Move)	imageUrl + videoUrl
Swap the person in footage	2.2 Animate (Replace)	videoUrl + imageUrl
Change aspect ratio	2.2 Reframe	videoUrl
Expand the canvas	2.2 Outpainting	videoUrl + prompt

Parameters do not port across generations. At authoring time: the 2.7 generators take resolution (720p or 1080p) and whole-second duration from 2 to 15, with aspectRatio (16:9 through 3:4) on T2V only. 2.6 and 2.5 T2V instead take one size string such as "1920*1080", with duration from a fixed grid (5, 10, or 15 on 2.6; 5 or 10 on 2.5), while their I2V members take resolution. Video Edit takes a 2 to 10 second source; its duration is optional and truncates. Reframe and Outpainting have no duration: length is numFrames (81 to 241) at framesPerSecond (5 to 30), and output matches the source only when matchInputNumFrames and matchInputFramesPerSecond are true (both default false). Outpainting guides new content with up to 10 refImageUrls.

On 2.7 I2V, image and video are mutually exclusive, and endImage (last frame) is only valid alongside image. Resolution and duration carry cost (1080p ran about triple 720p on 2.7 T2V at authoring time), so dry_run before batches.

Prompt the member's job

T2V wants a full scene in natural language: subject action, one camera move, lighting, style tags. I2V wants motion only: the still already fixes look and composition, so re-describing it wastes the prompt. Video Edit wants instructions, one clear change per sentence with a precise target, plus preservation language ("keep subject identity and camera path"); referenceImage carries a target look. In-frame readable text is unreliable on every member; push artifacts into negativePrompt (Video Edit and Animate lack it).

2.6 alone does multi-shot: with multiShots true (the default, but active only while enablePromptExpansion is true), open with a global style line, then Shot 1 [0-3s] ..., Shot 2 [3-7s] ..., keeping characters and location continuous. enablePromptExpansion defaults true on the generators (helps thin prompts, adds latency) and false on Reframe and Outpainting.

Where the audio comes from

The 2.7 generators always deliver audio: upload an audio file to drive it, or omit it and a synced track is generated from whatever sound the prompt describes, so prompt the ambience you want. On 2.6 and 2.5, audio comes only from an uploaded file (3 to 30 seconds, 15 MB or less at authoring time), which 2.6 lip-syncs. Video Edit's audioSetting decides the track: origin keeps the source audio, auto lets the model regenerate it. Animate carries the source track while mergeAudio stays true (the default).

Worked example: bracket a shot between two stills

1. search with target="models", query="wan image to video", public=true. Prefer the newest non-deprecated hit, e.g. model_wan-2-7-i2v (a live hit at authoring time: re-discover each session).
2. model_schema_get with that id: exclusivity notes, caps, defaults.
3. upload_asset the opening and closing stills (see the scenario skill) for asset ids.
4. model_run with that model_id, dry_run=true, and parameters={"prompt": "She turns from the window and walks toward the door, coat swaying; slow push-in. Room tone and distant traffic.", "image": "asset_open", "endImage": "asset_close", "duration": 5, "resolution": "1080p"} for the cost estimate; re-estimate after changing duration or resolution.
5. Re-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 review motion and audio together.

Common mistakes

• Combining image with video on 2.7 I2V, or passing endImage without image: the schema forbids both.
• Porting a 2.7 payload to 2.6 or 2.5 T2V: they take size ("1280*720" style), not resolution plus aspectRatio, and off-grid durations fail.
• Expecting Reframe or Outpainting to preserve length: defaults render 81 frames at 16 fps; set the two match flags.
• Outpainting's expand flags default to all four sides: set the unwanted directions (expandTop, expandBottom, expandLeft, expandRight) to false.
• Writing shot brackets with enablePromptExpansion false: multiShots is inert without it; no other member reads brackets.
• Re-describing the still or requesting content changes in an I2V prompt: I2V animates; Video Edit changes content.