Scenario Vidu Video

Overview

Vidu, Shengshu Technology's video family on Scenario, ships one model per mode per tier instead of folding every mode into one model: the model name carries both the mode (T2V, I2V, Reference2V) and the tier (Q3, Q2, Q1, 2.0). Fifteen members were live at authoring time, so search the family, pick by name, and treat model_schema_get as the contract, since caps disagree at every tier boundary.

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

Choose the model line by what you hold (input names from the live schemas):

You hold	Line	Key inputs
Text only	T2V	prompt, aspectRatio, duration, resolution
One still to animate	I2V	generationType: "image_to_video", images with one id
Start and end frames	I2V	generationType: "start_end_to_video", images with two
Identity references	Reference2V	images, up to 7

In start/end mode array order decides: the first id opens the clip, the second closes it, and generationType must match the image count. No I2V member exposes aspectRatio: geometry follows the source stills. Reference2V carries identity (characters, props, palette), not the opening state; to open on an exact frame, switch to an I2V member. One Reference2V member (Q2 Pro at authoring time) also took videos: up to 2 reference videos (one of 8 seconds or two of 5, 100MB each, aspect between 1:4 and 4:1), requiring images or videos.

Every member takes seed. audio (default false) is a background-music toggle, not sound design, and on most Q2 members it silently does nothing at 9 or 10 seconds.

Tier picks the caps

At authoring time: Q3 ran 1 to 16 seconds; Q2 ran 1 to 10, with start/end mode capped at 8; 2.0 I2V took exactly 4 or 8 seconds, and 8 only at 720p; Q1 exposed no duration or resolution at all, and 2.0 Reference2V no duration. movementAmplitude (auto, small, medium, large) exists only on the Q1 and 2.0 lines, and the general or anime style selector only on Q1 I2V members; on Q2 and Q3, motion energy and style go in the prompt. Turbo and Fast variants cut latency at the same duration and resolution caps as their Pro siblings. duration and resolution both move cost, several-fold on one member alone, so dry_run the exact job before a batch.

Two prompt shapes

T2V and I2V want a 120 to 180 word single-shot paragraph in present tense: scene and lighting, one or two motions, one camera move in plain verbs (pan, dolly, zoom in), emotion as physical cues ("her shoulders slump"), no cuts, and in I2V nothing that is not already in the still. Reference2V wants the opposite: one short sentence, roughly 20 to 40 words, naming which element comes from which reference by position, "the character from image1 wearing the armor from image2, walking through a snowstorm". Prompt length caps differ per member, 2000 to 5000 characters.

Worked example: bridging two keyframes

1. search with target="models", query="vidu", public=true. Prefer the newest tier in the mode you need, e.g. model_vidu-i2v-q3-pro (a live hit at authoring time: re-discover each session).
2. model_schema_get with that id: duration bounds, resolutions, defaults.
3. upload_asset the start and end stills (see the scenario skill) for two asset ids.
4. model_run with that model_id, dry_run=true, and parameters={"generationType": "start_end_to_video", "images": ["asset_start", "asset_end"], "prompt": "The knight slowly lowers his sword as dusk settles over the courtyard. The camera dollies in on his face.", "duration": 8, "resolution": "1080p"}; expand the prompt to the full paragraph shape above in real runs, and re-estimate after any change to duration or resolution.
5. Re-run 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 check both anchor frames landed.

Common mistakes

• Two images with generationType: "image_to_video", or one with start/end: the type must match the image count, and order sets start versus end.
• Expecting a Reference2V clip to open on a reference image: references fix identity, not frame one; use an I2V member for exact opening frames.
• Carrying caps across tiers: 16 seconds is Q3 only; 2.0 I2V accepts only 4 or 8, and 8 forces 720p.
• Prompting dialogue or sound effects: audio only adds background music, and it drops out at 9 or 10 seconds on most Q2 members.
• A 150-word paragraph on Reference2V, or a bare sentence on T2V: the two modes want opposite prompt shapes.
• Passing videos to any Reference2V member: at authoring time only one took reference videos; read the schema first.