scenario-vidu

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Scenario Vidu Video

Scenario Vidu 视频生成

Overview

概述

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.

Vidu是生数科技在Scenario平台上的视频生成模型系列，每个模式、每个等级都对应独立的模型，而非将所有模式整合到单个模型中：模型名称同时包含模式（T2V、I2V、Reference2V）和等级（Q3、Q2、Q1、2.0）信息。撰写本文时该系列已有15个模型上线，因此请通过

search

查询该模型系列，按名称选型，并以

model_schema_get

返回的结果为准，因为每个等级的功能上限都存在差异。

连接方式与核心流程：参见本仓库中的

scenario

技能；与模型无关的视频处理工作：参见

scenario-video

技能。如果你的可用技能中缺少这里提到的同类技能，请让用户安装（

npx skills add scenario-labs/skills --skill <name>

）；在无人值守模式下，可基于工具schema继续操作并标注缺失项。

Quick reference

快速参考

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

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.

根据你拥有的素材选择模型系列（输入参数名称以实时schema为准）：

你拥有的素材	模型系列	关键输入参数
仅文本	T2V	`prompt` 、 `aspectRatio` 、 `duration` 、 `resolution`
单张静态图片（需生成动效）	I2V	`generationType: "image_to_video"` 、包含单个id的 `images` 数组
首尾帧	I2V	`generationType: "start_end_to_video"` 、包含两个id的 `images` 数组
身份参考素材	Reference2V	`images` ，最多7张

在首尾帧模式下，数组顺序决定了帧的位置：第一个id对应视频开头，第二个对应视频结尾，且

generationType

必须与图片数量匹配。所有I2V系列模型都不支持

aspectRatio

参数：视频尺寸由源静态图片决定。Reference2V系列保留身份特征（角色、道具、配色），但不保证开头帧与参考图一致；如果需要精确的开头帧，请切换到I2V系列模型。有一个Reference2V模型（撰写本文时为Q2 Pro）还支持

videos

参数：最多2个参考视频（单个8秒或两个5秒，每个100MB，宽高比在1:4到4:1之间），需要传入

images

或

videos

参数。

所有模型都支持

seed

参数。

audio

（默认值为false）是背景音乐开关，而非音效设计功能，且在大多数Q2系列模型中，当视频时长为9或10秒时，该参数会静默失效。

Tier picks the caps

等级决定功能上限

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.

撰写本文时的参数限制：Q3系列支持1到16秒时长；Q2系列支持1到10秒，其中首尾帧模式上限为8秒；2.0 I2V系列仅支持4秒或8秒时长，且8秒仅支持720p分辨率；Q1系列完全不支持

duration

和

resolution

参数，2.0 Reference2V系列不支持

duration

参数。

movementAmplitude

（auto、small、medium、large）仅在Q1和2.0系列中可用，通用/动漫

style

选择器仅在Q1 I2V系列中可用；在Q2和Q3系列中，运动强度和风格需要通过prompt描述。Turbo和Fast变体在与Pro版本相同的时长和分辨率上限下，延迟更低。

duration

和

resolution

都会影响成本，部分模型的成本差异可达数倍，因此批量运行前请先对具体任务执行

dry_run

。

Two prompt shapes

两种Prompt格式

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.

T2V和I2V系列需要120到180词的现在时单段描述：包含场景和光线、一到两个动作、一个用简单动词描述的镜头运动（pan、dolly、zoom in），用肢体动作表现情绪（比如“她的肩膀耷拉下来”），不要有镜头切换；且I2V的prompt中不能出现静态图片里没有的内容。Reference2V则正好相反：需要一句约20到40词的短句，按位置说明每个元素来自哪张参考图，比如“image1中的角色穿着image2中的盔甲，在暴风雪中行走”。不同模型的prompt长度上限不同，从2000到5000字符不等。

Worked example: bridging two keyframes

实操示例：衔接两个关键帧

```
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).
```
model_schema_get
```
with that id: duration bounds, resolutions, defaults.
```
upload_asset
```
the start and end stills (see the
```
scenario
```
skill) for two asset ids.

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.

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
```
.
```
asset_display
```
the output and check both anchor frames landed.

使用
```
target="models"
```
、
```
query="vidu"
```
、
```
public=true
```
执行
```
search
```
。优先选择你所需模式下的最新等级模型，例如
```
model_vidu-i2v-q3-pro
```
（撰写本文时为有效结果：每次会话都需要重新查询）。
用该模型id调用
```
model_schema_get
```
：获取时长范围、分辨率、默认值等信息。
上传首尾静态图片（参见
```
scenario
```
技能的
```
upload_asset
```
），获取两个资产id。

调用

model_run

，传入

model_id

、

dry_run=true

，以及

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"}

；实际运行时请将prompt扩展为上述的完整段落格式，且任何时长或分辨率的变更后都需要重新估算成本。

重新调用
```
model_run
```
，设置
```
wait=false
```
，然后用返回的job id调用
```
jobs_wait
```
，超时则用
```
pending_job_ids
```
重新调用，绝对不要重复调用
```
model_run
```
。
用
```
asset_display
```
展示输出结果，检查两个锚点帧是否正确呈现。

Common mistakes

常见错误

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.

两张图片搭配
```
generationType: "image_to_video"
```
，或者单张图片搭配首尾帧模式：类型必须与图片数量匹配，且顺序决定了开头和结尾。
期望Reference2V生成的视频以参考图开头：参考图仅用于固定身份特征，而非第一帧；需要精确开头帧请使用I2V系列模型。
跨等级套用功能上限：16秒仅Q3支持；2.0 I2V仅支持4秒或8秒，且8秒强制为720p。
在prompt中加入对话或音效：
```
audio
```
仅添加背景音乐，且在大多数Q2系列模型中，9或10秒时长时会失效。
给Reference2V写150词的段落，或者给T2V写短句：两种模式需要的prompt格式完全相反。
给任意Reference2V模型传
```
videos
```
参数：撰写本文时仅有一个模型支持参考视频；请先查看schema。