scenario-grok-imagine-video

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Scenario Grok Imagine Video

Overview

概述

Grok Imagine, xAI's video family on Scenario, splits its modes across single-purpose members: text and first-frame generation, reference-to-video, prompt editing, and clip extension each live in their own model, so picking the member is picking the mode. Discover them with

search

and treat

model_schema_get

as the contract: members agree on prompt style and disagree on every cap. The Grok Imagine image models belong to the

scenario-grok-imagine-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.

Grok Imagine是xAI在Scenario平台上的视频模型系列，其不同功能模式由各自的专用成员模型提供：文生视频与首帧生成、参考生视频、提示词编辑、片段延伸分别对应独立模型，因此选择成员模型即选择功能模式。可通过

search

查找这些模型，并将

model_schema_get

视为约定：所有成员模型的提示词风格一致，但参数限制各不相同。Grok Imagine图像模型属于

scenario-grok-imagine-image

skill。

连接方式与核心流程：参见本仓库中的

scenario

skill；通用视频处理工作流：参见

scenario-video

skill。如果你的可用skill中缺少此处提到的同类skill，请让用户安装（命令：

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

）；若处于无人值守模式，可基于工具schema继续操作并标记缺失项。

Quick reference

快速参考

Pick the member by intent (input names from the live schema):

Intent	Member	Inputs
Text-to-video	Video 1.5	`prompt` alone; `aspectRatio` auto lands on 16:9
Animate a still	Video 1.5	`image` + `prompt` ; auto follows the image ratio
Identity, opening free	Video 1.5 R2V	`referenceImages` + `prompt` tagging `@image1`
Restyle or swap in place	Edit Video	`video` + a prompt naming only the changes
Continue a clip	Extend Video	`video` + a continuation prompt; `duration` is new seconds

Caps are per member, so read them off

model_schema_get

. At authoring time the generation members took 1 to 15 seconds, 1 to 4

numOutputs

, and eight

aspectRatio

values including auto; 1080p existed only on Video 1.5's text and first-frame modes, while R2V stopped at 720p and defaulted to 480p. The earlier unversioned Grok Imagine Video takes the same inputs as Video 1.5, also caps at 720p, and runs cheaper per

dry_run

. R2V took up to 7

referenceImages

; Extend generated 2 to 10 new seconds; Edit exposed no duration or resolution and preprocessed its source down to 8.7 seconds at 720p, so trim to the segment first (see

scenario-video

). No seed, negative prompt, or camera parameter exists anywhere in the family.

根据使用意图选择对应成员模型（输入参数名称以实时schema为准）：

用途	成员模型	输入参数
文生视频	Video 1.5	仅需 `prompt` ； `aspectRatio` 默认为16:9
静态图片动效化	Video 1.5	`image` + `prompt` ；自动适配图片比例
保持主体身份、首帧自由生成	Video 1.5 R2V	`referenceImages` + 标注了 `@image1` 的 `prompt`
原地重设风格或替换元素	Edit Video	`video` + 仅描述修改内容的提示词
延伸视频片段	Extend Video	`video` + 续写提示词； `duration` 为新增的秒数

各成员模型的参数限制不同，请通过

model_schema_get

查询。在本文撰写时，生成类模型的耗时为1至15秒，

numOutputs

范围为1至4，支持包括自动在内的8种

aspectRatio

取值；1080p分辨率仅支持Video 1.5的文生视频和首帧生成模式，R2V最高支持720p，默认480p。更早的无版本号Grok Imagine Video输入参数与Video 1.5相同，最高也支持720p，且单次

dry_run

成本更低。R2V最多支持7张

referenceImages

；Extend可生成2至10秒的新内容；Edit不提供时长或分辨率参数，会将源视频预处理为720p、8.7秒以内的片段，因此请先裁剪到目标片段（参见

scenario-video

）。全系列模型均不支持种子值、负向提示词或相机参数设置。

First frame or references

首帧与参考图

image

locks frame one: the video opens on that exact composition and animates out of it.

referenceImages

(its own member, R2V) carries people, products, and clothing across the shot without deciding how it opens. Tags bind by array order:

@image1

referenceImages[0]

(

<IMAGE_1>

also works), and one clean subject per reference keeps control; a busy reference dilutes it. If the opening must match a composition exactly, render that still and pass it as

image

; when only a face, product, or outfit must stay consistent, use R2V and tag each reference where it acts.

image

参数用于锁定首帧：视频会以该图片的精确构图开场，并由此展开动画。

referenceImages

（对应独立成员模型R2V）用于在整个镜头中保持人物、产品、服装的一致性，但不决定视频开场方式。标签按数组顺序绑定：

@image1

对应

referenceImages[0]

（也支持

<IMAGE_1>

写法），每张参考图对应一个清晰的主体能更好地控制效果；参考图内容过于杂乱会削弱一致性效果。如果开场必须完全匹配某个构图，请渲染该静态图并作为

image

传入；如果仅需保持人脸、产品或服装的一致性，请使用R2V并在提示词中对应位置标注参考图标签。

Sound is prompted, not switched

音效由提示词控制，无单独开关

Generation members produce native audio and no parameter controls it: the prompt is the whole mixing desk. Left unaddressed, the track tends toward generic background music, so end every prompt with named sounds ("Audio: rain on glass, distant traffic") or "no music". Dialogue written in quotes with a delivery verb lip-syncs: She says calmly: "We're live." Structure the rest like a director: scene, camera, style, motion, audio, in present tense, one continuous shot, one dominant camera move, physical actions instead of named emotions. 80 to 150 words touching at least three of those layers beat a one-line scene; editing prompts run shorter, naming only what changes. Extension prompts describe the next beat, not a new scene: open with a bridge ("the shot continues"), keep the established light and cast, and restate the audio.

生成类成员模型会自带原生音频，且没有单独的控制参数：提示词就是全部的调音台。如果不做指定，音轨通常会是通用背景音乐，因此请在每个提示词末尾加上具体音效说明（例如："Audio: rain on glass, distant traffic"）或注明"no music"。带引号的对话加上表述动词可实现口型同步：例如 She says calmly: "We're live."。其余内容请按照导演脚本的结构组织：场景、相机、风格、动作、音频，使用现在时，单个连续镜头，一个主导的相机运动，用物理动作而非情绪形容词来表达。包含至少三个上述层面、长度在80到150词的提示词效果远好于单行场景描述；编辑类提示词可以更短，仅说明修改内容即可。延伸类提示词应描述下一个情节而非新场景：以衔接语开头（例如"the shot continues"），保持已有的光线和角色设定，并重申音频要求。

Worked example: dialogue over an animated hero still

实操示例：为静态英雄角色图片生成带对话的动画视频

```
search
```
with
```
target="models"
```
,
```
query="grok imagine video"
```
,
```
public=true
```
. Match the member by name: first-frame animation wants e.g.
```
model_xai-grok-imagine-video-1-5
```
(a live hit at authoring time: re-discover each session).
```
model_schema_get
```
with that id: fields, caps, and defaults before anything else.
```
upload_asset
```
the hero still (see the
```
scenario
```
skill) to get its asset id.

model_run

with that

model_id

dry_run=true

, and the exact

parameters={"prompt": "The frame shows a knight on a cliff at dusk. She lowers her sword, turns to the camera, and says quietly: 'It ends tonight.' Slow dolly-in, wind lifting her cloak. Audio: wind, distant thunder, her line clear. No music.", "image": "asset_abc", "duration": 8, "resolution": "1080p"}

for the cost estimate; re-estimate after any change to duration, resolution, or

numOutputs

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
```
.
```
asset_display
```
the output and watch it with sound: the audio direction is judged by ear.

调用
```
search
```
接口，参数为
```
target="models"
```
、
```
query="grok imagine video"
```
、
```
public=true
```
。按名称匹配对应成员模型：首帧动画场景可使用例如
```
model_xai-grok-imagine-video-1-5
```
（本文撰写时的有效模型ID：每次会话都需重新查找确认）。
传入该ID调用
```
model_schema_get
```
：优先确认字段、参数限制和默认值。
调用
```
upload_asset
```
上传英雄角色静态图（参见
```
scenario
```
skill），获取资源ID。

调用

model_run

，传入

model_id

、

dry_run=true

以及如下

parameters

参数，以获取成本估算：

{"prompt": "The frame shows a knight on a cliff at dusk. She lowers her sword, turns to the camera, and says quietly: 'It ends tonight.' Slow dolly-in, wind lifting her cloak. Audio: wind, distant thunder, her line clear. No music.", "image": "asset_abc", "duration": 8, "resolution": "1080p"}

；若修改了时长、分辨率或

numOutputs

，需重新估算。

再次调用
```
model_run
```
，设置
```
wait=false
```
，然后用返回的任务ID调用
```
jobs_wait
```
；若超时，使用
```
pending_job_ids
```
重新调用，切勿重复发起
```
model_run
```
请求。
调用
```
asset_display
```
展示输出结果并开声音观看：音频效果需通过听觉判断。

Common mistakes

常见错误

Leaving audio to the default: generic music appears; direct the sound or state "no music" in every prompt.
Prompting an exact opening frame in R2V: references guide identity, not frame one; pass the still as
```
image
```
on Video 1.5.
Expecting 1080p everywhere: at authoring time it existed only on Video 1.5 text and first-frame runs, and R2V silently defaults to 480p.
Feeding Edit Video a long clip and expecting full-length output: the source is preprocessed down (8.7 seconds at authoring time); trim to the segment first.
Reading Extend's
```
duration
```
as total length: it counts only new footage.
Stacking camera moves or contradictory directions ("zoom in as the camera pulls back"): one dominant move per shot.

让音频保持默认设置：会生成通用背景音乐；请在每个提示词中指定音效或注明"no music"。
在R2V中提示精确的开场帧：参考图仅用于控制主体身份，不控制首帧；如需固定首帧，请在Video 1.5模型中传入静态图作为
```
image
```
参数。
认为所有模型都支持1080p：本文撰写时，1080p仅支持Video 1.5的文生视频和首帧生成模式，R2V默认会静默使用480p。
向Edit Video传入长片段并期望输出完整长度：源视频会被预处理压缩（本文撰写时为8.7秒）；请先裁剪到目标片段。
将Extend的
```
duration
```
参数理解为总时长：该参数仅表示新增的视频时长。
叠加多个相机运动或使用矛盾的指令（例如"zoom in as the camera pulls back"）：每个镜头仅设置一个主导的相机运动。