scenario-kling

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Scenario Kling Video

Scenario Kling 视频

Overview

概述

Kling, Kuaishou's video family on Scenario, ships as specialists, eighteen at authoring time: V3 (Omni plus dedicated T2V and I2V tiers), O1, 2.6, motion control, lipsync, and a talking avatar. Pick the member whose conditioning matches the job, then treat
model_schema_get
as the contract: the same parameter changes shape, default, and legality between members. Kling Video to Audio is audio output, the
scenario-audio
skill's domain.
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.
Kling是快手在Scenario平台上推出的视频模型家族,在本文撰写时共有18个专用模型:V3(含Omni以及专用的T2V、I2V档位)、O1、2.6、动作控制、口型同步和数字人模型。请选择适配任务需求的模型,将
model_schema_get
作为接口约定:同一参数在不同模型中的格式、默认值和合法性可能不同。Kling视频转音频的输出功能属于
scenario-audio
技能的范畴。
连接方式与核心流程:参见本仓库中的
scenario
技能;通用视频处理功能:参见
scenario-video
技能。如果你的可用技能中缺少此处提到的关联技能,请提示用户安装(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 by job; discover ids with
search
(
target="models"
,
query="kling"
,
public=true
):
JobMembers
Every mode in one model, tier via
mode
V3 Omni (
standard
720p,
pro
1080p,
4k
)
Text to video at a fixed tierV3 T2V Standard / Pro / 4K, 2.6 T2V Pro
Animate a still, optional last frameV3 I2V Standard / Pro / 4K, O1 I2V, 2.6 I2V Pro
Consistent characters from photosO1 Reference Images, V3 I2V
elements
Edit or restyle footage by promptO1 Video Editing / Reference Video, Omni (
videoReferenceType: "base"
)
Motion from a video onto a character imageMotion Control (V3 Pro / Std, V2.6)
Talking headAI Avatar (image plus audio), Lipsync (video plus audio or text)
Schema traps (names from live schemas, caps at authoring time):
  • V3 dedicated lines take
    prompt
    or
    multiPrompt
    , never both;
    multiPrompt
    is an array of
    {prompt, duration}
    shots, and
    shotType
    (
    customize
    or
    intelligent
    ) exists only on the V3 T2V members and I2V 4K. Omni keeps
    prompt
    required; its
    multiPrompt
    is a JSON string of up to 6 shots whose durations sum to
    duration
    .
  • duration
    is a string enum on most members (
    "5"
    , not
    5
    ); Omni alone takes a number, 3 to 15. The V3 dedicated lines reach 15 seconds; O1 and 2.6 stop at 10.
  • On O1 and V3 I2V an element is a
    frontalImage
    plus up to 4 angle
    referenceImages
    : both required on O1, each optional on V3 I2V, where a
    video
    can define the element instead.
  • Budgets shrink beside a video: O1 Reference Images takes 7 total (elements plus images), the O1 video members 4, and Omni's
    referenceImages
    drops from 7 to 4 next to
    referenceVideo
    . Tags bind by order (
    @Element1
    ,
    @Image1
    ); Omni prompts use
    <<<image_1>>>
    and
    <<<video_1>>>
    .
  • A reference clip runs 3 to 10 seconds on Omni and the O1 video members, 2 to 10 on Lipsync (trim first, per
    scenario-video
    ). Omni's
    referenceVideo
    defaults to
    videoReferenceType: "feature"
    , lending style and camera to a new clip;
    "base"
    edits the clip and ignores
    duration
    ;
    4k
    mode refuses a reference video either way.
  • cfgScale
    (0 to 1, default 0.5): raise toward 0.8 for storyboard fidelity, drop toward 0.3 to let the model invent.
  • Prompt in director order: scene, subject, action, one camera move, then audio and style; natural sentences beat tag lists, and complex scenes hold together near 5 seconds, not 15.
根据任务选择模型;可通过
search
接口查询模型ID(参数:
target="models"
、
query="kling"
、
public=true
):
任务类型可用模型
单模型支持所有模式,通过
mode
切换档位
V3 Omni(
standard
720p、
pro
1080p、
4k
)
固定档位的文生视频V3 T2V 标准版 / 专业版 / 4K版、2.6 T2V 专业版
静态图片动效化,支持可选尾帧V3 I2V 标准版 / 专业版 / 4K版、O1 I2V、2.6 I2V 专业版
基于照片生成一致性角色O1 参考图模式、V3 I2V
elements
参数
基于提示词编辑或重制视频风格O1 视频编辑 / 参考视频模式、Omni(
videoReferenceType: "base"
)
将视频动作迁移到角色图片上动作控制(V3 专业版 / 标准版、V2.6)
数字人播报AI 数字人(图片+音频)、口型同步(视频+音频/文本)
Schema注意事项(名称来自线上Schema,限制为撰写时的版本):
  • V3专用系列模型只能传入
    prompt
    或
    multiPrompt
    二者之一,不可同时使用;
    multiPrompt
    是由
    {prompt, duration}
    格式的镜头组成的数组,
    shotType
    (可选
    customize
    或
    intelligent
    )仅在V3 T2V系列和I2V 4K版中支持。Omni模型要求必须传入
    prompt
    ;其
    multiPrompt
    为JSON字符串,最多支持6个镜头,所有镜头的时长总和需等于
    duration
    。
  • 大多数模型的
    duration
    为字符串枚举类型(例如
    "5"
    ,而非数字
    5
    );仅Omni模型接受数字类型,范围为3到15。V3专用系列最长支持15秒;O1和2.6版本最长支持10秒。
  • 在O1和V3 I2V模型中,元素(element)由1张
    frontalImage
    (正面图)和最多4张不同角度的
    referenceImages
    (参考图)组成:O1模型要求两者都必须提供,V3 I2V中两者均为可选,也可通过
    video
    来定义元素。
  • 当使用视频输入时,参考资源的配额会减少:O1参考图模式总共支持7个(元素+图片),O1视频类模型支持4个;Omni模型的
    referenceImages
    配额在搭配
    referenceVideo
    使用时会从7个降至4个。标签按顺序绑定(
    @Element1
    、
    @Image1
    );Omni的提示词使用
    <<<image_1>>>
    和
    <<<video_1>>>
    格式。
  • Omni和O1视频类模型的参考视频时长要求为3到10秒,口型同步模型为2到10秒(需先按照
    scenario-video
    技能的说明裁剪)。Omni的
    referenceVideo
    默认
    videoReferenceType: "feature"
    ,用于为新视频赋予风格和镜头运动;
    "base"
    模式用于编辑原视频,此时会忽略
    duration
    参数;
    4k
    模式下无论哪种方式都不支持参考视频。
  • cfgScale
    (范围0到1,默认0.5):调高至0.8左右可提升分镜还原度,调低至0.3左右可让模型更具创造性。
  • 提示词建议按照导演逻辑排序:场景、主体、动作、一个镜头运动,最后是音频和风格;自然语句的效果优于标签列表,复杂场景在5秒左右时长下能保持连贯,不建议使用15秒。

Audio is a per-member switch

音频功能因模型而异

generateAudio
defaults true on the V3 dedicated lines and 2.6 T2V, false on Omni and 2.6 I2V, and does not exist on O1 (editing members carry source audio via
keepAudio
). It is refused beside Omni's reference video and beside
lastFrameImage
on 2.6 I2V; the 4K members bill per second either way. Voice output is Chinese and English, other languages auto-translate to English: put dialogue in quotes in the prompt, lowercase for English speech, uppercase for acronyms. For new speech on existing footage use Lipsync: an audio file, or
text
with a
voiceId
, never both.
generateAudio
参数在V3专用系列和2.6 T2V模型中默认为true,在Omni和2.6 I2V中默认为false,O1模型不支持该参数(编辑类模型通过
keepAudio
保留原音频)。在Omni模型使用参考视频、或2.6 I2V模型使用
lastFrameImage
时,无法开启该功能;4K版本无论是否开启音频都按秒计费。语音输出支持中文和英文,其他语言会自动翻译为英文:请将对话内容放在提示词的引号中,英文语音用小写,首字母缩略词用大写。若需要为现有视频添加新语音,请使用口型同步模型:可传入音频文件,或传入
text
和
voiceId
,二者不可同时使用。

Motion control reads two inputs

动作控制需要两个输入

Motion members require a character
image
and a driving
video
;
characterOrientation
names which input holds the character and, on V2.6, caps the driving clip at authoring time: 10 seconds for
image
, 30 for
video
. Defaults differ (V3 Std says
video
, the others
image
), so set it explicitly.
keepOriginalSound
decides whether the source audio survives.
动作控制类模型需要角色
image
(图片)和驱动
video
(视频)两个输入;
characterOrientation
参数用于指定角色所在的输入源,且在V2.6版本中,该参数会限制驱动视频的时长(撰写时的规则):角色为
image
时驱动视频最长10秒,角色为
video
时最长30秒。不同模型的默认值不同(V3标准版默认为
video
,其他模型默认为
image
),因此请显式设置该参数。
keepOriginalSound
参数用于决定是否保留原音频。

Worked example: two-shot character clip from a still

示例:从静态图片生成双镜头角色视频

  1. search
    with
    target="models"
    ,
    query="kling image to video"
    ,
    public=true
    . Prefer the newest non-deprecated hit, e.g.
    model_kling-v3-i2v-pro
    (a live hit at authoring time: re-discover each session).
  2. model_schema_get
    with that id: multiPrompt shape, duration values, element caps.
  3. upload_asset
    the start frame and the character's frontal photo (see the
    scenario
    skill).
  4. model_run
    with that
    model_id
    ,
    dry_run=true
    , and
    parameters={"startImage": "asset_s", "multiPrompt": [{"prompt": "Wide shot, @Element1 crosses the plaza, slow dolly-in", "duration": "5"}, {"prompt": "Close-up, @Element1 smiles and says: 'we made it'", "duration": "4"}], "generateAudio": true, "elements": [{"frontalImage": "asset_f"}]}
    for the cost estimate; re-estimate after changing durations, tier, or audio.
  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 review both shots with sound.
  1. 调用
    search
    接口,参数为
    target="models"
    、
    query="kling image to video"
    、
    public=true
    。优先选择最新的未弃用结果,例如
    model_kling-v3-i2v-pro
    (撰写时的有效结果:每次会话都需重新查询)。
  2. 传入该ID调用
    model_schema_get
    接口:确认multiPrompt格式、duration可选值、元素配额等信息。
  3. 调用
    upload_asset
    上传起始帧和角色正面照(参见
    scenario
    技能)。
  4. 调用
    model_run
    接口,传入
    model_id
    、
    dry_run=true
    和以下
    parameters
    来预估费用:
    {"startImage": "asset_s", "multiPrompt": [{"prompt": "Wide shot, @Element1 crosses the plaza, slow dolly-in", "duration": "5"}, {"prompt": "Close-up, @Element1 smiles and says: 'we made it'", "duration": "4"}], "generateAudio": true, "elements": [{"frontalImage": "asset_f"}]}
    ;修改时长、档位或音频设置后需重新预估。
  5. 再次调用
    model_run
    接口,设置
    wait=false
    ,然后使用返回的任务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。
  6. 调用
    asset_display
    展示输出结果,带音审核两个镜头的效果。

Common mistakes

常见错误

  • Passing both
    prompt
    and
    multiPrompt
    to a V3 dedicated line: exclusive there (Omni alone keeps
    prompt
    required).
  • Numeric durations: outside Omni,
    duration
    is the string
    "5"
    , not
    5
    .
  • Expecting audio beside a last frame (2.6 I2V) or a reference video (Omni): exclusive pairs.
  • Carrying one member's caps to another: 15 seconds,
    elements
    , and 4K each exist on one member, not the next.
  • Compound camera moves ("dolly in while orbiting"): one dominant move per shot; split the rest across
    multiPrompt
    shots.
  • Overloaded negative prompts: a short artifact list steers; a long one stiffens motion.
  • Prompting Omni
    base
    to extend a clip: it edits in place, and no Kling member extends at authoring time. Continue from the clip's
    lastFrame
    id (
    asset_get
    ) as the start frame of a new I2V run, then concatenate, per
    scenario-video
    .
  • Skipping
    dry_run
    on 4K or avatar runs: at authoring time 4K ran several times Standard and avatar cost spanned a 100x range. Iterate on a Standard tier; no V3, O1, or 2.6 schema carries a seed, so a 4K re-run of the keeper is a new take, while upscaling it (
    scenario-video
    ) keeps the take.
  • 为V3专用系列模型同时传入
    prompt
    和
    multiPrompt
    :二者是互斥的(仅Omni模型要求必须传入
    prompt
    )。
  • 使用数字类型的duration:除Omni外,其他模型的
    duration
    必须是字符串
    "5"
    ,而非数字
    5
    。
  • 在使用尾帧(2.6 I2V)或参考视频(Omni)时期望开启音频:这些组合是互斥的。
  • 套用其他模型的配额限制:15秒时长、
    elements
    参数、4K分辨率这些特性分别只在特定模型中支持,并非所有模型都有。
  • 使用复合镜头运动(例如“边推进边环绕”):每个镜头建议只保留一个主要运动;其余运动可拆分到
    multiPrompt
    的多个镜头中。
  • 负面提示词过多:简短的瑕疵列表可以引导模型,过长的负面提示词会导致动作僵硬。
  • 提示Omni
    base
    模式延长视频:该模式是原地编辑,且撰写时所有Kling模型都不支持视频延长。可通过
    asset_get
    获取视频的
    lastFrame
    ID,作为新的I2V任务的起始帧生成后续内容,再按照
    scenario-video
    技能的说明拼接。
  • 在4K或数字人任务中跳过
    dry_run
    :撰写时4K的费用是标准版的数倍,数字人费用差异可达100倍。建议先在标准版上迭代;V3、O1、2.6版本的Schema都不支持种子(seed)参数,因此对满意的结果重新跑4K会生成新的版本,而通过
    scenario-video
    技能超分辨率则可以保留原版本。