Scenario Kling Video

Overview

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.

Quick reference

Pick by job; discover ids with search (target="models", query="kling", public=true):

Job	Members
Every mode in one model, tier via mode	V3 Omni (standard 720p, pro 1080p, 4k)
Text to video at a fixed tier	V3 T2V Standard / Pro / 4K, 2.6 T2V Pro
Animate a still, optional last frame	V3 I2V Standard / Pro / 4K, O1 I2V, 2.6 I2V Pro
Consistent characters from photos	O1 Reference Images, V3 I2V elements
Edit or restyle footage by prompt	O1 Video Editing / Reference Video, Omni (videoReferenceType: "base")
Motion from a video onto a character image	Motion Control (V3 Pro / Std, V2.6)
Talking head	AI 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.

Audio is a per-member switch

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.

Motion control reads two inputs

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.

Worked example: two-shot character clip from a still

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.

Common mistakes

• 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.