Kling 3.0 Motion Control API: transfer a reference video's motion and audio onto one character image for AI dance videos and talking avatars. Try it online.
Provider: Kuaishou
Category: video generation
Endpoint: /v1/tasks
Status: Available
Cost: See live page pricing
Kling 3.0 Motion Control API: transfer a reference video's motion and audio onto one character image for AI dance videos and talking avatars. Try it online.
Pricing
Standard Usage
Ready for generation
Configure your parameters and click "Run" to see the output here.
REAL API RESULTS
Every example is a real result generated through the HiAPI endpoint: one character reference image plus one motion reference video in, one AI video out with identity from the image and motion plus audio from the clip.
Talking gesture
The portrait supplies identity and wardrobe; the clip supplies lip sync, gestures and delivery. The output keeps the sweater and face, follows the clip's timing and carries its original audio track.
Prompt
The woman presents directly to camera in a warm studio, steady framing, natural expression
Dance cover
Large multi-directional motion with fast footwork: identity and outfit follow the image, timing, beat and ambience follow the clip. This is the video-orientation path, which allows longer references.
Prompt
The dancer performs the street groove on the rooftop at dusk, camera steady, keep the outfit and city backdrop
Explainer delivery
With character_orientation=image the character keeps the orientation of the photograph, so the presenter never turns away from the lens. This is the fixed-camera path for lessons, product walkthroughs and spoken explainers.
Prompt
The man explains at his desk in the bright office, camera holds his orientation from the photo, calm delivery
When you only have a still to animate, need several references composed into a new shot, or want a script-driven talking avatar, these video generation APIs on HiAPI cover it.
Kling 3.0 Motion Control is the motion control model in the Kuaishou Kling AI 3.0 family, an image-to-video model driven by a motion reference: upload one character image and one reference video, the image decides who the character is and what they wear, the clip decides how they move, emote and pace. The output is an AI video of the pictured character re-enacting the reference motion, with no motion prompt engineering required.
Compared with Kling 2.6 Motion Control, 3.0 keeps facial identity steadier through head turns, brief occlusion and longer sequences and can keep the reference audio, which suits AI dance videos, talking avatars, mascot animation and previsualization: same motion, different performer.
Playground and API share one model ID: kling-3.0/motion-control. character_orientation chooses whether the character faces like the clip or like the image and sets the length cap; keep_original_sound controls whether the clip's audio is carried over.
These specifications match the current HiAPI request contract: input limits, output length, resolution and per-second pricing, settled on the actual output length.
Character image
Motion clip
Output length
Resolution
Audio
Model ID
Endpoint
Three steps to a motion transfer video: prepare the inputs, choose the parameters, submit the task. Validate with a short clip at 720p first, then move to 1080p or a longer reference.
Use a portrait with head, shoulders and torso visible, and a 3–30s single-shot clip of one unobstructed person. Upload both and pass them as image_url and video_url.
Pick video orientation for big or multi-angle moves (≤30s) and image orientation for fixed-camera talking (≤10s). Validate at 720p, then render 1080p; disable keep_original_sound when you do not want the clip audio.
Create an async task with kling-3.0/motion-control, then receive the result via callback.url or poll GET /v1/tasks/:id. Billing settles on the real output seconds; generation usually takes a few minutes.
One POST /v1/tasks request submits a motion transfer video task; the cURL, Python and Node.js samples run as-is. See the API docs for every parameter, callbacks and error handling.
curl -X POST "https://api.hiapi.ai/v1/tasks" \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "kling-3.0/motion-control",
"input": {
"image_url": "https://static.hiapi.ai/model-examples/kling-3.0-motion-control/v3/2026/09/28/1790607181108-36686f5400675348-p1.png",
"video_url": "https://static.hiapi.ai/model-examples/kling-3.0-motion-control/v3/2026/09/28/1790608051171-c882228bfe66fe03-c1.mp4",
"resolution": "1080p",
"character_orientation": "video",
"keep_original_sound": true,
"prompt": "The woman presents directly to camera in a warm studio, steady framing, natural expression"
}
}'KLING 3.0 MOTION CONTROL API
Kling 3.0 Motion Control is the motion control video generation model in Kuaishou's Kling 3.0 family. It transfers the motion, expressions and delivery of a reference video onto a single character image: the image decides who the person is and what they wear, the clip decides how they move, speak and pace the performance. Supply those two inputs and the model handles the rest, with no motion prompt engineering.
The output follows the clip's length and timing and keeps its native audio, so spoken delivery reuses the original track and dance covers keep the on-site beat without a dubbing pass. Identity holds through head turns, raised hands and brief occlusion.
The sections below cover six capabilities with one asset set: how motion transfer splits identity and performance, one character with a new motion, one motion on a new character, the two character_orientation modes, keep_original_sound audio transfer, and what the reference material needs to look like. Every example is a real result generated through HiAPI.
Each group answers one question: who appears, what they perform, how the camera behaves, and what happens to the sound. The assets come from the same real task set and can be compared directly.
01
The portrait carries identity only: the face, the hair, the polo shirt and the office behind it. The clip carries performance only: the rhythm of speech, when the head turns, when the smile lands, how the hands move. Keeping them separate lets one character appear in many clips without the likeness drifting.
That split is also the generation order: an image model produces the portrait, a video model turns it into a performance reference, and this model performs the transfer. Every example on this page was produced in that order.
Record the script once, then re-skin it across several likenesses or languages with the original tone and pauses intact.
Put a trending routine onto your own character: motion and beat come from the clip, the look comes from the image.
Keep one character sheet and let it perform across many clips with a consistent face.
Rehearse motion and timing on cheap material, then move to higher resolution or a longer reference.
Sibling models cover different starting material: a still only, several references composed into a new shot, or a script-driven avatar each have a better fit.
Kling 3.0 Motion Control
You have one portrait and one real performance clip and want that performance transferred.
Kling 3.0 图生视频
You only have a still and want it animated from a prompt, with no reference performance.
Seedance 2.5 参考生视频
You want several images, clips or audio tracks composed into a brand-new shot rather than one performance replayed.
HeyGen Avatar V
The picture is driven by a script and synthesized speech for standardized avatar delivery.
One API key covers both the Playground and production calls: find a combination you like in the page, then move the same request body into your service without switching accounts or endpoints. Tasks are submitted asynchronously and can report their terminal state through a callback or a compensating query.
Billing settles on the actual output seconds: a hold sized by the orientation cap is taken at creation, then adjusted to the real duration, with no separate charge for the reference material. Outputs are hosted for you to download or route into your own storage.
Video, image and audio models share one credential and usage view.
The hold follows the cap and is corrected to the real output seconds.
Upload a portrait and a reference clip in the Playground, pick the orientation and resolution, and get a clip with its original audio in a few minutes.
Two required inputs: a character reference image (image_url) and a 3-30 second motion reference video (video_url). The image defines appearance and identity; the clip defines motion, expressions and timing. A prompt is optional for scene, style and camera guidance.
No duration parameter is needed. Output follows the reference clip: up to 30s with character_orientation=video and up to 10s with image. Very fast or complex motion may yield a shorter clip because only usable motion segments are extracted.
video: the character faces the way the reference clip does, best for dance, walking and multi-angle motion. image: the character keeps the orientation of the reference image, best for fixed-camera talking or explainer content, capped at 10s.
mp4/mov, up to 100MB, 3-30 seconds, one continuous shot with the subject's head, shoulders and torso clearly visible; prefer a single person, no cuts and no occlusion. The image also needs head, shoulders and torso visible, both sides over 340px, aspect ratio between 2:5 and 5:2.
Yes by default (keep_original_sound=true), which suits talking and dance covers that reuse the original track; pass false for video-only output.
Per actual output second, with different rates for 720p and 1080p. A hold sized by the orientation cap is taken at creation and adjusted to the real output seconds after completion; see the live pricing block on this page.
02
Changing the performance means changing only the clip. The portrait stays fixed while the dance reference pulls the same face into a different motion range and tempo, with wardrobe, hair and features preserved, which suits series content built around one character.
03
The reverse also holds: keep the clip and swap the portrait and someone else delivers the same speech and gestures. One recorded performance can be re-skinned across several likenesses for recasts, character variants, or the visual half of a multi-language dub.
04
character_orientation=video lets the character face the way the clip does, which suits dance, blocking and large motion, up to 30 seconds. character_orientation=image holds the orientation of the photograph, which keeps the frame steadier for fixed-camera explainers and presentations, up to 10 seconds.
The same material behaves differently in each mode: following the clip absorbs more complex motion, while following the photo is better for product demos and lesson videos where the subject must not turn away.
05
keep_original_sound defaults to on, so the clip's audio track lands in the output unchanged: the tone, pauses and breath of a spoken take need no re-recording, and dance footage keeps its on-site beat, ready to cut or caption.
The waveform below comes from the example output's own audio track and shows where the spoken passages sit. Pass false when you want picture only.
06
The clip should be mp4/mov, up to 100MB, 3 to 30 seconds, one continuous shot, with the subject's head, shoulders and torso clearly visible, one person in frame and few cuts or occlusions. The image needs the same head, shoulders and torso framing, both sides over 340px, and an aspect ratio between 2:5 and 5:2.
When motion is extremely fast or the shot changes often, the model extracts only the usable segments and the output can be shorter than the clip, which is why settlement is based on the actual output seconds.
Example inputs and golden outputs are downloadable for verification.