
Kling Video 4.0 Reference-to-Video API by Kuaishou
Kling 4.0 Omni Reference combines up to 15 images, videos, subjects and voices to generate a consistent 3-30 second video.
Kling Video 4.0 Reference-to-Video is developed by Kuaishou. Atlas Cloud (operated by Atlas Cloud AI LLC) provides access to it and does not own it. All trademarks belong to their respective owners.
Kling Video 4.0 Reference-to-Video
Kling Video 4.0 Reference-to-Video (Omni Reference) generates a 3–30 second video that stays consistent with everything you give it: reference images, reference videos, Kling subjects and voices — up to 15 references in one request.
Why Choose This?
- Omni reference. Mix up to 10 images, 5 videos, 7 subjects and 3 voices (15 items in total).
- Motion and style transfer. Reference videos carry motion rhythm, camera work or style into the new clip.
- Consistent characters. Define subjects (elements) from images or a short video clip in the request, and they stay consistent throughout the clip.
- Keyframes too. Optionally pin a first, middle or last frame.
Parameters
| Parameter | Required | Description |
|---|---|---|
| model | Yes | kwaivgi/kling-v4.0/reference-to-video |
| prompt | Yes | Scene description. Refer to inputs with @image_N, @video_N, @element_N, @voice_N. |
| reference_images | No | Reference images (URLs or Base64). |
| reference_videos | No | Feature reference videos. MP4 / MOV, 3–30s in total. |
| elements | No | Subjects created for this request from images or a video clip (see Notes). |
| voice_ids | No | Kling voice IDs (custom or preset), up to 3. |
| image / middle_images / last_image | No | Optional keyframes. |
| duration | No | Video length in seconds, integer 3–30. Default 5. |
| aspect_ratio | No | auto (default, adapts to the references), 16:9, 9:16, 1:1, 21:9. |
| resolution | No | Native 720p or 1080p (default); Tencent ESR 1080p-esr, 2k-esr, or 4k-esr. |
Reference numbering: images are numbered keyframes first (first → middle → last), then reference_images; videos, subjects and voices are numbered in array order. If the prompt mentions none of the reference images, videos, subjects or voices, all of them are referenced automatically.
How to Use
curl -X POST "https://api.atlascloud.ai/api/v1/model/generateVideo" \ -H "Authorization: Bearer $AIP_API_KEY" \ -H "Content-Type: application/json" \ --data-raw '{ "model": "kwaivgi/kling-v4.0/reference-to-video", "prompt": "The man in @image_1 walks through a park at golden hour, following the motion rhythm of @video_1. Photorealistic style.", "reference_images": ["https://p2-kling.klingai.com/kcdn/cdn-kcdn112452/kling-qa-test/man.jpeg"], "reference_videos": ["https://p2-kling.klingai.com/kcdn/cdn-kcdn112452/kling-qa-test/animals-video-5s.mp4"], "duration": 5, "aspect_ratio": "auto", "resolution": "1080p" }'
Generation is asynchronous. The response contains a request id; poll GET https://api.atlascloud.ai/api/v1/model/prediction/{id} until status is completed, then read the video URL from outputs.
Notes
- Preview. Kling 4.0 is in early access upstream; pricing and limits follow Kling's preview terms and may change.
- The Kling 4.0 API has no audio on/off or
multi_shotparameter; describe sound and shot structure in the prompt. durationmust be an integer from 3 to 30. Supported resolutions are720p,1080p,1080p-esr,2k-esr, and4k-esr.- ESR choices generate a Kling source video, then enhance it with Tencent ESR:
1080p-esruses a native 720p source;2k-esrand4k-esruse a native 1080p source. These are enhanced outputs. 2k-esrtargets 1440p (2560×1440 at 16:9);4k-esrtargets 2160p (3840×2160 at 16:9). The source aspect ratio and frame rate are preserved; portrait outputs use the corresponding short-edge target.- ESR pricing adds an enhancement charge for output seconds to the source-generation price. Video-input billing still applies to source generation; the enhancement charge is not multiplied by input-video duration. ESR requires an additional processing stage, so completion can take longer than native generation.
- The output video is transferred to AtlasCloud storage before it is returned.
- Subjects (
elements): describe each subject in the request withelement_name(and an optionalelement_description) plusreference_type—image_refer:frontal_image+ 1–3refer_images;video_refer: one 3–8srefer_videosclip of a real person. The subjects are created first and the video is generated right after, in the same request; this adds a short wait before generation starts. - Subjects are created anew for every request and are not saved: no subject ID is returned, and there is no
element_idto reuse. To keep a character consistent across generations, send the same subject images or clip again. - Limits per request: up to 10 images (keyframes + reference images), 5 videos (3–30s in total), 7 subjects (at most 3 video character subjects), 3 voices, and 15 items in total.
- The
refer_videosclips of video subjects count toward the billed input video duration.
Related Models
- Kling Video 4.0 Text-to-Video: generate from text only.
- Kling Video 4.0 Image-to-Video: animate from first, middle and last keyframes.
- Kling Video 4.0 Video-Edit: edit an existing clip.


















