Kling Video Generation

Generate AI videos through AceDataCloud's Kuaishou Kling API.

Setup: See [authentication](../_shared/authentication.md) for token setup.

Quick Start

curl -X POST https://api.acedata.cloud/kling/videos \
  -H "Authorization: Bearer $ACEDATACLOUD_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"action": "text2video", "prompt": "a cat playing piano on a rooftop at sunset", "model": "kling-v3", "mode": "std", "duration": 5}'

Async: See [async task polling](../_shared/async-tasks.md). Poll via POST /kling/tasks with {"id": "..."}.

Models

Quality Modes

Workflows

1. Text-to-Video

POST /kling/videos
{
  "action": "text2video",
  "prompt": "a futuristic city with flying cars",
  "model": "kling-v3",
  "mode": "std",
  "duration": 5,
  "aspect_ratio": "16:9"
}

2. Image-to-Video

Animate a still image. Optionally specify an ending frame.

POST /kling/videos
{
  "action": "image2video",
  "prompt": "the scene slowly comes alive with movement",
  "start_image_url": "https://example.com/scene.jpg",
  "end_image_url": "https://example.com/end-scene.jpg",
  "model": "kling-v3",
  "mode": "pro"
}

3. Omni References

Use kling-o1 or kling-v3-omni with reference images and/or one reference video. Cite each item in the prompt using its one-based token.

POST /kling/videos
{
  "action": "text2video",
  "prompt": "turn <<<video_1>>> into hand-painted animation while preserving its motion",
  "model": "kling-o1",
  "mode": "std",
  "duration": 5,
  "video_list": [
    {
      "video_url": "https://example.com/source.mp4",
      "refer_type": "base",
      "keep_original_sound": "no"
    }
  ]
}

Use refertype: "feature" to reference style, motion, or a neighboring shot. Use refertype: "base" to edit the supplied video. A base video cannot be combined with first/end frames.

4. Extend Video

Continue an existing video with additional seconds.

POST /kling/videos
{
  "action": "extend",
  "video_id": "existing-video-id",
  "prompt": "the camera pulls back to reveal the full landscape",
  "model": "kling-v2-5-turbo"
}

5. Motion Control

Apply precise camera/motion control from an image + reference video.

POST /kling/motion
{
  "image_url": "https://example.com/subject.jpg",
  "video_url": "https://example.com/motion-reference.mp4",
  "mode": "std",
  "character_orientation": "image"
}

6. Lip Sync

Create a lip-synced video from a source video plus either an audio track or input text.

POST /kling/lip-sync
{
  "video_url": "https://example.com/source.mp4",
  "mode": "audio2video",
  "audio_url": "https://example.com/voiceover.mp3"
}

7. Talking Photo

Animate a still portrait from an image plus an audio track.

POST /kling/talking-photo
{
  "image_url": "https://example.com/portrait.jpg",
  "audio_url": "https://example.com/voiceover.mp3",
  "model": "kling-v2-1-master",
  "duration": 5,
  "mode": "pro"
}

Parameters

Gotchas

• kling-o1 supports duration=5 only; kling-v3 and kling-v3-omni support flexible 3–15 seconds; most other models support 5 or 10
• mode=4k is only available for kling-v3 and kling-v3-omni and is incompatible with camera_control
• generate_audio enables synchronized audio generation (supported by kling-v3, kling-v3-omni, and kling-v2-6 in pro mode)
• endimageurl is only for image2video action — it defines the last frame
• Omni references are supported only by kling-o1 and kling-v3-omni; cite them as <<<imageN>>> / <<<video1>>>
• Omni reference requests do not support negativeprompt, cfgscale, camera_control, or mode=4k
• With videolist, generateaudio must be false; a base video cannot be combined with first/end frames
• elementlist is intentionally unavailable because upstream Element IDs are not tenant-scoped; use imagelist for subject references
• Motion control (/kling/motion) is a separate endpoint from video generation
• Lip-sync is a separate endpoint (/kling/lip-sync) and requires mode; use audio_url for audio2video or text + voice fields for text2video
• Talking-photo is a separate endpoint (/kling/talking-photo) and requires both imageurl and audiourl
• pro mode costs roughly 2x std mode but generates faster with better quality
• Task states use "succeed" (not "succeeded") — check for this value when polling
• negative_prompt helps avoid unwanted elements (e.g., "blurry, low quality, text")