Kling 4.0 API reference

Precios

Without video input (text, reference images, image subjects, voices, keyframes) the billed seconds equal the output length. With a reference video or video subject the billed seconds are output + max(total input video, output); each clip is rounded up to whole seconds and measured by us when the task is created. Reference images, subject images and keyframes are not billed separately; the smallest request (720p, 3 s) is $0.471. Charged only on success; failed generations are not charged.

Autenticación

All requests carry the API key in the header:

Authorization: Bearer YOUR_API_KEY

Generate a video

# Reference-to-video: images + a reference video, mentioned by position
curl -X POST https://api.apimodels.app/v1/video/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kling-4.0",
    "prompt": "The dancer in @image_1 performs the routine from @video_1 on the rooftop in @image_2 at dusk",
    "image_urls": ["https://example.com/dancer.png", "https://example.com/rooftop.jpg"],
    "video_urls": ["https://example.com/routine.mp4"],
    "keep_original_sound": false,
    "duration": 10,
    "resolution": "720p",
    "aspect_ratio": "9:16"
  }'

# Keyframe-to-video: first + middle + last frame (a 5 s clip takes up to 6 frames)
curl -X POST https://api.apimodels.app/v1/video/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kling-4.0",
    "prompt": "The girl in @first_frame walks out, crosses the street in @keyframe_1 and arrives at the party in @last_frame",
    "first_frame_url": "https://example.com/first.png",
    "keyframe_urls": ["https://example.com/street.png"],
    "last_frame_url": "https://example.com/party.png",
    "duration": 5,
    "resolution": "1080p"
  }'

# Response (async — a task is created):
# { "code": 200, "msg": "success", "data": { "taskId": "clxxx", "state": "pending" } }

# Poll until completed (result files are kept 30 days)
curl "https://api.apimodels.app/v1/video/generations?task_id=TASK_ID" \
  -H "Authorization: Bearer YOUR_API_KEY"
# { "code": 200, "data": { "state": "completed",
#     "resultUrls": ["https://r2.apimodels.app/videos/clxxx.mp4"] } }

Two routes, one model id

• Reference-to-video (including text-only) — prompt alone is text-to-video; add any mix of image_urls (≤10, @image_1…), video_urls (≤5, @video_1…, 2-30 s each, ≤30 s together with video subjects), elements (≤7, @Name) and voices (≤3, @Name), 15 materials at most. keep_original_sound decides whether the reference videos keep their own audio. Without aspect_ratio the clip follows the references, or 16:9 for a prompt alone.
• Keyframe-to-video — first_frame_url (image and image_url also work), keyframe_urls (≤8 middle frames, in time order) and last_frame_url, mentioned as @first_frame, @keyframe_1 and @last_frame. Total frames ≤ min(10, duration + 1): 4 at 3 s, 5 at 4 s … 10 from 9 s up. The clip follows the first frame's aspect ratio.
• Frames cannot be combined with image_urls, video_urls, elements or voices; mixing them returns 400. Every @image_N, @video_N, @keyframe_N, @first_frame and @last_frame in the prompt must match an input in the request, numbered by array position; every @Name must match a name in elements or voices.

Named subjects (elements) and voices

elements and voices are API-only for now (the playground covers reference images, reference videos and keyframes). Image subject: image_urls with 2-4 images, the first being the main view. Video subject: video_url with one 3-8 second single-shot character clip, shortest side at least 700 px; at most 3 video subjects. tag is optional: character, animal, prop, costume, scene, effect or others (video subjects can only be character); description up to 100 characters; voice_url only on character subjects. voices are standalone {name, audio_url} clips, up to 3, and must contain a recognizable voice. Names are 1-20 characters (Chinese, letters, digits, underscore), unique across elements and voices regardless of case, and cannot be image_N, video_N, keyframe_N, object_N, audio_N, first_frame or last_frame.

{
  "model": "kling-4.0",
  "prompt": "@Hero opens the door of @image_1 and greets @Robot; @Robot answers in the voice of @Narrator",
  "image_urls": ["https://example.com/door.jpg"],
  "elements": [
    {
      "name": "Hero",
      "image_urls": ["https://example.com/hero-front.jpg", "https://example.com/hero-side.jpg"],
      "tag": "character",
      "description": "Young man in a red jacket",
      "voice_url": "https://example.com/hero-voice.mp3"
    },
    {
      "name": "Robot",
      "image_urls": ["https://example.com/robot-front.png", "https://example.com/robot-back.png"],
      "tag": "prop"
    }
  ],
  "voices": [
    { "name": "Narrator", "audio_url": "https://example.com/narrator.mp3" }
  ],
  "duration": 10
}

Parámetros

Notas

• Sound is always generated and cannot be turned off: generate_audio: false returns 400. With reference videos, keep_original_sound is the only audio switch.
• Negative prompts, multi-shot storyboards and video extension are not supported; output stops at 1080p. Kuaishou has announced 4K and 10-bit HDR, but neither is available yet.
• Reference images need each side to be at least 300 px and an aspect ratio between 0.4 and 2.5. The top-level image_urls, video_urls, first_frame_url, keyframe_urls and last_frame_url take public URLs or data URIs; media inside elements and voices must be public http(s) URLs.
• Reference video lengths are measured when the task is created; a video whose length cannot be read (login wall, expired link, unsupported format) returns 400 instead of being billed on a guess.
• Asynchronous: create returns a taskId; poll GET /v1/video/generations?task_id= every 10-15 seconds until completed, or pass callback_url. Result files are deleted after 30 days, so download or re-host promptly.
• Content safety checks run upstream; rejected generations are not charged. Only upload faces and voices you have the rights to use.

Try it in the Playground. What is new, worked price examples and a comparison with Kling V3, Kling V3 Omni and Vidu Q4 are on the Kling 4.0 model page.