Docs / Model Capabilities

Model Capabilities and Parameters

Every video model on RelayDance is called through the same request shape (model, prompt, seconds, image, metadata), but the families behind those model IDs accept very different inputs. This page is the contract: for each family, which fields are honoured, which are forwarded as-is, which are silently ignored, and which make the request fail before anything is billed. When a field is not listed as supported for a family, assume it does nothing.

#Capability matrix

#Seedance 2.5

Model IDs: doubao-seedance-2-5-720p, doubao-seedance-2-5-480p. The most capable family: longest clips, the most reference media, and the only family that combines a character reference with a motion reference video in one request.

Honoured fields

Reference media roles

{
  "model": "doubao-seedance-2-5-720p",
  "prompt": "@image1 rises onto tiptoes, holds for a beat, then lowers the heels, matching the motion of the reference video",
  "seconds": "5",
  "metadata": {
    "ratio": "adaptive",
    "generate_audio": false,
    "content": [
      { "type": "image_url", "role": "reference_image", "image_url": { "url": "https://your-cdn.example.com/character.png" } },
      { "type": "video_url", "role": "reference_video", "video_url": { "url": "https://your-cdn.example.com/motion.mp4" } }
    ]
  }
}

#Seedance 2.0

Model IDs: doubao-seedance-2-0-480p, doubao-seedance-2-0-720p, doubao-seedance-2-0-1080p, doubao-seedance-2-0-4k, doubao-seedance-2-0-fast-260128 (720p, faster), doubao-seedance-2-0-mini-480p, doubao-seedance-2-0-mini-720p, plus seedance-1-5-pro-with-audio and seedance-1-5-pro-no-audio. Same request shape and the same metadata fields as 2.5, with tighter limits.

#Kling

Model IDs: kling-v2-6, kling-v2-6-pro, kling-v3, kling-v3-pro, kling-v3-omni, kling-v3-omni-pro. The -pro suffix selects the 1080p tier; without it you get 720p. Kling is a text-to-video and first-frame image-to-video route with a deliberately small surface: it does not take reference images, reference videos, audio, or any motion-transfer input. Requests that include those are rejected before billing rather than quietly downgraded to text-to-video.

Honoured fields

Rejected (HTTP 400, nothing billed)

• video_url or audio_url entries in metadata.content: there is no reference-video, motion-control or video-edit mode on this route
• More than one first-frame image, or an end frame without a first frame
• Images given as base64 data URLs or asset:// references: Kling needs a publicly reachable https URL

Ignored

generate_audio (Kling output has no audio track), seed, camera_fixed, watermark, output_format, resolution (locked by SKU). The Seedance @imageN syntax means nothing to Kling and is stripped from the prompt.

Output and billing

mp4, 24 fps, no audio. Standard tier renders 1280×720 (or the portrait / square equivalent), Pro renders 1080p; with a first frame the output keeps the image's aspect ratio. Typical turnaround is one to three minutes. Billing is per second of the normalised duration and is charged once at submission with no post-adjustment; a failed task is refunded in full. Download the result within 24 hours. Kling has a limited concurrency budget: an HTTP 429 means the queue is full, retry with a short backoff.

{
  "model": "kling-v3-omni",
  "prompt": "The boy rises onto his toes, holds briefly, then lowers his heels. Static camera, full body in frame.",
  "seconds": "5",
  "image": "https://your-cdn.example.com/character.png"
}

#MiniMax H3

Model IDs: minimax-h3-768p, minimax-h3-2k. Every clip comes with natively generated stereo audio; there is no silent option and generate_audio has no effect.

Ignored: resolution (locked by SKU), seed, camera_fixed, watermark, output_format, last-frame and video/audio references. Billed per second; failed tasks are not billed.

#Grok Video

Model IDs: grok-video-1.5, grok-image-video (lighter, cheaper). Image-to-video only: a request without an image is rejected. Output is 720p with no audio.

Ignored: generate_audio, seed, camera_fixed, watermark, last frame, reference images beyond the first, video and audio references. Billed per second of the requested length, charged at submission; failed tasks are refunded.

#HappyHorse

Model IDs: happyhorse-1.0-t2v, happyhorse-1.0-i2v, happyhorse-1.0-r2v, happyhorse-1.0-video-edit. The variant decides how the single image field is interpreted: first frame for i2v, reference image for r2v, and the source clip URL for video-edit. t2v takes the prompt only. Watermarks are always disabled. Other metadata keys are forwarded as given. Billed per second, charged at submission; failed tasks are refunded.

#What happens to fields a model does not support

If you are building a client that switches between families, key your request builder on the family, not on the shared request shape: a payload that is perfect for Seedance 2.5 is a text-only render on MiniMax and a hard error on Kling.

#Choosing a model

• Character consistency from a reference image plus a motion reference video: Seedance 2.5
• Clips longer than 15 seconds, or more than 9 reference images: Seedance 2.5
• Native 4K: Seedance 2.0 Pro, or Super-Resolution on an existing clip
• Generated soundtrack: Seedance 2.5 / 2.0 with generate_audio, or MiniMax H3 (always on)
• Animate one still image at a fixed 5 or 10 seconds, lowest cost: Kling standard tier
• Quick image-to-video at 6 to 15 seconds without audio: Grok Video

#Next steps

• Video Generation : Submit, poll and webhook flow
• Generate Video reference : Full request and response schema
• Image-to-Video example : Reference media walkthrough
• Error Codes : What a 400 or a failed task means