Reference-to-Video API

reference-to-video is Seedance 2.0's most powerful mode. A single request can include up to 9 reference images + 3 reference videos + 3 reference audio clips, and the model composes a new video guided by all of them.

Typical scenarios:

• Style reference — a handful of images defining a specific art style; the new video mirrors that style
• Character / product reference — keep the same virtual character or product appearing in new scenes and actions
• Cinematography reference — a demo video that conveys the camera pacing and motion you want
• Music-driven pacing — a reference audio clip that drives the visual rhythm and mood
• Video editing / extension — continue, extend, or rewrite existing footage

Endpoint

POST https://api.evolink.ai/v1/videos/generations

Model ID: seedance-2.0-reference-to-video

The Fast variant is seedance-2.0-fast-reference-to-video — same parameter structure.

Request Parameters

Key constraint: image_urls, video_urls, and audio_urls can all be empty (equivalent to pure text-to-video), but providing only audio_urls is not allowed. Whenever audio is supplied, you must also provide at least one image or one video as a visual anchor.

Using the Prompt to Assign Roles

This model has no tag syntax (there are no @Image1, @Video1, or similar tags). You assign roles to each asset using natural language, and the model understands references like "image 1 / video 1 / audio 1" based on array order.

Common patterns:

You can freely combine these patterns in a single prompt. The order of the assets doesn't affect validity, but it does affect how the model interprets "image 1 / image 2" — keep it stable for reproducibility.

Input Asset Limits

Images

Videos

Audio

Overall

Request Examples

cURL — Three-modal composition (image + video + audio)

curl -X POST https://api.evolink.ai/v1/videos/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedance-2.0-reference-to-video",
    "prompt": "Replicate video 1's first-person perspective and camera pacing. Use audio 1 as the soundtrack for the entire video. Scene: a young rider weaving through a rain-soaked city street at night, neon reflections on wet asphalt.",
    "image_urls": ["https://example.com/rider-style.jpg"],
    "video_urls": ["https://example.com/pov-reference.mp4"],
    "audio_urls": ["https://example.com/synthwave-bgm.mp3"],
    "duration": 10,
    "quality": "720p",
    "aspect_ratio": "16:9"
  }'

Python — Images only (up to 9)

import requests

response = requests.post(
    "https://api.evolink.ai/v1/videos/generations",
    headers={
        "Authorization": "Bearer YOUR_API_KEY",
        "Content-Type": "application/json"
    },
    json={
        "model": "seedance-2.0-reference-to-video",
        "prompt": "The overall art style references the color palette and texture of the 3 provided images. Scene: a small-town summer market at dusk, warm tones.",
        "image_urls": [
            "https://example.com/style-ref-1.jpg",
            "https://example.com/style-ref-2.jpg",
            "https://example.com/style-ref-3.jpg"
        ],
        "duration": 8,
        "aspect_ratio": "16:9"
    }
)

task = response.json()
print(f"Task ID: {task['id']}")

Node.js — Video-only reference (camera replication)

const res = await fetch("https://api.evolink.ai/v1/videos/generations", {
  method: "POST",
  headers: {
    "Authorization": "Bearer YOUR_API_KEY",
    "Content-Type": "application/json"
  },
  body: JSON.stringify({
    model: "seedance-2.0-reference-to-video",
    prompt: "Replicate video 1's orbital camera movement and velocity curve. Subject: a classical sculpture in a museum hall at dusk.",
    video_urls: ["https://example.com/orbit-shot.mp4"],
    duration: 8,
    quality: "720p",
    aspect_ratio: "16:9"
  })
});

const task = await res.json();
console.log("Task ID:", task.id);

Response

{
    "id": "task-unified-1774857405-abc123",
    "object": "video.generation.task",
    "created": 1774857405,
    "model": "seedance-2.0-reference-to-video",
    "status": "pending",
    "progress": 0,
    "type": "video",
    "task_info": {
        "can_cancel": true,
        "estimated_time": 180,
        "video_duration": 10
    },
    "usage": {
        "billing_rule": "per_second",
        "credits_reserved": 60,
        "user_group": "default"
    }
}

Billing Notes

• Per-second billing based on the output video's duration
• Reference video input duration also counts toward billing (a 10-second reference video bills at 10 seconds of input)
• Audio generation itself is free of extra charge

FAQ

Do the reference assets appear directly in the output? No. The model treats them as signals for style / composition / motion / rhythm; the final output is fully generated new content.

Can I send the request without any reference assets? Yes — this acts like pure text-to-video. But if you have no references, use the cheaper text-to-video directly.

Does asset order matter? Yes. If your prompt says "video 1", the model maps that to video_urls[0]. Keeping a stable order makes experiments reproducible.

Related

• Models Overview
• Text-to-Video API
• Image-to-Video API
• Fast Models — seedance-2.0-fast-reference-to-video
• Async Tasks / Webhooks