Seedance Says First/Last Frame Cannot Be Mixed with Reference Media

Short Answer

Seedance only reads the role of each item in the request. It does not see which switches you ticked in your tool, and it does not read your prompt for this. An image without a role is treated as a first frame (image-to-video). Add a reference video to that, and the request becomes “first frame + reference media”, which the provider rejects. The typical error comes back as a 400 at submission time. No task is created and nothing is billed:

The parameter `content` specified in the request is not valid:
first/last frame content cannot be mixed with reference media content.

The fix: submit through the native Seedance endpoint POST /seedance/api/v3/contents/generations/tasks, set "role": "reference_image" on images and "role": "reference_video" on videos, and pass the video as a public URL the provider can download directly, or register it in the asset library and reference it as asset://.

A Real Case

A customer was using a local creative tool they had built. They passed one character image and one video, ticked “omni reference”, left “first/last frame” unticked, and even wrote “image 1 is not the first frame, do not use first/last frame mode” in the prompt. Every run still failed with the error above. We captured the raw request on the gateway side (image Base64 truncated):

{
  "model": "doubao-seedance-2-0-260128",
  "prompt": "Replace the person in video 1 with the character in image 1 ... (image 1 is not the first frame) ... do not use first/last frame mode",
  "duration": 14,
  "ratio": "9:16",
  "resolution": "720p",
  "images": ["data:image/jpeg;base64,/9j/4AAQ..."],
  "videos": ["/assets/input/ai_ref_xxxx.mp4"]
}

The request has two problems. Either one is enough to make it fail:

Problem	Explanation
The image has no `role`	The request went to the generic video endpoint `/v2/videos/generations`, which only has `images` and `videos` arrays and no field that can say “this is a reference image”. A single image is a first frame by rule, which conflicts with the reference video. The tool’s “omni reference” switch never made it into the request, and prompt text has no effect on parameter validation
The video is a local path	`/assets/input/ai_ref_xxxx.mp4` is a path inside the tool on the customer’s own computer. The provider’s servers cannot reach it, so even with the first-frame problem fixed, the video would never reach the model

The customer’s browser console also showed 415 Unsupported Media Type for the same mp4, coming from the tool’s own local preview endpoint (/api/media-preview/... on localhost). The tool had failed to turn the video into a usable address and put the local path into the request as-is.

Do Not Submit Seedance Through the Generic Video Endpoints

/v2/videos/generations (and /v1/videos, /v1/video/generations) are the gateway’s generic video endpoints. Their fields are the common subset shared by several video models, and they cannot express Seedance’s input modes:

No role: they cannot tell first frame, first/last frame and multimodal reference apart, so one image plus one video always ends up as “first frame + reference”
Resolution is not fully passed through: in our tests, 2.5 requested at 480p came out at 720p, and the 2.0 series requested at 1080p came out at 720p, while billing follows the resolution actually generated
Parameters unique to 2.5 (omni_reference_task_type, output_format, and so on) have no matching field

So always use the native endpoints for Seedance:

Step	Endpoint
Submit a task	`POST /seedance/api/v3/contents/generations/tasks`
Query a task	`GET /seedance/api/v3/contents/generations/tasks/{id}`

If you use a third-party tool, check whether its Seedance channel offers a “native” or “Volcengine Ark” format. A tool that only speaks the generic video API cannot run tasks with a reference video.

The Right Way: Set role on Every Item

import os, requests

BASE = "https://api.apiyi.com/seedance/api/v3/contents/generations/tasks"
headers = {"Authorization": f"Bearer {os.environ['APIYI_API_KEY']}"}

body = {
    "model": "doubao-seedance-2-0-260128",
    "content": [
        {"type": "text", "text": "Replace the person in @video1 with the character in @image1, keeping the motion, expressions and background music unchanged"},
        {"type": "image_url", "image_url": {"url": "https://cdn.example.com/character.png"},
         "role": "reference_image"},
        {"type": "video_url", "video_url": {"url": "https://cdn.example.com/source.mp4"},
         "role": "reference_video"},
    ],
    "resolution": "720p",
    "ratio": "9:16",
    "duration": 10,
}

r = requests.post(BASE, headers=headers, json=body, timeout=60)
print(r.status_code, r.json())   # on success: {"id": "cgt-..."}, then poll with that id

Key points:

The three input modes are mutually exclusive: first/last frame (2 images, first_frame / last_frame), first frame (1 image), and multimodal reference (reference_image / reference_video / reference_audio). As soon as there is a reference video or reference audio, every image must be reference_image
No role means first frame: a single image without role is the same as first_frame
In the prompt, refer to items by the order they were passed in, for example @image1 and @video1
Reference limits: up to 9 images + 3 videos + 3 audio clips for the 2.0 series, and up to 30 images + 10 videos + 10 audio clips for 2.5

How to Pass a Reference Video

The provider downloads your media on its own servers, so the video must be at an address the provider can reach directly.

Method	Works?	Notes
Public direct URL (preferred)	✅	Host it on object storage or a CDN (OSS, S3, R2, TOS, etc.) with no login, cookie or extra headers required, and keep it valid until the task finishes
Asset ID `asset://...`	✅	Use it when the same video is reused, or when real people appear in it. Assets must be mp4 / mov, 2–15 seconds, under 50MB. See Asset Library
Base64	⚠️ Not recommended	Videos are an order of magnitude larger than images, and inlining one in the request body is the easiest way to make submission time out. See Asset-First Workflow
A path on your machine (`/assets/...`, `C:\...`, `file://...`)	❌	The provider cannot read files on your computer
A private network address (`localhost`, `127.0.0.1`, `192.168.x.x`)	❌	Same as above
A link that needs a login to download	❌	The provider does not carry your login session when fetching

Before submitting, you can check the link from any machine with internet access (replace <URL> with your video link):

curl -s -o /dev/null -w 'code=%{http_code} type=%{content_type} size=%{size_download}\n' '<URL>'

You should see code=200, a type of video/mp4 or video/quicktime, and a size that matches the original file. An HTML page, a JSON body or any 4xx means the provider cannot use that link. For the full link check on images, see Image Link Opens but Fails.

Video Generation API

The native endpoint, plus the content layout and role values for each input mode

Asset-First Workflow

Why reference videos should not be sent as Base64, and how to register assets

Basics

Basic API

Image API (Official)

Video API (Official)

Multimodal Understanding API

Text Generation

Audio API (Official)

Retrieval & Safety

Seedance Says First/Last Frame Cannot Be Mixed with Reference Media

Short Answer

A Real Case

Do Not Submit Seedance Through the Generic Video Endpoints

The Right Way: Set role on Every Item

How to Pass a Reference Video

Video Generation API

Asset-First Workflow

​Short Answer

​A Real Case

​Do Not Submit Seedance Through the Generic Video Endpoints

​The Right Way: Set role on Every Item

​How to Pass a Reference Video

​Related Docs

Video Generation API

Asset-First Workflow

Short Answer

A Real Case

Do Not Submit Seedance Through the Generic Video Endpoints

The Right Way: Set role on Every Item

How to Pass a Reference Video

Related Docs