Doubao Seedance 2.0 Series Tutorial
The Doubao Seedance 2.0 series (referred to as the Seedance 2.0 series below) models support multi-modal content input including images, videos, audio, and text, with capabilities for video generation, video editing, and video extension. They can faithfully reproduce object details, voice timbre, effects, styles, and camera movements while maintaining stable character features, giving users director-level control. This document introduces the exclusive capabilities of the Seedance 2.0 series models, helping you quickly implement Video Generation API calls.
- Register a platform account and log in to the platform console.
- Obtain an API Key: Create an API Key in the platform console and keep it safe. Do not disclose it to others (see [API Authentication](/en/api-guides/authentication)).
Model Capabilities
The Seedance 2.0 series currently includes Doubao Seedance 2.0 (referred to as Seedance 2.0 below), Doubao Seedance 2.0 Fast (referred to as Seedance 2.0 Fast below), and Doubao Seedance 2.0 Mini (referred to as Seedance 2.0 Mini below). The three models support basically the same features; the main difference lies in the trade-off between generation quality and cost:
For the highest generation quality, Seedance 2.0 is recommended;
To balance cost and generation speed without demanding ultimate quality, Seedance 2.0 Fast is recommended;
For lower cost, Seedance 2.0 Mini is recommended.
| Model Name | Seedance 2.0 | Seedance 2.0 Fast | Seedance 2.0 Mini | |
|---|---|---|---|---|
| Model ID | doubao-seedance-2-0-260128 | doubao-seedance-2-0-fast-260128 | doubao-seedance-2-0-mini-260615 | |
| Text-to-video | ✓ | ✓ | ✓ | |
| Image-to-video (first frame) | ✓ | ✓ | ✓ | |
| Image-to-video (first & last frame) | ✓ | ✓ | ✓ | |
| Full-modal reference【New】 | Image reference | ✓ | ✓ | ✓ |
| Video reference | ✓ | ✓ | ✓ | |
| Combined reference * Image + audio * Image + video * Video + audio * Image + video + audio | ✓ | ✓ | ✓ | |
| Edit video【New】 | ✓ | ✓ | ✓ | |
| Extend video【New】 | ✓ | ✓ | ✓ | |
| Generate audio-containing video > "generate_audio": "true" | ✓ | ✓ | ✓ | |
| Web search tool【New】 | ✓ | ✓ | ✓ | |
| Draft mode | ✗ | ✗ | ✗ | |
| Return the last frame of the video output > "return_last_frame": > "true" | ✓ | ✓ | ✓ | |
| Output video specs | Output resolution > "resolution": "720p" | * 480p (8-bit color depth) * 720p (8-bit color depth) * 1080p (8-bit color depth) * 4k (10-bit color depth) | * 480p (8-bit color depth) * 720p (8-bit color depth) | * 480p (8-bit color depth) * 720p (8-bit color depth) |
| Output aspect ratio > "ratio":"16:9" | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 | |
| Output duration > "duration": 5 | 4~15 seconds | 4~15 seconds | 4~15 seconds | |
| Output video format | mp4 | mp4 | mp4 | |
| Offline inference > "service_tier": "flex" | ✗ | ✗ | ✗ | |
| Online inference rate limits | Max RPM | Non-4k resolution: * Enterprise users: 600 * Individual users: 180 4k resolution: * Enterprise users: 15 * Individual users: 15 | * Enterprise users: 600 * Individual users: 180 | * Enterprise users: 600 * Individual users: 180 |
| Max concurrency | Non-4k resolution: * Enterprise users: 10 * Individual users: 3 4k resolution: * Enterprise users: 1 * Individual users: 1 | * Enterprise users: 10 * Individual users: 3 | * Enterprise users: 10 * Individual users: 3 | |
| Offline inference rate limits | TPD | - | - | - |
Basic Usage
Full-modal Reference
Input text, reference images, videos (with audio tracks), and audio to generate a new video. It can inherit the character image, visual style, and composition from reference images; the subject content, camera movement, actions, and overall style from reference videos; and core information such as voice timbre, music melody, and dialogue from reference audio.
Preview (visit the model card for more examples):
| Input: Text | Input: Images, video, audio | Output |
|---|---|---|
| Use the first-person perspective composition of Video 1 throughout, and use Audio 1 as the background music throughout. A first-person fruit tea advertisement, Seedance brand "Ping Ping An An" apple fruit tea limited edition; the first frame is Image 1, your hand picks an Aksu red apple with morning dew, with crisp apple collision sounds; 2-4 seconds: quick cuts, your hand puts apple pieces into the shaker, adds ice and tea base, shakes hard, ice collision sounds synced with upbeat drums, background sound: "Freshly cut and shaken"; 4-6 seconds: close-up of the finished product in first person, layered fruit tea poured into a transparent cup, your hand gently spreads milk foam on top, sticks a pink logo on the cup, camera zooms in on the layered texture of the milk foam and fruit tea; 6-8 seconds: holding the cup in first person, you bring the fruit tea from Image 2 to the camera (simulating the perspective of handing it to the viewer), the logo on the cup is clearly visible, background sound "Take a fresh sip", the last frame freezes on Image 2. The background voice is a female voice throughout. |
- You can freely combine the following modal contents. Note that "text + audio" and "pure audio" input are not supported.
- Text
- Images: 0~9
- Videos: 0~3
- Audio: 0~3
- Advanced usage: Full-modal video generation can specify a reference image as the first/last frame via the prompt, indirectly achieving the "first & last frame + full-modal reference" effect. To strictly ensure the first/last frames match the specified images, prefer image-to-video (first & last frame) (configure role as first_frame/last_frame).
- For input requirements of each modality, see Full-modal Reference.
Edit Video
You can provide the video to be edited, reference images, or audio, combined with prompts, to complete various video editing tasks, such as: replacing the video subject, adding/removing/modifying objects in the video, partial repainting/restoration, etc.
Preview (visit the model card for more examples):
| Input: Text | Input: Video & image | Output |
|---|---|---|
| Replace the perfume in the gift box in Video 1 with the face cream in Image 1, keeping the camera movement unchanged |
Extend Video
Extend the original video forward or backward, or concatenate multiple video clips (up to 3 clips) into one coherent video.
Preview (visit the model card for more examples):
| Input: Text | Input: Videos to be extended | Output |
|---|---|---|
| The arched window in Video 1 opens, entering the gallery interior, then connects to Video 2, after which the camera enters the painting, connecting to Video 3 |
- When extending 1 video forward or backward, the generated video generally only contains the ending frames of the original video. However, you can flexibly control it via the prompt to include the original video content. For example: Extend Video 1 forward, [description of the extended content...], and finally connect back to Video 1.
- Pass in 2~3 videos to fill in the transition parts in between; the generated video will contain both the original video content and the newly generated video content.
Output 4k Video
Only supported by Seedance 2.0
Seedance 2.0 supports 4k video output with 10-bit color depth encoding, fully preserving rich color gradations and smooth gradient transitions, meeting the requirements of professional film production and HDR video content.
| Preview 1 | Preview 2 |
|---|---|
Note: The demonstration videos are stitched from multiple shots generated by Seedance 2.0, not directly generated by the sample code below.
Using Web Search
The web search capability only applies to pure text input
Configure the tools.type parameter to web_search to use the web search tool.
After enabling web search, the model will autonomously decide whether to search internet content (such as products, weather, etc.) based on your prompt. This improves the timeliness of the generated video but also adds some latency.
The actual number of searches can be obtained from the usage.tool_usage.web_search field returned by the Query Video Generation Task API; 0 means no search was performed.
| Input: Text | Output |
|---|---|
| Macro lens focuses on an emerald green glass frog on a leaf. The focus gradually shifts from its smooth skin to its completely transparent belly, where a bright red heart is beating powerfully and regularly. Note Search the web for the appearance features of the glass frog. |
More Capabilities
The Seedance 2.0 series models also support common basic capabilities such as text-to-video, first frame image-to-video, first & last frame image-to-video, and custom video output specifications (including: resolution, aspect ratio, video duration, and whether the video contains a watermark).
Effortless Creation
The Seedance 2.0 series models do not support directly uploading reference images/videos containing real human faces. To facilitate the use of portraits, the platform provides the following solutions. For detailed tutorials, see Effortless Creation of Portrait Videos:
| Solution | Description |
|---|---|
| Trust model outputs as input materials | Original outputs containing human faces generated by some models under this account can be used as input materials to call the Seedance 2.0 series models again for secondary creation, without triggering input moderation interception. |
| Use preset virtual avatars | The platform provides a preset virtual avatar library, offering creators free, compliant, and diverse portrait materials. Suitable for scenarios that require realistic-style faces without specifying a particular person, pursuing zero compliance risk and fast creation. |
| Use authorized real-person materials | Supports using authorized real-person portrait materials for video generation. |
Prompt Tips
Prompt Skill
The platform provides the Seedance 2.0 prompt optimization skill to help you fine-tune your prompts.
Configuration: It is recommended to run the following command in your local project to install
sd2-pevia the Skills catalog:bashnpx --yes skills@latest add "https://arkdocs.tos-cn-beijing.volces.com/skills/" \ --skill sd2-pe \ --yesFor manual installation, download the following
SKILL.mdfile. Taking OpenClaw as an example, copy the full content of the file into the chat input box and send "Please help me install this skill", then wait for the tool to complete the installation.Usage: Enter
/sd2-pe + your promptin the AI chat box to start tuning your prompt.
Prompt Rules
The prompt must reference materials using the "material type + sequence number" format, where the sequence number is the ordering of that material among materials of the same type in the request body. For example, 「Image n」 refers to the n-th reference image with
type="image_url"in thecontentarray (counting from 1 in array order). Note that referencing materials by Asset ID is not supported.For prompt formulas and detailed rules of different tasks, refer to the Doubao Seedance 2.0 series prompt guide.
Usage Limits
See the "Usage Limits" section of the Seedance 2.0 series model documentation.
FAQ
10-bit Color Depth and H.265/HEVC Encoded Video Playback Compatibility
The following are the playback compatibility test results of 10-bit color depth and H.265/HEVC encoded videos across browsers and players on various platforms. Actual results may vary depending on device configurations.
Recommended:
macOS: Safari and Chrome are recommended browsers; VLC, mpv, and QuickTime Player are recommended players
Windows: Edge and Chrome are recommended browsers; VLC and mpv are recommended players
Windows Platform
About conditional support: Higher hardware decoding capability is required. Known to play normally on Intel i7 + NVIDIA RTX 4070 + Windows 11 or higher configurations; for other configurations, please test yourself.
macOS Platform
About conditional support: Higher hardware decoding capability is required. Known to play normally on Apple M2 or later models; for M1 and earlier models, please test yourself.
Flickering in Generated Videos
Typical symptom
In first frame image-to-video and first & last frame image-to-video scenarios, some frames of the generated video show stretching, compression, and other flickering issues.
Root cause
The resolution and aspect ratio of the input image do not match those of the output video, causing frame-to-frame flickering.
Solution
Crop the input image: refer to the pixel dimension table supported by the Seedance 2.0 series models (see the ratio field in Create Video Generation Task API), and crop the input image to the target pixel dimensions.
Set the ratio field of the API to
adaptive.Use the Seedance 2.0 series models to start the first frame / first & last frame image-to-video task again.