Seedance 2.5 API · Volcengine
Seedance 2.5 is a ByteDance video model for generating, editing and extending clips from text and image, video or audio references. Current Volcengine specifications offer 480p, 720p, 1080p output. Use XiuAI to review applicable prices and access the asynchronous video API.[1][2][5]
doubao-seedance-2-5-260628Model overview
Published provider specifications. Available features depend on the current API route.
- Input
- Text, image URLs, video URLs and audio URLs
- Output
- Video with optional synchronized audio
- Duration
- 4–30 seconds
- Resolution
- 480p / 720p / 1080p
- Reference limits
- 30 images · 10 videos · 10 audio clips
- Frame rate
- 24 fps
- API routes
- Ark API / Unified video API
Where Seedance 2.5 fits
For product or scene demonstrations, describe the subject, motion and camera movement. Use a first-frame image when you already have a starting composition. With multiple references, explain whether each input should guide appearance, action, camera movement or sound.[1][2]
Existing footage can be used for editing or extension. Describe what to add, remove or replace. Inspect product shapes, readable text and motion continuity before using the result; more references do not guarantee a better output. Your application handles version selection, review and storage.[1][2][6]
Choosing between 2.5 and Flash
Start with output requirements and reference counts. Seedance 2.5 offers 480p / 720p / 1080p, while Flash offers 480p / 720p. Seedance 2.5 accepts up to 30 images, 10 videos and 10 audio clips, subject to input-combination requirements.[2]
If Flash meets your delivery requirements, compare current rates and published examples under the same input and output conditions. Evaluate processing time and output quality with your own requests; a model name does not guarantee a fixed turnaround or quality level.[2][5]
Input modes and key parameters
Text-to-video, first-frame, first/last-frame and multimodal reference generation use different input roles. Choose the mode before assigning roles in content. Do not mix first/last-frame roles with multimodal reference roles. URLs must remain accessible while the task is processed; reference counts do not replace format, size or combined-duration requirements.[2]
Use ratio=adaptive for first-frame, first/last-frame, editing and extension tasks. Editing also requires duration=-1 so output follows the reference clip. Duration describes clip length, not processing time. Audio references and generate_audio are separate choices; check both before submission.[2]
Understanding generation costs
Charges depend on video-token usage and the applicable rate. Resolution, actual output and reference video affect usage; video inputs can also have minimum billable usage. A lower reference-video rate does not necessarily produce a lower total charge. Compare input conditions, specifications and usage together.[3]
Prices and published examples on this page come from the current catalog. Each example has stated specifications and usage assumptions; it is not a quote for your request or a universal per-second rate. Each deliberate new generation is billed separately. Reconcile recorded charges using task and request records.[5][3]
API access
Using the model through XiuAI
Reuse your XiuAI account, balance and an authorized API key. Choose an available protocol listed on this page and copy its matching example with the exact model ID. Ark and unified video API fields differ. Supply accessible reference URLs and save the generated results to your own storage.[5][6]
Save the task ID returned on creation and query that task until it finishes. Continue from task records after leaving the playground, then reconcile charges in request records. If acceptance is unclear, check existing tasks before creating another one. A temporary query error does not establish generation failure.[4][6]
Replace the example first-frame URL with an accessible URL. Query the returned task ID for the result.
POST /api/v3/contents/generations/taskscurl 'https://router-api.xiu.ai/api/v3/contents/generations/tasks' \
-H "Authorization: Bearer $XIUAI_API_KEY" \
-H 'Content-Type: application/json' \
--data-raw '{
"model": "doubao-seedance-2-5-260628",
"content": [
{
"type": "text",
"text": "Slowly push the camera toward a product on a desk, keeping its packaging text sharp."
},
{
"type": "image_url",
"image_url": {
"url": "https://assets.example.com/product.jpg"
},
"role": "first_frame"
}
],
"resolution": "720p",
"duration": 5,
"ratio": "adaptive",
"generate_audio": true
}'Query the task
curl https://router-api.xiu.ai/api/v3/contents/generations/tasks/$TASK_ID \
-H "Authorization: Bearer $XIUAI_API_KEY"Read the API documentationView task and charge recordsTasks and data handling
- Asynchronous result
- Query by task ID after submission. Closing the page does not cancel generation.
- Result download
- Use the expiry returned with the task and save the result to your own storage before it expires.
- Input data
- Use accessible URLs you manage. XiuAI does not host input assets or manage creative workflows.
Frequently asked questions
Can I generate a video without an image?
Yes. Choose text-to-video with a prompt. Use first-frame mode for a starting image, first/last-frame mode for starting and ending images, or reference mode when several assets guide the result.[2]
Which output resolutions are supported? Do input and output limits match?
Current Volcengine output options for 2.5 are 480p, 720p, 1080p. Input-video resolution and output resolution are separate limits; accepting a higher-resolution reference does not imply that output option is available.[2]
Are all combinations within the reference counts valid?
No. Counts are only one constraint. Check formats, sizes, roles and combined duration as well. First/last-frame roles cannot be mixed with multimodal reference roles; check audio references against the current request requirements.[2]
Why does editing use automatic ratio and duration?
Editing uses ratio=adaptive and duration=-1 so output follows the selected reference clip. The prompt must clearly describe an edit, and the source must meet the duration requirements. A mismatch between intent and parameters can still cause failure.[2]
Why can a lower reference-video rate still mean a higher total charge?
Charges depend on video-token usage and the applicable rate. Resolution, actual output and reference video affect usage; video inputs can also have minimum billable usage. A lower reference-video rate does not necessarily produce a lower total charge. Compare input conditions, specifications and usage together.[3]
Why does a successful create request return a task ID?
Creation confirms acceptance, not completed generation. Query the same task and retrieve its result after success. Queue and generation times vary with specifications and service load; clip length is not processing time.[4]
Should I create another task after a timeout?
Save the task ID returned on creation and query that task until it finishes. Continue from task records after leaving the playground, then reconcile charges in request records. If acceptance is unclear, check existing tasks before creating another one. A temporary query error does not establish generation failure.[4][6]
Will XiuAI keep the generated video for me?
XiuAI does not provide asset or output hosting. Download results to your own storage. Result-link availability and task-record retention are separate; an unavailable link does not mean its record was deleted. See Data Privacy for asynchronous task data handling.[6][4]
Related models
Seedance 2.5 Flash · VolcengineSources and updates
Specifications reference provider documentation; prices and availability come from the current XiuAI catalog. Specifications checked: 2026-11-15
Content reviewed: 2026-11-16
[1] ByteDance model introduction[2] Volcengine request parameters[3] Volcengine billing guide[4] Volcengine task retrieval[5] Current XiuAI model and pricing data[6] XiuAI Data Privacy