ByteDance Seedance 2.5 API

Overview

Seedance 2.5 is ByteDance’s audio-video generation model for moving beyond short clips toward more complete creative sequences, with stronger long-form storytelling, multimodal reference, and editing.

It can generate up to 30 seconds in one pass and supports two video extensions, while improving shot continuity, subject stability, motion quality, and audio-video consistency.

The model also expands white-model reference, motion reference, timestamp control, green-screen editing, camera editing, and reference editing. On iCreat, output is 480P or 720P for 4–30 seconds.

bytedance/seedance-2-5

Key Features

30-Second Story Generation. Seedance 2.5 can generate up to 30 seconds in one pass, giving setup, progression, transitions, and resolution more room inside one coherent sequence.

Two Video Extensions. Existing results can be extended twice while maintaining subject features, scene context, audio-visual style, and narrative rhythm across a longer connected piece.

Large Multimodal Reference Set. Seedance 2.5 supports up to 30 images, 10 videos, and 10 audio files, giving complex scenes more reference material for control.

White-Model and Motion Reference. White-model references can guide layout, subject paths, camera position, and lighting, while motion references help transfer movement and shot behavior more deliberately.

Timestamp Control. Prompts can assign story beats, actions, viewpoints, camera changes, and pacing to specific time ranges, giving longer videos clearer structure and more predictable timing.

Targeted Video Editing. Seedance 2.5 can change characters, actions, sound, story, camera movement, green-screen scenes, and reference-based content without rebuilding the whole video.

More Natural Visual Results. The model improves materials, skin, eyes, lighting, saturation, motion quality, and uncontrolled subtitles or background music for cleaner, more natural-looking output.

Model Comparison

Seedance 2.5 vs Seedance 2.0 vs Seedance 2.0 Fast

Seedance 2.5 vs MiniMax H3

Inputs

A Text Prompt is required. Seedance 2.5 can also use images, video, and audio as multimodal references.

Parameters

Pricing

Seedance 2.5 uses different unit prices depending on Resolution and whether a Reference Video is included.

Total Cost = Unit Price × (Output Video Duration + Reference Video Duration)

For requests without a Reference Video, Reference Video Duration is zero.

Quick Start

This example uses a Reference Video and Reference Image to replace the subject while keeping the original movement and scene.

curl --fail-with-body --connect-timeout 10 --max-time 60   -X POST https://api.icreat.ai/v1/task/submit/bytedance/seedance-2-5   -H "Authorization: Bearer ${ICREAT_API_KEY}"   -H "Content-Type: application/json"   -d '{
    "content": [
      {
        "type": "text",
        "text": "Replace the cat in this video with the dog in the picture. Keep the original movement, camera behavior, timing, and scene."
      },
      {
        "type": "video_url",
        "video_url": {
          "url": "https://example.com/video.mp4"
        },
        "role": "reference_video"
      },
      {
        "type": "image_url",
        "image_url": {
          "url": "https://example.com/dog.jpg"
        },
        "role": "reference_image"
      }
    ],
    "generate_audio": true,
    "ratio": "16:9",
    "resolution": "720p",
    "duration": 10,
    "watermark": false
  }'

A successful request returns:

{ "task_id": "task-xxx" }

Output

submit → task_id → query-status → get-result

Submit the request, poll query-status with the returned task_id, and call get-result with the same ID after the task reaches SUCCEEDED.

Use Cases

Long Narrative Scenes. Use the 30-second window for scenes that need setup, interaction, camera transitions, and resolution without splitting every story beat into separate short generations.

Reference-Heavy Advertising. Combine product images, character references, motion examples, audio, and scene guidance when a commercial video must keep several creative constraints aligned.

White-Model Previsualization. Use a simple 3D white model to define layout, camera position, subject movement, and spatial relationships before Seedance 2.5 renders the final visual style.

Video Revision and Reframing. Edit selected characters, actions, scenes, sound, or camera movement when most of an existing video already works and only specific elements need changing.

Multi-Character Storytelling. Large image, video, and audio reference sets help control different characters, voices, props, and scenes when a longer sequence contains multiple important subjects.

Limitations

Chinese Prompts are limited to 2,000 characters; English Prompts are limited to 2,000 words.

Reference videos can total no more than 15 seconds, even when several files are uploaded.

Reference audio can total no more than 15 seconds across all uploaded files.

Set need_review=true when reference media contains faces or copyrighted IP, or generation is likely to fail.

Complex physical motion and scenes with very many interacting subjects can still lose stability.

FAQ

When should I choose Seedance 2.5 instead of Seedance 2.0?

Choose 2.5 for 30-second storytelling, larger reference sets, extensions, and editing. Choose 2.0 when 1080P or 4K output matters more than longer generation.

Should I use 480P or 720P on iCreat?

Use 480P for cheaper testing and repeated revisions. Use 720P when the concept is selected and higher visual detail is worth the higher per-second price.

When should I add a Reference Video?

Add one when motion, camera behavior, timing, or an existing scene should guide the result. Remember that Reference Video Duration also contributes to total billing.

Should I use duration=-1 or set a fixed duration?

Use -1 when the model can choose the length. Set a fixed duration when timing, budget, or a planned story structure needs tighter control.

Should I choose Seedance 2.5 or MiniMax H3?

Choose Seedance 2.5 for longer sequences, extension, and targeted editing. Choose H3 when direct 2K output and generalized mixed-reference control are more important.