Authors and ModelsByteDance

Seedance 2.5

Seedance 2.5
ByteDance · Video Generation
POST /v1/videos/generations

Long-form video model for coherent clips up to 30 seconds, with up to 50 reference images, videos, and audio clips, native audio, editing, and extension.

At a glance

FieldValue
Model idseedance-2-5
Model release date2026-08-07
Input modalitiesText, Image, Video, Audio
Output modalitiesVideo
Context window-
Weight precision-
RegionMalaysia
Featurestext_to_video, image_to_video, reference_to_video, video_editing, multimodal, audio_sync, camera_control, character_consistency, draft_mode
Native inferenceNo
NewYes
Supported endpointsPOST /v1/videos/generations
Alternate model idsbytedance/seedance-2.5, seedance-2.5

Pricing

ChargeSpecRate
T2V/I2V 480Pper second$0.206
T2V/I2V 720Pper second$0.462
T2V/I2V 1080Pper second$1.137
Video Input 480Pper second of input plus output$0.123
Video Input 720Pper second of input plus output$0.276
Video Input 1080Pper second of input plus output$0.680

Example request

curl https://api.empiriolabs.ai/v1/videos/generations \
  -H 'Authorization: Bearer $EMPIRIOLABS_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{"model": "seedance-2-5", "prompt": "sunrise over the ocean", "duration": 6}'

Parameters

ParameterTypeRequiredDefaultDescription
promptstringyes-Scene description. Address a specific attachment inline as @Image1, @Video1, or @Audio1 to say which reference a shot should follow.
modeenumno"auto"Selects how your inputs are used. auto detects it from what you attach. t2v is text to video. i2v_first animates a first frame. i2v_both animates between a first and a last frame. reference builds from up to 30 images, 10 videos and 10 audio clips. To edit or continue an attached clip, say so in the prompt. · Allowed: auto, t2v, i2v_first, i2v_both, reference
resolutionenumno"720p"Output resolution. Draft mode always renders at 480p, and a draft’s final render is always 1080p, so this is ignored on both. · Allowed: 480p, 720p, 1080p
aspect_ratioenumno"adaptive"Output aspect ratio. adaptive lets the model choose. When you supply a first frame, the output follows that image’s aspect ratio, so aspect_ratio must stay adaptive. Reference jobs, including ones with a reference video, accept any ratio. · Allowed: adaptive, 16:9, 9:16, 1:1, 4:3, 3:4, 21:9
custom_durationbooleannotrueIf false, the model decides clip length.
durationnumberno5Clip length in seconds. Only used when custom_duration=true. Video editing always keeps the source clip length. · Range: 4 – 30
generate_audiobooleannotrueGenerate native audio with the video.
output_formatenumno"mp4"mp4 plays everywhere. mov carries higher color precision for color grading, keying, and compositing, and some players cannot open it. · Allowed: mp4, mov
draftbooleannofalseRender a 480p draft to try the shot cheaply, then render the one you like at native 1080p from the same prompt, references and seed.
draft_task_idstringno-The ID of a draft to render at native 1080p. The draft keeps its prompt, references, length, aspect ratio, seed and audio. Drafts stay available for 7 days, and each draft’s result shows its ID.
return_last_framebooleannofalseAlso return the final frame as a PNG, ready to use as the first frame of the next clip.
imagestringno-Reference / first-frame image URL.
image_endstringno-End-frame image URL for i2v_both.
videostringno-Reference video URL for reference / edit / extend.
negative_promptstringno""What to avoid.

Notes

Next-generation Seedance built for long-form storytelling: one request can produce a coherent clip of up to 30 seconds, draw on up to 50 reference assets, and edit or extend existing footage.

Modes

  • Text to video, first frame, or first and last frame.
  • Multimodal reference to video from up to 30 images, 10 videos, and 10 audio clips. Audio-only input works on its own.
  • To edit an attached clip or continue it, describe that in your prompt. The model reads the intent from your wording.

Output

  • 480p, 720p, and 1080p at 24fps, 4 to 30 seconds, with native synchronized audio.
  • MP4 by default. MOV carries higher color precision for grading and compositing, and some players cannot open it.
  • When you supply a first frame, the output follows that image’s aspect ratio, so aspect_ratio stays adaptive. Reference jobs, including ones with a reference video, accept any ratio.

Draft mode

  • Set draft to true to render a 480p draft while you are still trying shots. Drafts render at 480p only.
  • The response returns the draft’s id as usage.draft_task_id. Send that back as draft_task_id to render the same shot at 1080p.
  • The final render takes the prompt, reference assets, duration, aspect ratio, seed and audio setting from the draft, so anything you send alongside draft_task_id is ignored. A draft stays promotable for 7 days.

Input limits

  • Images: jpeg, png, webp, bmp, tiff, gif, heic, or heif, 300 to 6000 px per side, under 30 MB each.
  • Videos: MP4 or MOV, 2 to 30 seconds each and 30 seconds in total, under 200 MB each.
  • Audio: WAV or MP3, 2 to 30 seconds each and 30 seconds in total, under 15 MB each.
  • Reference uploads that contain real human faces are rejected by the model.

Billing

  • Without a reference video, you are billed per second of generated video at the rate for the output resolution.
  • With a reference video, the model prices the reference clip’s seconds alongside the generated seconds, at the lower video-input rate. Billed seconds are therefore reference duration plus output duration, so a request that includes a reference video costs more in total than the same output generated from text alone.
  • Draft mode is not a discounted rate. A draft bills as a normal 480p generation and its final bills as a normal 1080p generation, so the saving comes from trying shots at 480p and rendering only your pick at 1080p.
  • Worked example at 480p: a 4 second clip from a text prompt bills 4 seconds. A 4 second clip generated from a 5 second reference video bills 9 seconds at the video-input rate.

Uploaded media preprocessing

  • Video inputs are capped to 30 seconds.
  • Uploaded video inputs are normalized to provider-compatible MP4 when needed. Set preprocess_video to false to send a clip through untouched.

Machine-readable schema: GET https://api.empiriolabs.ai/v1/models/seedance-2-5.