Skip to main content
Seedance is a flagship multimodal video family on Venice for text-, image-, and reference-driven video generation. Seedance 2.0 (plus Fast) and Seedance 2.5 share the same R2V prompt-routing model: a single reference-to-video endpoint handles four distinct workflows (Reference, Edit, Extend, Stitch) — the workflow is inferred from the shape of your prompt. This guide covers variants, the four workflows, public API media policy, family-specific multimodal limits, output bitrate, Seedance 2.5 draft mode, pricing, and curl examples.
Person-bearing media is not supported on the public Seedance API. Public *-basic models do not use consent attestation (consents.seedance / needs_consent). Such inputs may be rejected upstream as a content-policy or provider error. Use the Venice app or Studio for the full Seedance feature set.

Variants

All variants are async. Submit via POST /api/v1/video/queue, then poll POST /api/v1/video/retrieve until the response body is video/mp4. See Video Generation for the general queue flow. Pass resolution as one of: 480p, 720p, 1080p, or 4k (lowercase). Seedance 2.0 (non-Fast) and 2.5 accept all four; 2.0 Fast accepts 480p / 720p only. Discover live model IDs and capabilities with GET /models?type=video — do not hardcode availability.

Output bitrate

You can set the output encode bitrate on every public Seedance 2.0 (including Fast) and 2.5 model. Pass bitrate_mode on /video/queue yourself — Venice does not pick this automatically. This is a queue-only encoding option: it does not change price or token cost, and /video/quote does not accept bitrate_mode. Other video families (Wan, Kling, LTX, and so on) do not support this field.

The “one model, four workflows” model

The reference-to-video variants (seedance-2-0-reference-to-video-basic, its Fast sibling, and seedance-2-5-reference-to-video-basic) use the same prompt-routing pattern. The model infers the task from the prompt prefix and the shape of your inputs. There is no task or workflow field — the prompt syntax is the routing. The prompt syntax is canonical and case-sensitive: angle brackets, capital first letter, single space before the number — <Video 1>, <Image 1>, <Audio 1>.

Workflow patterns

Reference workflow

Use the uploaded reference files as donors — subject, scene, motion, style, vocal timbre — to generate a brand-new video. Canonical prompt patterns:
Examples:
  • Refer to <Subject 1> in <Image 1> to generate a 5-second clip of the same character riding a horse through snow.
  • Refer to the camera scene in <Video 1> to generate a similar establishing shot of a futuristic city at dawn.
  • Refer to <Subject 1> in <Image 1> and use the timbre in <Audio 1> for the narrator describing the scene. (audio donors must be paired with at least one image or video reference — audio alone is rejected)

Edit workflow

Modify a single input video. Anything not explicitly named in the prompt is preserved. Use this when you want a localized change (subject swap, weather/color change, element add/remove) rather than a wholly new video. Canonical prompt pattern:
Sub-patterns for finer control:
Examples:
  • Strictly edit <Video 1>, changing its weather from sunny to a heavy rainstorm.
  • Add snacks such as fried chicken and pizza to the countertop in <Video 1>.
  • Remove the red car from <Video 1>, keeping the rest of the video content unchanged.
  • Replace the perfume featured in <Video 1> with the face cream from <Image 1>, with all original motions and camera work preserved.
The last example combines Edit with an image reference — perfectly legal, the model uses <Image 1> as a visual donor for the replacement.

Source-matched aspect ratio and duration

For Seedance reference-to-video edit / extend, you can ask the output to follow the source clip instead of picking a fixed ratio or length: Requirements:
  • Queue: any source-matched value requires reference_video_urls.
  • Quote: any source-matched value requires reference_video_total_duration. Source-matched duration bills ceil(reference_video_total_duration) seconds.
  • Ratio and duration are independent — you can match one without the other.
  • For extend, prefer a fixed duration (how long to generate) and optionally aspect_ratio: "adaptive". Source-matched duration is for edit-style “same length as source” jobs.

omni_reference_task_type (Seedance 2.5 R2V)

BytePlus classifies Seedance 2.5 all-modal reference jobs as reference, edit, or extend from the prompt. You can optionally pre-guide that classification on /video/queue with: Aliases editing→edit and extension→extend are accepted. When omitted, Venice infers the value from the prompt if reference_video_urls are present. The prompt must still match the chosen type — a conflict can still fail asynchronously.

Extend workflow

Continue a single clip forward or backward in time. By default Seedance returns only the new content — not the original input concatenated with the extension. This is by design, for transition continuity; if you want the input clip preserved alongside the extension, say so explicitly:
Transition handling: the model automatically extracts the transition frames for seamless blending, and the original segments of the input video are not re-generated. Examples:
  • Extend <Video 1>, generate a dramatic chase scene through narrow alleys at dusk.
  • Extend <Video 1> backward, the same character walking toward the camera before the original shot begins.
  • Extend <Video 1>, start with <Video 1>, then the camera pulls back to reveal a vast landscape.

Stitch workflow (Track Completion)

Connect input clips with AI-generated transitions. Respect the family-specific combined-duration and clip-count caps in Multimodal input limits (Seedance 2.0: ≤3 clips / ≤15 s combined; Seedance 2.5: higher video caps). Canonical prompt pattern:
Examples:
  • <Video 1> + a smooth seamless cut + followed by <Video 2>
  • <Video 1>. The moment a leaf falls to the ground, it sets off a special effect of golden particles. A gust of wind blows by, leading into <Video 2>.
  • <Video 1> + a wisp of smoke transforms into a flock of birds + followed by <Video 2> + a slow dolly-in + followed by <Video 3>
The model auto-trims connecting segments at the join points for continuity.

Universal prompt formula

Across all four workflows, the recommended authoring formula is:
  • Subject + Motion: the logical foundation — define “Who” is performing “What action”
  • Environment + Aesthetics: spatial background, lighting, visual style
  • Camera: explicit shot type or movement
  • Audio: ambient sound effects or vocal direction for immersive output
Layering this on top of a workflow prefix (e.g., Strictly edit <Video 1>, changing its <subject + motion + environment + ...>) produces the highest-quality outputs.

Multimodal input limits

Values below are what the Venice API accepts. Requests outside these ranges are rejected at the schema layer with a 400 before reaching inference. Seedance 2.0 and Seedance 2.5 use different caps. Always check the column for the model family you are calling.

Shared media floors

Family comparison

Reference audio is supported on the R2V variants only. Each entry is forwarded to the model as a role: "reference_audio" content item that the prompt addresses as <Audio 1>, <Audio 2>, … — the model uses each clip for vocal timbre, sound effects, or background music depending on how the prompt frames it. The legacy singular audio_url field maps to the same content shape and is now equivalent to passing a one-element reference_audio_urls.
reference_audio_urls cannot be the only reference input. The model requires at least one image or video reference alongside any audio donor. Pair reference_audio_urls with reference_image_urls or reference_video_urls — audio-only submissions are rejected, and so are submissions that pair audio with only image_url or video_url.

Request size

The queue endpoint accepts JSON bodies up to 35 MB. Inline data URLs for large videos can push past this — for multi-clip Stitch in particular, prefer URLs over inline base64.

Draft mode (Seedance 2.5)

Seedance 2.5 models that support draft mode, such as seedance-2-5-text-to-video-basic, can render a 480p draft and later upgrade that draft to 1080p or 4k. The draft and the upgrade are separate requests and are billed separately. Call POST /api/v1/video/quote for each price before you queue it. Sending draft to any other model returns 400 “This model does not support draft”. Sending draft_task_id to any other model returns 400 “This model does not support draft_task_id”.

Create a draft

Set draft: true on POST /api/v1/video/queue or POST /api/v1/video/quote. resolution defaults to 480p. Any other value returns 400 “Draft mode requires resolution 480p”. A 5 second 16:9 draft is 854×480. Quote the draft:
Queue the draft:
The queue response, and a retrieve response while status is PROCESSING, include these fields. draft_task_id is the draft’s queue_id. Retrieve the draft while it is running:
The response also includes average_execution_time and execution_duration, as any other retrieve does. draft_expires_at is 7 days after the draft was created. When the body switches to video/mp4, the same draft values are response headers: x-venice-is-draft, x-venice-draft-task-id, x-venice-draft-expires-at, and x-venice-can-upgrade-to-hd. On that completed response, x-venice-can-upgrade-to-hd is true until draft_expires_at. A draft’s media is kept until the HD upgrade window ends, even when delete_media_on_completion is true.

Upgrade the draft to 1080p or 4k

Send a new POST /api/v1/video/queue with draft_task_id set to the draft’s queue_id. The draft must be your own, on the same model, and inside the 7-day window. resolution defaults to 1080p; send "resolution": "4k" for a 4k upgrade. Any other value returns 400 “Draft HD upgrade requires resolution 1080p or 4k”. duration and aspect_ratio are optional on the upgrade. Seed, duration, aspect ratio, and audio follow the draft. A duration or aspect_ratio that does not match the draft is ignored, and the draft’s values are used. Media inputs on the upgrade request are ignored. prompt is still required. The same draft can be upgraded more than once inside the window. Do not send draft and draft_task_id together. Quote the upgrade. This quote requires authentication. duration is required on the quote request, but the price uses the draft’s duration. Add "resolution": "4k" to quote a 4k upgrade:
Queue the upgrade:
Queue the same draft as a 4k upgrade:
A rejected draft or upgrade request is not charged.

Pricing

Call POST /api/v1/video/quote to get a quote for a given request shape before submitting it to /video/queue. The quote endpoint is the only authoritative source; pricing details may change and shouldn’t be cached or duplicated client-side. When reference video(s) are part of the request, also pass reference_video_total_duration (the sum of all reference clip durations in seconds) so the quote matches what /video/queue will charge:
Source-matched Seedance 2.5 edit quote (bills from the source length):

Complete examples

All examples assume VENICE_API_KEY is set in the environment.

Text-to-video

Seedance 2.0 text-to-video (4K)

Seedance 2.5 text-to-video (longer duration)

Image-to-video (first frame)

Seedance I2V models (seedance-2-0-image-to-video-basic, its Fast variant, and seedance-2-5-image-to-video-basic) do not accept aspect_ratio — the output aspect ratio is auto-derived from the input image’s dimensions. Passing the field returns a 400 with “This model does not support aspect_ratio”. Use the T2V or R2V variants if you need explicit aspect-ratio control.

Reference workflow — subject donor

Seedance 2.5 reference workflow — multi-image

Reference workflow — subject + audio donor

Edit workflow

Seedance 2.5 edit — source-matched duration and aspect

omni_reference_task_type: "edit" pre-guides BytePlus classification. duration: "auto" (or "-1") and aspect_ratio: "adaptive" (or "auto") make the output follow the source clip. See Source-matched aspect ratio and duration and omni_reference_task_type.

Edit workflow with image grounding

Extend forward

Stitch (3 clips)

Polling for completion

After every queue submission, save the returned queue_id and poll /video/retrieve until the response body is video/mp4:
The response is JSON ({ "status": "PROCESSING", ... }) while the job runs, and the body switches to video/mp4 bytes once it completes. status is one of PROCESSING or COMPLETED — there is no failed status. A failed job instead returns a non-200 response carrying the recorded error. See Video Generation for the full polling pattern.

Troubleshooting

At least one reference is required for this model

Reference-to-video submissions must include at least one of reference_image_urls, reference_video_urls, or image_references. Pure text-only generation isn’t a valid R2V workflow — use a text-to-video model ID instead. reference_audio_urls alone is not sufficient (see the Audio section above). Note that video_references is not accepted on Seedance — it is specific to models that support per-reference voice audio, and passing it returns “This model does not support per-reference video entities”.

Too many reference videos / images

Seedance 2.0 caps R2V at 9 images and 3 videos. Seedance 2.5 raises those caps to 30 images and 10 videos. If you exceed the family limit, trim inputs or stitch offline first.

Duration / aggregate duration errors

  • 2.0: per-clip reference video/audio [2, 15] s; combined video/audio ≤ 15 s; output 4–15 s.
  • 2.5: per-clip reference video/audio [2, 30] s; combined video/audio ≤ 30 s; output 4–30 s.
  • Source-matched duration (-1 / auto on Seedance 2.5): source clip must be 4–30 s, and reference_video_urls (queue) or reference_video_total_duration (quote) is required.
Trim clips client-side before submission.

Prompt routes to the wrong workflow

Workflow is inferred from prompt syntax. Common misroutings:
  • Wanting to Extend but writing Refer to ... → model treats your video as a donor, not a canvas to continue
  • Wanting to Stitch but writing Refer to ... → model picks one as the donor, ignores the others
  • Wanting to Edit but writing Generate a video based on <Video 1> → ambiguous; model may default to Reference
Use the canonical prefixes exactly as written: Strictly edit <Video 1>, ..., Extend <Video 1>, ..., <Video 1> + ... + followed by <Video 2>.

Unsupported person-bearing media

Public Seedance API models do not run a consent attestation flow (consents.seedance / needs_consent). Media with detectable persons may fail with a content-policy or provider error. Use the Venice app or Studio instead.

Quote doesn’t match the queued amount

If you included a reference video but didn’t pass reference_video_total_duration to /video/quote, the quote and the queued amount may differ. Always pass reference_video_total_duration (sum of all reference clip durations, in seconds) when reference videos are present.

Draft mode

These requests return 400 and are not charged. See Draft mode (Seedance 2.5).
  • draft on a model that does not support it: “This model does not support draft”
  • draft_task_id on a model that does not support it: “This model does not support draft_task_id”
  • draft and draft_task_id together: “draft cannot be combined with draft_task_id”
  • draft_task_id that is not a UUID: “draft_task_id must be a valid queue_id”
  • Unknown or unavailable id: “Draft queue id was not found or is no longer available.”
  • A queue id that is not a Seedance 2.5 draft: “Only a Seedance 2.5 draft can be upgraded to HD.”
  • A different model than the draft: “Draft HD upgrade must use the same model as the draft.”
  • Draft still generating: “Draft has not finished generating. Wait until it completes, then upgrade to HD.”
  • Draft not yet accepted by the provider: “Draft is still submitting. Wait until the draft task is accepted before upgrading.”
  • Older than 7 days: “Draft expired. Draft task IDs are valid for 7 days from creation.”
  • Draft resolution other than 480p: “Draft mode requires resolution 480p”
  • Upgrade resolution other than 1080p or 4k: “Draft HD upgrade requires resolution 1080p or 4k”

References