Skip to main content

Uncensored — This model declares no content filter. SpicyAPI forwards the request without adding a platform content scan; legal and consent duties remain with the customer.Browse every model at this filtering level
bytedance/seedance-2.0/reference-to-video

ByteDance's Seedance 2.0 reference-to-video builds one shot from your own material: up to 9 images, 3 video clips and 3 audio tracks, with no first frame. Output is 4 to 15 seconds at 480p, 720p, 1080p or 4K, with native audio.

From $0.043/s
Run model

Workbench

Input

Reference to video

Reordering updates prompt references. Removing media also removes its references; review your prompt.

Required, and here it also does the casting. References are cited in the order you send them, so name which image is which character, what each clip contributes and what the audio is for. Briefs are not reviewed on our side; where Seedance's own layer refuses the wording, a first frame on image-to-video is the route that still gets there.
Provide at least one of theseRequirement met
2/9

Drop files or media links, paste, or browse · Up to 9 files · Max 10 MB · JPG · PNG · WEBP · GIF

  • @Image1
  • @Image2
2 / 9 files
Up to 9 reference images, used to hold a character, a style or a set. Two or three angles of the same character usually work better than a stack of near-identical frames. This endpoint has no first- or last-frame parameter: reference mode and frame mode are mutually exclusive, so switch to image-to-video when the opening frame must be exact. Seedance's own input check may turn away a reference photo or clip that shows a real, identifiable person's face. Each side must be between 300 and 6000 pixels; a smaller plate is rejected by the input check before you are charged. at most 9 items
affects price0/3

Drop a video here, or browse · Up to 3 files · Max 90 MB · MP4 · WEBM

0 / 3 files
Up to 3 reference clips, each 2 to 15 seconds long and no more than 15 seconds combined, supplying motion, pacing or camera language. The length of a reference clip counts toward the billed quantity — not as a tier switch but as a linear multiplier of the same kind as `duration_seconds` — so trim to the seconds you actually want referenced. at most 3 items
0/3

Drop an audio file here, or browse · Up to 3 files · Max 90 MB · MP3 · WAV

0 / 3 files
Up to 3 reference audio files, each 2 to 15 seconds long and no more than 15 seconds combined, giving the model a sonic anchor — delivery, tempo, a musical phrase — that the motion and cut points align to. Distinct from `generate_audio`: this is an input, while whether the finished clip carries a soundtrack is still decided by `generate_audio`. at most 3 items
Defaults to `adaptive`: the model finds framing that accommodates every referenced subject, which matters most when the reference assets do not share a shape. Lock a fixed ratio only for a hard delivery requirement — `9:16` for a vertical feed, for example — and check that no referenced person or product ends up cropped. No effect on billing.
affects price
480p, 720p, 1080p or 4k, default 720p, and it feeds the billed quantity directly. It sets the output tier only and will not sharpen low-resolution references. Test whether the references combine at a low tier, then raise it for the final take. No upscaling or enhancement tiers exist here.
Durationaffects price
4 to 15 seconds, default 5, or -1 to let the model pick the length (the estimate assumes the 15-second maximum; the charge follows the delivered length). Fifteen seconds is the ceiling of this generation; for a longer single shot, move to the 2.5 endpoint. Note that this and the length of your reference clips are two independent contributions to the billed quantity, so maxing both is where a single request gets expensive. no lower than 4 · no higher than 15
Generate audioaffects price
On by default. Produces a native soundtrack synchronized to the picture. Even with `reference_audio_urls` supplied, turning it off means no soundtrack is rendered — reference audio shapes pacing and performance, it is not the output track. Switch it off when the edit will be scored in post.
Safety checkerOff

SpicyAPI does not run a safety checker on this model. Your prompt and media go to the model as written, and the model itself decides what comes back.

Checking your account before enabling submission and pricing.

Output

Ready for a new run
Example result
Extend
Video editing
Video upscaling

Related models and next tasks

Missing the model you need?

Tell us what to add and where you would use it. Requests go to the people who decide what gets onboarded next.

Give several approved assets separate jobs in one scene

Seedance 2.0's reference-to-video endpoint is the one that takes several kinds of material in a single call: up to 9 stills in reference_image_urls, 3 clips in reference_video_urls and 3 tracks in reference_audio_urls. prompt is required, and so is at least one of reference_image_urls or reference_video_urls, since audio alone is not enough to submit. The work sits in the prompt, which has to say what each supplied asset is responsible for.

Open the playground
Give several approved assets separate jobs in one scene
Required input
prompt, plus reference_image_urls or reference_video_urls
Reference caps
up to 9 images, 3 video clips, 3 audio tracks
Reference clip length
2-15s per clip, 15s combined per type
Duration
4-15 seconds, or -1 for smart duration
Resolution
480p / 720p / 1080p / 4k, default 720p
Not present here
No image_url / last_image_url, those belong to image-to-video
Multiple arrays, multiple assets
Each reference type accepts more than one URLreference_image_urls and reference_video_urls are arrays, so one generation can draw on several images and several clips at once. Neither of the other two endpoints in this family has an array field anywhere in its schema.
No assigned roles
The schema doesn't label what each array controlsNothing in the input forces one image to mean identity and another to mean setting. That division of labor has to be written into the prompt itself, referencing each supplied asset explicitly.
The only multi-asset endpoint here
text-to-video and image-to-video have no reference arraysThis is the endpoint built for combining several approved sources into one shot. image-to-video works from stills in two fixed roles, an opening frame and an optional closing one; text-to-video works from words alone. Neither can be handed a set of references to reconcile.

What the Seedance 2.0 API does well

Three reference channels in one call

Three reference channels in one call

Image, video and audio references travel together in one request, so a scene can take its look from one asset and its pacing from another without compositing them outside the API first. ByteDance describes Seedance 2.0 as accepting up to 9 images, 3 video clips and 3 audio clips alongside the instruction text, which is what this endpoint exposes.

Works with a partial set

Works with a partial set

reference_audio_urls is genuinely optional, and at least one of reference_image_urls or reference_video_urls is required by the schema's anyOf constraint. Supply exactly what the scene actually needs and skip the rest entirely rather than padding a request.

  • Same 4K/15-second ceiling as the rest of the family

    resolution and duration_seconds carry the same range as text-to-video and image-to-video, up to 4k and 15 seconds. This endpoint isn't a reduced-capability variant just because it accepts more kinds of input material than the other two.

  • adaptive applies here too

    aspect_ratio defaults to adaptive here as well, and on this endpoint it looks for framing that holds every referenced subject, which matters most when the assets you attach arrive in different shapes. Lock one of the six fixed ratios when a delivery format is already set, then check that no referenced person or product ends up cropped.

  • Reference audio and generate_audio are separate controls

    reference_audio_urls shapes the pacing and delivery the model follows while generating; generate_audio, on by default, decides whether the finished clip carries a rendered soundtrack at all. Supplying reference audio with generate_audio switched off returns a silent clip that still follows the referenced timing.

  • A natural fit for recombination

    Assets approved in earlier passes, such as a character plate or a location clip from an earlier shoot, can be reused directly as references instead of regenerating them from scratch for every new shot in a series.

Who Seedance 2.0 is for

  • Split identity from movement across two sources

    Character plus separate motion reference

    Split identity from movement across two sources

    One reference image can anchor appearance while a reference video demonstrates the motion the prompt should follow, letting the two contributions be approved, revised and swapped independently of each other across successive iterations of the same shot.

  • Let audio and image do different jobs

    A voice or sound reference paired with a visual anchor

    Let audio and image do different jobs

    reference_audio_urls can carry tone or pacing while reference_image_urls fixes what's on screen. Write both roles into the prompt explicitly, naming each asset by its position in the arrays, so the model knows which one governs which detail.

  • Reuse what already passed review instead of regenerating it

    Recombining approved assets from earlier passes

    Reuse what already passed review instead of regenerating it

    Feeding back a previously approved image or clip as a reference keeps continuity without redoing work that already cleared review, which is especially useful across a longer production with a recurring cast or recurring setting.

  • Lock aspect_ratio when the references don't share a shape

    A vertical or fixed-ratio delivery with mixed references

    Lock aspect_ratio when the references don't share a shape

    adaptive is the default, but when a delivery format is fixed and the references come in different shapes, choosing a fixed ratio ahead of time avoids leaving that mismatch for the model to resolve on its own.

How to call the Seedance 2.0 API

  1. State each asset's job before listing it

    Name what each supplied reference should control: identity, motion, sound, atmosphere, before the model has to guess from array order alone, since the schema doesn't assign roles automatically.

  2. Leave out references that decide nothing

    Each extra asset is one more thing the model has to reconcile, so attach the references that change the outcome and drop the ones that only repeat what another asset already shows.

  3. Start lean, add one reference at a time

    Begin with the single most important asset, confirm it holds up in the output, then add a second reference only if the scene genuinely needs it.

  4. Say which reference wins a shared detail

    Two images can disagree about hair, wardrobe or lighting, and nothing in the schema ranks array entries against each other. Name the one that governs in the prompt, using the image 1 and image 2 labels this endpoint applies to the assets you send.

Seedance 2.0 worked examples

  • One character reference, one ambience reference

    reference_image_urls carries a single approved character plate; reference_audio_urls carries a short room-tone sample the scene should match in pacing and mood.

    Keep the person's appearance and outfit exactly as shown in the reference image; match the quiet, echoing ambience of the reference audio while they walk slowly through an empty hallway

  • Two image references for shot continuity

    Two reference_image_urls entries, one for the character and one for the location, combined into a single new shot that places both in the same frame.

    Place the character from the first reference image into the location shown in the second reference image, walking toward camera at a relaxed pace, natural daylight matching the location reference

  • A motion reference paired with a character still

    One entry in reference_video_urls carries the camera move, one entry in reference_image_urls carries the person, and the prompt states which does what. The reference clip is trimmed first, because its length is billed alongside the output seconds.

    Follow the camera move of the reference video exactly, with the person from the reference image walking through a covered market at dusk, handheld feel, warm string lights overhead

Seedance 2.0 API questions

This endpoint's flexibility is also its cost: nothing in the schema assigns meaning to a reference array. The prompt is where identity, motion, sound and atmosphere get tied to sources, so treat it as a casting sheet rather than an afterthought, naming assets the way the endpoint numbers them, image 1, video 1, audio 1. Trim reference clips before uploading, since their seconds are billed alongside the output, and add a second reference only once the first one has proved it holds.

All Seedance models10

Related comparisons

Seedance 2.0 vs Seedance 2.5