Seedance 2.5
ByteDance Seedance 2.5. Generates video from a prompt alone, or from reference material: up to 5 reference images and videos in any mix (5 images, or 3 clips and 2 stills, or 5 clips), plus up to 10 reference audio tracks. Native audio generation, so the clip comes back with sound. Up to 30 seconds of output and up to 30 seconds of combined reference footage, at 480p or 720p (there is no 1080p tier on this model). 🔴 TASK-CLASSIFICATION TRAP, read this before writing a prompt: the model picks its task from the PROMPT TEXT. A prompt containing editing or extension trigger words (\"edit\", \"add\", \"remove\", \"replace\", \"extend\", \"continue\") is treated as video-editing or video-extension rather than generation, and that mode REQUIRES Aspect Ratio \"adaptive\" (and, for editing, duration -1). Sending a normal aspect ratio or duration with such a prompt fails after submission with InvalidParameter.TaskTypeConstraint. Wireflow documents this and does not enforce it: phrase generation prompts without those verbs, or set Aspect Ratio to adaptive.
Seedance 2.5
Node type: video:seedance_2_5
Category: video
Description
ByteDance Seedance 2.5. Generates video from a prompt alone, or from reference material: up to 5 reference images and videos in any mix (5 images, or 3 clips and 2 stills, or 5 clips), plus up to 10 reference audio tracks. Native audio generation, so the clip comes back with sound. Up to 30 seconds of output and up to 30 seconds of combined reference footage, at 480p or 720p (there is no 1080p tier on this model). 🔴 TASK-CLASSIFICATION TRAP, read this before writing a prompt: the model picks its task from the PROMPT TEXT. A prompt containing editing or extension trigger words ("edit", "add", "remove", "replace", "extend", "continue") is treated as video-editing or video-extension rather than generation, and that mode REQUIRES Aspect Ratio "adaptive" (and, for editing, duration -1). Sending a normal aspect ratio or duration with such a prompt fails after submission with InvalidParameter.TaskTypeConstraint. Wireflow documents this and does not enforce it: phrase generation prompts without those verbs, or set Aspect Ratio to adaptive.
Pricing
- Cost: ~26 credits per second
- Per-second rate: The figure is the rate at a 5-second reference run. Our margin is progressive and steepest on the smallest runs, so a clip shorter than that can cost more per second, and a model priced with a flat multiplier is exact at every length. The charge is always computed from the real duration.
- Notes: RETAIL prices (marginMultiplier 1.0 bypasses progressive tiers). Derivation, same h*w*24/1024 token formula as 2.0: 720p 16:9 (1280x720) = 21,600 tokens/s, x $10.7/M = $0.2311/s true cost, x1.13 margin = $0.26/s retail. 480p 16:9 (854x480) = 9,608 tokens/s, x $10.7/M = $0.1028/s true cost, x1.4 margin = $0.1439, rounded up to $0.15/s retail. Cross-check against BytePlus published examples: a 5s 720p clip with no reference video = 21,600 x 5 x $10.7/M = $1.156, matching their quoted $1.16. With-video-input runs bill a cheaper $6.4/M but ALSO bill the reference footage tokens: covered via tokenPricing, pre-charge adds a 30s allowance and settlement trues down to usage.total_tokens (never up). Margins copied from the 2.0 directive and NOT yet confirmed for 2.5. refVideoTokensPerSecond is PER GENERATION RESOLUTION (9,607.5 at 480p, 21,600 at 720p), derived from refDensityTable, not a flat constant: BytePlus bills a reference second at the same w*h*24/1024 density as an output second. CONFIRMED against BytePlus published with-video-input prices for a 5s output across its 4s..30s reference range -- 480p $0.553..$2.152, 720p $1.244..$4.838, 1080p $3.062..$11.907 -- every one of which is reproduced to four decimals by (outSeconds + refSeconds) x density x rate: 480p 30s ref = (5 + 30) x 9607.5 x $6.4/M = $2.1521; 720p = 756,000 x $6.4/M = $4.8384; 1080p = 1,701,000 x $7.0/M = $11.907. Six published numbers, six exact fits, which is why this is a derivation rather than the ceiling-shaped guess it replaced (#1316: the old flat 21,600 made 480p quote MORE than 720p).
Variant family
- Family:
bytedance_seedance - Is primary: no
- All variants:
video:seedance_2_5
Canvas ports
These appear as port handles on the left side of the node.
| ID | Label | Details |
|---|---|---|
prompt |
Prompt | TEXT (required) · multiline |
first_frame_url |
Start Frame | IMAGE — The video opens exactly on this image. Use it when you have already composed the shot (framing, background, camera position) and want it kept. Costs one of the 5 registered materials. Pair it with an End Frame for a start-to-end shot. It cannot be combined with any reference media (Reference Images, Reference Audio, Audio): ARK runs a start/end-frame task OR a reference task, never both (the mix is refused before you are charged). Wiring a Reference Video turns the run into an edit, and this image then becomes a reference too. To add a voice to a Start Frame shot, run Sync Lipsync v3 on the finished clip. |
image1 |
Reference Image 1 | IMAGE — An identity and style guide, not a framing guide. The model may reframe freely, so a carefully composed shot wired here comes back as a DIFFERENT shot. If you want the video to open on this exact image, use Start Frame instead. |
video_url |
Reference Video | VIDEO · list |
audio_url |
Reference Audio | AUDIO · list — A timbre and timing reference, not the soundtrack. Seedance re-voices it and never passes the track through, so the lips will not match it exactly. Cannot be combined with a Start Frame (the mix is refused before you are charged). mp3 or wav only. For an exact voice, run Sync Lipsync v3 on the finished clip. |
image_url |
Reference Image (from Seedance 2.0) | IMAGE — An image arriving from the Seedance 2.0 card. It is used as a reference, not as an opening frame: the model may reframe freely. To make the video open on an exact image, wire the Start Frame port above. |
end_image_url |
End Frame | IMAGE — Optional last frame. Beside a Start Frame the video ends exactly on this image (and, like the Start Frame, it cannot be combined with reference media). With no Start Frame it is treated as one more reference image, so it guides the ending rather than pinning it. |
reference_image_urls |
Reference Images | IMAGE · list — Identity and style guides. The model may reframe freely, so these steer WHO and WHAT is in the shot, not how it is framed; to pin the framing use Start Frame. Start Frame, End Frame and any reference VIDEO all draw on the same budget of 5 registered materials per run. |
audio_urls |
Audio | AUDIO · list — Optional reference audio, up to 10 tracks and 30 seconds combined. Shares the reference-audio slot with Reference Audio; wiring both concatenates them. A timbre and timing reference that Seedance re-voices, never passed through. Cannot be combined with a Start Frame. For an exact voice, run Sync Lipsync v3 on the finished clip. |
Sidebar config
These render as form fields in the right-side config panel when the node is selected.
| ID | Label | Details |
|---|---|---|
ratio |
Aspect Ratio | TEXT · options: 21:9, 16:9, 4:3, 1:1, 3:4, +2 |
duration |
Duration | TEXT · options: auto, 4, 5, 6, 7, +23 |
resolution |
Resolution | TEXT · options: 480p, 720p |
generate_audio |
Generate Audio | BOOLEAN — On: the clip comes back with sound the model generates, guided by any Reference Audio. Off: the clip comes back SILENT. Off never keeps your own track; add that afterwards (Sync Lipsync v3 for a voice). |
watermark |
Watermark | BOOLEAN |
contentFilter |
Content filter | BOOLEAN — On by default. BytePlus screens your prompt, your reference material and the finished clip, so a run can fail after it has been submitted when something gets flagged (copyright is the common one). Turning this off asks BytePlus to skip that safety and IP moderation. It costs 10 percent more per run, and you are responsible for holding the rights to whatever you generate. |
aspect_ratio |
Aspect Ratio override (set by Seedance 2.0) | TEXT · options: 21:9, 16:9, 4:3, 1:1, 3:4, +2 — Leave this empty. It is filled in automatically when the Seedance 2.0 node routes a run into this model, and while it has a value it OVERRIDES the Aspect Ratio field above ("auto" lets the model choose from your reference media). To change the shape of your video, use Aspect Ratio, not this. |
Outputs
| ID | Label | Type |
|---|---|---|
video |
Generated Video | VIDEO |
Start Frame, reference images and the budget
Start Frame is first_frame_url. Wire it when the video must open on an exact image you already composed. image_url is NOT the Start Frame: it is the port the Seedance 2.0 card bridges into, and it is always sent as a plain reference image, so the model may reframe freely. Setting image_url to get a first-frame run gets you a reference-mode run, and you pay for it.
References (Reference Images, Reference Video, and a Start or End Frame) draw on one shared budget of registered materials, images and videos together. The budget is 5 on the Vercel route and up to 8 inside a Trigger task (ark-createasset-budget.ts: ARK_VERCEL_ROUTE_REGISTRATIONS and SUPPORTED_REGISTRATIONS_PER_NODE).
Content filter, and what turning it off costs
BytePlus moderates a Seedance run in three separate places: your prompt, the reference images and video you register, and the FINISHED CLIP. The last one is the surprising one. A run can pass the first two checks, generate, and then fail after submission with a message like "the request failed because the output video may be related to copyright restrictions". Nothing you can see on the canvas predicts it.
Content filter is on by default and that is the behaviour every existing
board already has. Switch it off and Wireflow sends content_filter: false,
which asks BytePlus to skip that safety and IP moderation.
Two things to know before you do:
- It costs 10 percent more. BytePlus bills the run at a higher rate, so the quote on the node, the credit gate before the run, the charge and the final settlement all go up by 10 percent. The number on the card is the number you pay, either way.
- The rights are yours to hold. Skipping the filter does not give you a licence to anything. You are responsible for having the rights to the material you feed in and to whatever comes out.
Leave it on unless a run has actually been refused and you know the footage is yours to use.
Realistic talking-head video
Animating an avatar still? Make a Realistic AI Avatar has the tested recipe: reference mode for head turns, and ship the raw 720p output with no video upscale.
Auto-generated from the Wireflow node registry.