Seedance 2.5

ByteDance Seedance 2.5. Generates video from a prompt alone, or from reference material: up to 8 reference images and videos in any mix (5 on the older inline route), plus up to 2 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, 720p, 1080p or 4K. 1080p and 4K are made as 10-bit H.265/HEVC and delivered as an H.264 copy so every browser can play them. 4K is the dearest and slowest tier: Seedance makes one 4K clip at a time for all of Wireflow, so a 4K run can wait in a queue first. 🔴 EDIT MODE, read this before writing a prompt: with a Reference Video wired, the model picks its task from the PROMPT TEXT. Editing or extension words (\"edit\", \"add\", \"remove\", \"replace\", \"extend\", \"continue\") make the run an edit or extension of @Video1 rather than a new generation. Wireflow never sends Aspect Ratio on a run with a Reference Video, so that field is ignored (no need to set it to adaptive). Duration does not set the length of an edit either (the model derives it from @Video1), so leave Duration at its default (the default 5 is not sent with a Reference Video): the model refuses a set Duration on an edit. To lip sync an edit, wire the audio of the clip being edited into Reference Audio, turn Generate Audio off, and add this line to the prompt, without which the audio can pull a wide shot into a close-up: \"Match the framing of @Video1 in every frame: same face size and position, same distance from the camera. Never move closer or crop tighter.\" The mouth then follows the timing and big shapes of the track, close but not frame perfect. Before laying the audio back in, check the length of the take against the clip: short clips have come back about 6 percent fast, so retime the take in the Video Editor when it is off. Set Model to Seedance 2.5 Flash for about 35 percent less per second: 480p or 720p only, no draft mode, and at most 5 reference videos. On Flash, a run with a reference clip holds credits for the longest output it could make until it finishes, then settles on what was used.

Try Seedance 2.5 in the playground

Last updated View as MarkdownAgent setup

Seedance 2.5

Node type: video:seedance_2_5
Category: video

Description

Seedance 2.5 is ByteDance's video generation model. It makes a clip with sound from a prompt alone, or from reference material: up to 8 reference images and videos in any mix (5 on the older inline route), plus up to 2 reference audio tracks. Output runs up to 30 seconds, with up to 30 seconds of combined reference footage, at 480p, 720p or 1080p (no 4k on this model); 1080p comes back as 10-bit H.265/HEVC, which some browsers and players cannot play. Its node type is video:seedance_2_5. Call it from code by running a saved workflow that contains it with POST /api/v1/workflows/{id}/run, or let an agent build and run it over the Wireflow MCP server. It costs about 26 credits per second. That is the rate for a 5-second clip; shorter clips can cost a bit more per second, and you are always charged for the real length. Edit mode, read this before writing a prompt: with a Reference Video wired, the model picks its task from the PROMPT TEXT. Editing or extension words ("edit", "add", "remove", "replace", "extend", "continue") make the run an edit or extension of @Video1 rather than a new generation. Wireflow never sends Aspect Ratio on a run with a Reference Video, so that field is ignored (no need to set it to adaptive). Duration does not set the length of an edit either (the model derives it from @Video1), so leave Duration at its default (the default 5 is not sent with a Reference Video): the model refuses a set Duration on an edit. To lip sync an edit, wire the audio of the clip being edited into Reference Audio, turn Generate Audio off, and add this line to the prompt, without which the audio can pull a wide shot into a close-up: "Match the framing of @Video1 in every frame: same face size and position, same distance from the camera. Never move closer or crop tighter." The mouth then follows the timing and big shapes of the track, close but not frame perfect. Before laying the audio back in, check the length of the take against the clip: short clips have come back about 6 percent fast, so retime the take in the Video Editor when it is off. To see a Reference Video swap end to end, watch Seedance 2.5 swap the dancer in a real dance clip for an AI character on YouTube, with a Nano Banana 2 character sheet wired in as the reference image.

Pricing

  • Cost: ~27 credits per second
  • Per-second rate: The figure is the rate for a 5-second clip. Shorter clips can cost a bit more per second. You're always charged for the real length.

Variant family

  • Family: bytedance_seedance
  • Is primary: no
  • All variants: video:seedance_2_5, video:seedance_2_5_flash

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 registered materials (up to 8 per run, 5 on the older inline route). 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): Seedance runs a start/end-frame task OR a reference task, never both (the mix is refused before you are charged). With a Reference Video wired, this image is sent as one more reference image instead, so the video no longer opens on it. A Start Frame shot cannot take your own voice: with Generate Audio on, the model makes its sound.
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. Audio the model listens to; this track never plays in the result. In an edit (a Reference Video wired and a prompt that changes it, like "replace the man in @Video1"), wire the audio of that same clip here, turn Generate Audio off, and add this line to the prompt: "Match the framing of @Video1 in every frame: same face size and position, same distance from the camera. Never move closer or crop tighter." Without it, the audio can pull a wide shot into a close-up. The mouth then follows the timing and big shapes of this track, close but not frame perfect, and the clip comes back silent. Check its length against the clip (short clips have come back about 6 percent fast), retime it in the Video Editor if it is off, then lay the audio back in. Keep Generate Audio off when the track is a copyrighted song, because with it on Seedance can refuse the generated audio for copyright. With Generate Audio on, Seedance re-voices the track (timbre, timing and words), so the lips will not match it exactly. Without a Reference Video, it cannot be combined with a Start Frame (the mix is refused before you are charged). mp3 or wav only.
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 registered materials per run (up to 8, or 5 on the older inline route).
audio_urls Audio AUDIO · list. Optional reference audio. Shares the reference-audio slot with Reference Audio, and wiring both concatenates them: up to 2 tracks and 30 seconds between the two. It behaves exactly like Reference Audio (see that port for lip sync in an edit) and never plays in the result. Without a Reference Video, it cannot be combined with a Start Frame.

These render as form fields in the right-side config panel when the node is selected.

ID Label Details
quality Model TEXT · options: standard, flash · default: "standard". Seedance 2.5 Flash costs about 35 percent less per second. A run with a reference clip holds credits for the longest output it could make until it finishes, then settles on what was used. It makes 480p or 720p only (no 1080p or 4K), has no draft mode, and takes at most 5 reference videos per run.
ratio Aspect Ratio TEXT · options: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, adaptive · default: "16:9"
duration Duration TEXT · options: auto, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30 · default: "5"
resolution Resolution TEXT · options: 480p, 720p, 1080p, 4k · default: "720p". 4k costs about 4.7 times as much per second as 720p and is the slowest tier: Seedance makes one 4k clip at a time for all of Wireflow, so a 4k run can wait in a queue behind another before it starts.
generate_audio Generate Audio BOOLEAN · default: true. On: the clip comes back with sound the model generates, guided by any Reference Audio. Off: the clip comes back SILENT, and it never keeps your own track. For lip sync in an edit, turn this off, wire the audio of the clip being edited into Reference Audio, and add this line to the prompt: "Match the framing of @Video1 in every frame: same face size and position, same distance from the camera. Never move closer or crop tighter." The mouth then follows the timing and big shapes of that track, close but not frame perfect; check the take length against the clip and retime it in the Video Editor if it is off, then lay the audio back in. Keep it off when that audio is a copyrighted song, because with it on Seedance can refuse the generated audio for copyright. Outside an edit, with this on, Seedance re-voices any Reference Audio, so the lips will not match your own track exactly.
watermark Watermark BOOLEAN · default: false
contentFilter Content filter BOOLEAN · default: true. On by default. Seedance 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 Seedance 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, 9:16, auto. 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 up to 8 per run, or 5 on the older inline route.

Content filter, and what turning it off costs

Seedance moderates a 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 Seedance to skip that safety and IP moderation.

Two things to know before you do:

  • It costs 10 percent more. The run is billed 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.

4K

Set Resolution to 4k (resolution: "4k" over the API or MCP) for a 3840x2160 clip at 16:9, or the same pixel count at the other ratios. It works in every mode the card has: prompt only, Start and End Frame, reference images, and a Reference Video. Credits with no reference video, by length:

Length 720p 1080p 4k
5 s 135 400 640
10 s 270 800 1280

A run with a Reference Video holds more up front and settles on what the run used, as at every resolution. 4k comes back from Seedance as 10-bit H.265, and Wireflow delivers an H.264 copy so it plays everywhere. It is the slowest tier: Seedance makes one 4k clip at a time for all of Wireflow, so a 4k run can wait in a queue before it starts.

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.

Seedance 2.5 Flash

Seedance 2.5 Flash is a setting on the Seedance 2.5 card, not a separate node. Set Model to Seedance 2.5 Flash in the sidebar. Over the API or MCP, set quality: "flash" in the Seedance 2.5 node's config.

Flash costs about 35 percent less per second than Seedance 2.5. Credits with no reference video, by length:

Length 480p, 2.5 480p, Flash 720p, 2.5 720p, Flash
5 s 75 50 135 90
10 s 150 100 270 180

A run with a reference video holds credits for the longest clip it could make until it finishes, then settles on what the run used, so the difference comes back. With Duration on Auto that is 30 seconds. Duration 5 is the default and holds like Auto, 30 seconds. Any other length holds that length or the reference clip's, whichever is longer. That hold can be larger than Seedance 2.5's for the same short clip: at 720p with a 5 second reference clip and Duration on Auto, Flash holds 860 credits where Seedance 2.5 holds 611, and you need that balance to start the run.

What Flash does: everything the card does (a prompt, a Start Frame and End Frame, reference images, video and audio, native audio, 4 to 30 seconds or Auto), at 480p or 720p.

What it does not do: 1080p, 4k, draft mode, or more than 5 reference videos. Flash takes at most 10 reference images, 5 videos and 5 audio tracks; the card already takes at most 8 images and videos together and 2 audio tracks, so the video cap is the one you can reach. Each of those is refused before you are charged; set Model back to Seedance 2.5 for those shots.


Auto-generated from the Wireflow node registry.

For AI agents: the full documentation index is at /llms.txt, and most docs pages are available as Markdown by adding .md to the URL.

© 2026 Wireflow. All rights reserved.