Seedance 2.5: Three Video Modes and Token Costs

TL;DR. Seedance 2.5 is a ByteDance video model released on July 31, 2026. It works in three modes: from text, from a single photo and from a set of references that the prompt refers to as @Image1, @Video1, @Audio1. A clip runs from 4 to 30 seconds in one shot, at 480p, 720p or 1080p, and sound is generated together with the picture without affecting the cost. The price is counted per second: a 4-second clip in 480p is 64 tokens, thirty seconds in 720p is 1020.

Three modes, not one

Because of the hype around "fifty references", half of the texts about 2.5 describe only the third mode. There are three modes, and the first two cover most tasks.

Mode	What you provide	When to use it
From text	only a prompt	you have an idea but no source material
From a photo	one image + a prompt	animate a specific frame, product, portrait
From references	up to 50 files: images, video, audio	you need control over the character, style, motion and sound

In the studio this is the "Network type" switch (Text2Video / Image2Video) and the generation choice, with the reference mode placed on a separate line, Seedance 2.5 Reference. All three have the same settings: duration, resolution, aspect ratio, sound, bitrate.

What changed compared with 2.0

	Seedance 2.0	Seedance 2.5
Duration in one pass	4 to 15 seconds	4 to 30 seconds
References	up to 12	up to 50 (30 images, 10 videos, 10 audio files)
A fast and cheap tier	yes: Fast and Mini	no

The last row matters more than it seems. 2.0 has lightweight versions for drafts, 2.5 does not: every attempt costs the full price. So iterating on ideas with 2.5 is expensive, and for that there is Seedance 2.0 Mini.

ByteDance also claims "20% better prompt following", but does not publish the comparison method, so the claim cannot be checked.

Multi-reference: files are addressed by number

Each uploaded file gets a number, and in the prompt you refer to it directly. That is the point of the third mode: you cannot describe a specific face in words, but you can easily show it with an image.

The character from @Image1 walks down the street from @Image2.
The camera follows, medium shot, evening light.
The character of the motion is like in @Video1, the rhythm follows @Audio1.

Don't do this: "make a video with the person from the first picture in the style of the second".

The model will not match "the first picture" to a file on its own. Without the @Image1 token the references stay plain attachments and generation goes by the prompt text: you will get a clip and the tokens will be charged, but there will be no resemblance to your source.

The division of labor is simple: a reference answers "who and what", the prompt answers "what is happening".

Show with a reference	Describe in words
the face, figure, clothes of the character	what the character does
a specific object, packaging, logo	how the camera moves
style and color palette	where the light comes from
the character of the motion (from a video)	a change of state in the frame

How many tokens it takes

The cost is counted for every second of the clip, and resolution adds its own surcharge to each second. Hence the rule: duration is the most expensive setting, not quality.

Duration	480p	720p	1080p
4 seconds	64 tokens	136	252
5 seconds	80	170	315
10 seconds	160	340	630
30 seconds	480	1020	1890

The logic is simple: a second in 480p costs 16 tokens, in 720p 34, in 1080p 63. The price grows in proportion to the length, with no thresholds or volume discounts. The minimum clip is 4 seconds; the model does not make shorter ones. The estimate is recalculated as soon as you move the duration slider or change the resolution and is always visible before the run, so check against it rather than this table: we review the prices in model groups. What tokens are and how they are charged is on the pricing page.

Important about the reference mode. If you provide a video as input, you pay for it too: ByteDance bills the sum of input and output seconds, and we count it the same way. A five-second clip with a five-second video reference costs twice as much as the same clip without a reference. Images and audio on the input are not billed this way, only video. The total is always visible on the button before the run.

Sound is the exception: the "Audio" toggle does not change the cost, the soundtrack is born in the same pass as the picture.

How long to wait for the result

Video generation is not instant, and it is worth planning for. According to our production data: a 4-second 480p clip from text took 60 seconds; 5 to 8-second clips from a photo took from 3 to 7 minutes. The output file is 854×480 at 24 frames per second for 480p; if the sound toggle is off, there is no audio track in the file at all.

The spread is wide and depends on the provider's queue, so "I'll run ten variants over lunch" is a bad plan. One well-thought-out prompt and one run is better.

The order that saves tokens

1. Collect the references in advance and look at them: a small or blurry face on the input gives the same on the output.
2. The first run: 4 seconds and 480p. The goal of this take is not beauty but a check: whether the right objects were picked up by the @Image1 tokens.
3. Once you are sure the references read, raise the resolution. There is no point paying for 720p before the check.
4. Increase the duration last. A composition mistake at thirty seconds costs seven and a half times more than at four.
5. Do not ask for two actions at once in a four-second clip. A camera push-in plus the character turning around do not fit: by the third second the hero drifts out of the frame. One action per clip.

FAQ

How many references does the Reference mode accept? Up to 50 files: 30 images, 10 videos and 10 audio files. In the interface the attachment counter shows how many are already attached.

Is sound generated together with the video? Yes, with the "Audio" toggle: the picture and the sound are computed in one pass, and the token cost does not grow because of it. If you provide your own audio as a reference, the motion adapts to it.

Why does the price change when I move the duration slider? Because the cost is counted in seconds: each added second adds to the estimate, and the resolution raises the price of the second itself. The total is always visible on the button before the run.

Can I get 1080p? The interface has this option. But the model's own ceiling is 720p with native 1080p added in August; anything above that is a separate upscaling step.

How is 2.5 different from 2.0 in one sentence? A clip twice as long and four times more references, but no cheap tier, and that is what you pay extra for.

---

You can build your first clip in the video studio: choose Seedance in the model list, generation 2.5, set 4 seconds and look at the estimate before the run. The full description of the family is in the Seedance card. No account yet? Create one: the starting tokens are enough for a trial image, and for video you will need a top-up.