What Is Seedance 2.5? ByteDance's New Video Model

Seedance 2.5 is the video generation model released on 2026-07-31 by the Seed research team at ByteDance. It produces clips of up to 30 seconds in one pass, with synchronized audio and speech generated alongside the picture, and it can hold subjects, products and style steady across a shot using reference images. On VIDEO AI ME it runs in English, in the browser.

Both panels above were given the same 15 second script. Seedance 2.5 is on top, Seedance 2.0 below. The spot opens on a woman walking on a mirror calm ocean at sunset, cuts to a founder at a laptop, and ends on a brand card. Watching the two run side by side is the fastest way to understand what actually changed, and the rest of this article explains why.

What Is Seedance 2.5, Technically?

Seedance 2.5 is a text and image conditioned video model. You give it a written description of a shot, optionally one or more images, and it returns a video file with an audio track already attached. It sits in the same family as Seedance 2.0, ByteDance's previous release, but it is not a tuned version of it. The generation ceiling, the audio behaviour and the way it handles references are all different.

Three things define it in practice:

1. Length in a single pass. Up to 30 seconds is generated as one continuous piece of reasoning about the shot, rather than assembled from shorter clips that were rendered separately and joined afterwards.
2. Native audio. Sound effects, ambient sound and lip synced dialogue come out of the same generation. There is no second pass, no separate voice model, no manual sync.
3. Multimodal referencing. Attach images and the model carries what is in them across the whole shot: a face, a product, a colour treatment.

ByteDance launched the model inside its own consumer apps, Jimeng AI and Doubao. Both in practice expect a Chinese phone number and a Chinese language interface, and API access was announced as arriving later through BytePlus ModelArk. That gap between "the model exists" and "you can use the model" is why most people searching for it never actually reach it. You can read the Seed team's own research index at seed.bytedance.com and BytePlus's platform documentation at byteplus.com.

What Is Seedance 2.5 Able to Do That 2.0 Could Not?

The honest summary is that 2.0 makes shots and 2.5 makes scenes. Here is the difference in the terms that matter when you are producing ad creative.

Roughly double the cost per second, for length, sound and consistency you previously had to build by hand. If you want the full breakdown of that trade, we wrote it up in our Seedance 2.5 versus Seedance 2.0 comparison, and the older model still has its own Seedance 2.0 explainer if you are coming to this fresh.

The 30 Second Single Pass

Most video models are short by design. You generate five or eight seconds, then generate another five or eight, then try to make the two match on the cut. Anyone who has done this for a client knows the failure mode: the jacket changes shade, the light moves, the room grows a window.

Seedance 2.5 reasons about the whole duration at once. On VIDEO AI ME you pick a duration from 4, 8, 12, 15, 20, 25 or 30 seconds, and a 30 second selection is generated as one piece. That means the model knows at second two what has to be true at second twenty eight. Practically, you get scene changes that hold: a subject can turn, walk out of frame, and reappear in a different setting with the same face and the same wardrobe.

This changes how you write. A prompt that works well at eight seconds is usually a description of one moment. A prompt that works well at thirty seconds is a small story: a setup, a development, a turn, a resolution, with the transitions named out loud. Our Seedance 2.5 prompt guide covers the structure in detail.

Native Synchronized Audio

Audio is generated with the picture, not bolted on afterwards, and it is included in the price. There is no toggle and no surcharge on VIDEO AI ME. What you get is three layers at once:

• Speech. Dialogue you write in quotes is spoken by the on screen subject, with lip movement that matches.
• Sound effects. Footsteps, a laptop lid, water, a door.
• Ambient bed. The acoustic character of the space the shot is set in.

The implication for advertisers is that a hook line no longer needs a separate voiceover step. Write the line into the prompt and the person in the frame says it. If you do not want music, say so in the prompt, because the model will often add a bed if you leave the question open.

Multimodal Referencing and Why It Matters for Ads

Reference to video is the mode that ecommerce operators tend to care about most. On VIDEO AI ME the mode is chosen automatically by how many images you attach:

In reference to video you address the images in the prompt as @Image1, @Image2, @Image3 and @Image4. That explicit binding is the whole trick. "Keep the creator from @Image1 holding the bottle from @Image2, kitchen counter, morning light" gives you a repeatable creator and a product that stays the right shape. Vague references get blended, which is the most common reason a first attempt looks wrong.

Model Capabilities Versus What VIDEO AI ME Ships

This distinction matters, so it gets its own section. ByteDance's launch material describes model level features including video references, audio references, acceptance of many more reference images, timestamp level editing, green screen editing and multi round extension. Those are properties of the model as ByteDance describes it. On VIDEO AI ME today, what you get is text to video, image to video, and reference to video with up to four images, at 480p or 720p, in durations from 4 to 30 seconds, with audio always generated.

Aspect ratios available on VIDEO AI ME are auto, 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16, except in image to video where the output follows the input image. That is the honest boundary. Plan your production around the two lists in this section, not around the launch announcement.

Realism: Textures, Skin, Eyes and Light

The realism work in 2.5 is the part that is hard to put a number on and easy to see. In our runs, the differences that show up most consistently are in materials and faces. Fabric holds a weave instead of turning into a smooth painted surface. Skin keeps pore level texture under strong light rather than going waxy. Eyes track and hold a catchlight through a camera move, which is the single detail most likely to break the illusion when it fails. Light behaves like light: a low sun blooms and halates, reflections sit in the right place on a wet surface.

We are describing what we observed in the comparison render above, not a benchmark score. Do not trust anyone quoting precise realism percentages, including us. Generate your own shot and look at the faces.

What It Costs on VIDEO AI ME

Credits are the billing unit and one credit equals one cent of generation cost. Seedance 2.5 costs 23 credits per second at 480p and 48 credits per second at 720p. Audio is included in that.

Plans are Starter at $29 a month with 1,400 credits, Pro at $99 with 5,600, and Premium at $199 with 12,000. One thing to plan around: a full 30 second clip at 720p costs 1,440 credits, which is more than the whole Starter allowance. Thirty second 720p work starts at Pro. Starter handles 30 second clips at 480p comfortably at 690 credits each. Video generation requires an active subscription.

Should You Use It?

If you are producing paid social creative, the case is straightforward: longer coherent takes and generated dialogue remove two production steps you were previously paying for in time. If you are testing hooks in volume at ten variants a day, cheaper per second models are still the right tool for the testing round, and you promote the winner to 2.5 for the version you actually run. Our hands on Seedance 2.5 review goes through where it held up and where it cost us, and the ranked list of AI video models puts it next to everything else worth considering.

To generate with it, open the model picker in the editor and select Seedance 2.5. The step by step walkthrough covers settings, duration and the reference workflow.

Frequently Asked Questions

What is Seedance 2.5?

Seedance 2.5 is a video generation model released on 2026-07-31 by ByteDance's Seed team. It generates clips of up to 30 seconds in a single pass with synchronized audio, including sound effects, ambient sound and lip synced speech, and can carry subjects and products across a shot using reference images.

Who made Seedance 2.5?

ByteDance's Seed research team built it. ByteDance is also the company behind TikTok and the Jimeng AI and Doubao consumer apps, where the model first appeared.

How long can a Seedance 2.5 video be?

On VIDEO AI ME you can select 4, 8, 12, 15, 20, 25 or 30 seconds. A 30 second clip is generated in a single pass rather than stitched from shorter renders, which is why continuity holds across scene changes.

Does Seedance 2.5 generate sound?

Yes. Sound effects, ambient sound and lip synced dialogue are generated with the video and included in the price. There is no audio toggle and no audio surcharge.

How much does Seedance 2.5 cost?

On VIDEO AI ME it costs 23 credits per second at 480p and 48 credits per second at 720p, where one credit equals one cent of generation cost. A 15 second 720p clip is 720 credits. Plans start at Starter, $29 a month for 1,400 credits.

Can I use Seedance 2.5 in English?

Yes. On VIDEO AI ME the model runs in English in a normal browser on an active subscription. ByteDance's own apps, Jimeng AI and Doubao, are Chinese language and in practice expect a Chinese phone number.