Seedance 2.5 brings 30-second clips and 50 references

You spend an evening getting your AI avatar's face, outfit and favourite location just right, then the video model gives you eight seconds before the shot falls apart. On October 23, Dreamina announced Seedance 2.5, ByteDance's new video and audio model, with clips up to 30 seconds long and up to 50 references for a single clip.

• Seedance 2.5 generates clips up to 30 seconds long, which Dreamina says makes it easier to tell a story with smoother transitions and stronger character consistency.
• One clip can take up to 50 multimodal references: images, video, audio and text.
• You can edit specific timestamps, characters or other elements inside a finished video instead of starting over.
• On Dreamina, it rolls out to subscriber accounts of people over 16 in Europe, Asia, the Middle East and South America. The announcement does not list the United States for Seedance 2.5.

The jump that matters for avatar creators is the reference count. Seedance 2.0 already made consistent characters practical, but you had to choose which few references to keep. With 50 slots, you can give the model your avatar's face from several angles, the outfit, the room, a voice sample and a clip that shows how the camera should move, all in one request. The length helps too: a 30-second take is long enough for a hook, a point and a payoff in one shot, which is the shape of most reels.

The timestamp editing is the quieter change. Until now, one wrong hand gesture at second 19 meant generating the whole clip again and hoping the rest came out as well. Being able to point at a moment and change only that part saves both credits and the take you already liked. Seedance 2.5 is also showing up outside Dreamina: Runway lists it in Runway Creative, and ByteDance's cloud has a separate 4K version for film studios in an enterprise trial, with a 480p Draft mode for testing shots first.

How to use Seedance 2.5 now that clips run 30 seconds

The following images were generated using Nano Banana 2:

1. Build a reference pack before you write a prompt. One clean front-facing portrait, two three-quarter views, one full-body outfit shot, one empty location and one short voice sample. Fifty slots is a ceiling, not a target; a small, clean pack beats a big, messy one.
2. Give every reference one job. Say in the prompt what each one is for: this image is the face, this one is the jacket, this one is the room. References that overlap (two different jackets, two different rooms) make the model guess.
3. Plan the 30 seconds as beats. Write the clip as three or four short moments: the opening line, the reaction, the reveal, the closing look. A long take with no plan drifts; a long take with beats reads like a scene.
4. Fix one moment instead of rerolling. When most of the clip works, use the timestamp edit on the part that doesn't. Where a Draft mode is offered, test the idea at low resolution first and only render the final version once the beats land.

Worked example — the character reference. A plain, evenly lit portrait gives the model the face and nothing else to copy, so it can put him in any room you choose.

Prompt (Nano Banana 2, 1:1):
Studio portrait of a South Asian man in his late twenties with short black hair and a trimmed beard, wearing a plain olive crew-neck t-shirt, facing the camera straight on with a relaxed neutral expression, even soft front light from a large softbox, plain light grey seamless backdrop, head and shoulders, sharp focus on the eyes, natural skin texture, 85mm lens, no text, no logos

Worked example — the location reference. An empty room tells the model where the scene happens without adding a second person it might try to blend with your avatar.

Prompt (Nano Banana 2, 1:1):
Empty ceramics studio in the afternoon with no people, an unbranded potter's wheel with no lettering in the centre, shelves of unglazed bowls and vases along a white brick wall, large industrial window letting in warm sunlight, dust in the light beams, wide eye-level view, well-exposed, no text, no logos

Our honest take: the 50-reference limit will matter most to people who already keep a tidy character sheet. If your avatar lives in five random screenshots, start by making a proper reference pack; that is where the consistency comes from. Availability is also uneven for now, so check that your Dreamina plan and region have Seedance 2.5 before you plan a week of content around it.

Receipts: Dreamina's Seedance 2.5 announcement, TechNode on Seedance 2.5 4K, Runway Creative in Claude Marketplace.

The bottom line for AI avatar creators

Seedance 2.5 moves the limit from "how long can the shot hold" to "how well do you know your character". Thirty seconds and fifty references reward creators who have a fixed face, a fixed outfit and a clear plan for each clip, and the timestamp edit means one bad moment no longer costs you the whole take.

If you don't have a locked face for your avatar yet, that is the first job, before any video model: start with your avatar.

FAQs about Seedance 2.5

How long can a Seedance 2.5 clip be?

Up to 30 seconds per clip, according to Dreamina's announcement.

How many references can I use in one clip?

Up to 50 multimodal references, which can be images, video, audio and text.

Can I change one part of a clip without regenerating it?

Yes. Dreamina says you can edit specific timestamps, characters or other elements within a video.

Who can use Seedance 2.5 on Dreamina?

Dreamina is rolling it out to subscriber accounts of people over 16 in Europe, Asia, the Middle East and South America. The announcement does not list the United States for Seedance 2.5.

Are Seedance 2.5 videos labelled as AI?

Dreamina says its models use invisible watermarks, C2PA Content Credentials and visible AI labels, and that its policies forbid the unauthorised use of someone's likeness.