Seedance 2.5 Guide: What It Is, Where to Use It and How to Prompt It

Seedance is the video model family from ByteDance's Seed team, and version 2.5 is the one most people mean when they talk about reference-driven AI video in late 2026. It arrived on 31 July 2026 with two headline changes over Seedance 2.0: much longer single generations and a far larger reference set. This guide covers what it does, where to find it, how to prompt it and where it still falls short.

Seedance 2.5 at a glance

Where to use it

• Dreamina, ByteDance's creative app, is the main home for Seedance. It is where the Omni-Modal Reference mode lives. Our Dreamina image and video guide covers the app itself.
• CapCut offers Seedance 2.5 inside its video editor, which is handy if you edit there anyway: generate a shot and drop it straight into the timeline.
• BytePlus ModelArk is the API route for developers. Many third-party video platforms host the model through it, so you may also find Seedance 2.5 in a model picker elsewhere.

Access has rolled out in stages and can differ by account, region and plan. If you don't see Seedance 2.5 in the model list, that is usually the reason.

The modes

Text to video

Describe the shot and Seedance generates picture and sound together. Good for quick ideas, less predictable for anything with a specific look.

Image to video

Start from a still you have approved. This is still the most reliable way to control the look, because the model starts from your frame rather than inventing one. Our image-to-video comparison sets Seedance against the other models for this job.

Reference to video

This is where Seedance 2.5 stands out. You can attach up to 50 references: character stills, product shots, style frames, storyboard images, video clips for motion and audio for voice, music or effects. ByteDance also describes green-screen and white-model (plain 3D blocking) references, so you can block out movement roughly and let the model dress it. See consistent characters: which platform for what for how this compares.

Editing what you made

Seedance 2.5 can change one region or one moment of a generated clip, such as removing an object or changing a background, without regenerating everything else. CapCut calls this Intelligent Edit Mode.

How to prompt Seedance 2.5

Seedance reads long, structured prompts well. A layout that works:

1. Subject and setting in one sentence.
2. Action, in the order it happens. With 30 seconds to fill, say what happens first, next and last.
3. Camera: one move per shot, named plainly (slow push in, static wide, handheld follow).
4. Sound: ambience, effects and any spoken line in quotation marks. Keep dialogue short.
5. References: say what each one is for. "Use image 1 for the character's face and clothes, video 1 for the walking motion" works better than attaching files and hoping.

Two habits save credits. First, test at a short length before asking for the full 30 seconds; a 30 second clip that goes wrong at second 4 is expensive. Second, keep references consistent: three photos of a product in different lighting will confuse the model more than one clean photo.

Seedance 2.5 or another model?

Seedance 2.5 is the obvious choice when you have a lot of reference material to hold together, or when a shot needs a full 30 seconds with sound. For a single short shot from one still, Kling 4.0 and Hailuo H3 are just as capable, and for exact camera paths Luma Ray3.2's keyframes give more control. Most people get the best results by testing the same still in two models and keeping Seedance for the reference-heavy shots.

Limits to know

• Long Video Mode is beta. The 180 second projects are built to hold characters, lighting and sound together, but expect drift and plan to fix sections.
• Specs vary by surface. Resolution, length and reference limits depend on where you use the model; an API plan or a partner app may cap lower than Dreamina or CapCut.
• Lip sync is not perfect. Short lines work better than long speeches.
• Text in the frame, such as signs and labels, can still come out garbled. Add critical text in your editor.
• Usage rules apply. CapCut says Seedance videos can be used for ads and brand content within its terms of service; check the terms of the app and plan you use.

Try it

Give Seedance one character still and one short reference clip of someone walking, then run this prompt.

A woman in a mustard raincoat walks along a wet city pavement at dusk, neon reflections in the puddles. She stops, looks up at the rain and smiles. Camera: slow tracking shot beside her, then a gentle push in on her face. Sound: rain on the pavement, distant traffic, a soft "finally". Use image 1 for her face and coat, video 1 for the walking motion.

Check whether the coat stayed the same colour throughout, whether the walk matched your reference, and whether the word landed on her lips. Then run the same still through one other model to see where Seedance is stronger.

Reference-heavy work means a lot of inputs and versions. Dreamdrive, a free Chrome extension, saves each clip you make with its prompt, model and settings, so you can find the version that worked and reuse its prompt next time.