Seedance: ByteDance video AI — what 2.0 and 2.5 can do

What Seedance is

Seedance is the video models from the ByteDance Seed team, which also makes the image models Seedream. The idea of the family is one architecture for video and sound: the clip, speech, effects and music are generated together, not stuck on afterwards.

Seedance 2.0 came out on February 12, 2026. It was the first to take four kinds of input at once — text, images, video and audio — and it learned to edit and extend clips. A faster Seedance 2.0 Fast came out with it.

Seedance 2.5 came out on July 31, 2026 on the same architecture. The main difference is length: up to 30 seconds in one generation and two extensions. The model builds several connected shots inside the clip instead of stretching one moment.

What Seedance can do

Several shots in one generation. Describe the scene shot by shot — a wide shot, a medium shot, a tracking shot — and the model assembles them into one clip with cuts and a shared soundtrack.

A character from a reference. In Seedance 2.5 you can upload up to 30 images, 10 videos and 10 audio files at once and say in the prompt what to take from where. The model keeps the character’s face and look in a new scene.

Dialogue and sound. Lines in quotes become speech with synced lips, and the model builds the shot-reverse-shot cutting between the speakers on its own.

Motion and camera. The model holds fast action: sports, jumps, drone flyovers, a slow-down at the peak.

Editing and extension. Seedance 2.0 and 2.5 change individual fragments, characters, actions and the plot of a finished clip. 2.5 adds editing by timecode, green-screen work and control from a “white model” — a rough 3D blockout of the scene.

Seedance 2.5 and 2.0: how they differ

	Seedance 2.0	Seedance 2.5
Released	Feb 12, 2026	Jul 31, 2026
Length per generation	4–15 s	up to 30 s
Extension	yes	two times
References	up to 9 images, 3 videos, 3 audio	up to 30 images, 10 videos, 10 audio
Several shots per pass	yes	yes, a connected story
Editing	fragments, characters, actions	+ timecodes, green screen, “white model”
Faster version	2.0 Fast	—

ByteDance separately names the 2.5 improvements: object texture, skin and eyes, light, color saturation, and how tightly the sound matches the picture.

What Seedance is good for

• Short stories and ads of 20–30 seconds without stitching several generations together.
• Clips set to music: upload audio as a reference for rhythm and sound.
• A series with one character: dozens of photos of the character in one request.
• Sports and action: fast movement, flyovers, slow motion.
• Editing a finished clip: replace a character, an action or a span by timecode.

Seedance and alternatives

	Seedance 2.5	Veo 3.1	Gemini Omni 1.1 Flash
Developer	ByteDance Seed	Google DeepMind	Google DeepMind
Length per generation	up to 30 s	4, 6, 8 s	3–10 s
Extension	2 times	up to 148 s (API, 720p)	up to 40 s
Resolution	not disclosed	up to 4K	up to 4K (upscale)
References	30 photos, 10 videos, 10 audio	up to 3 images	images and up to 3 videos
Sound	yes	yes	yes
In AI2Face	yes	yes	yes

Seedance 2.5 leads on the length of a single clip and the number of references. Veo 3.1 gives 4K and a precise first-and-last-frame mode. Gemini Omni Flash is fast and flexible about what you put in.

How to write prompts for Seedance

• Describe the sound in as much detail as the picture: lines in quotes, music, the key effects.
• For several shots, number them: “Shot 1: wide… Shot 2: medium… Shot 3: tracking…”.
• Name references by number. In ByteDance’s official examples: “the character from @Image 2, the scene from @Image 3, the prop from @Image 4”.
• Test the idea on a short clip, then make the long one once the prompt is set: 30 seconds costs several times more than 5.

A short story in three shots. Shot 1: a wide shot of a fishing village at dawn, fog
over the harbor, boats rocking, seagulls calling. Shot 2: a medium shot of an old fisherman
in a yellow raincoat, he unties the rope of a wooden boat, looks at the sea and smiles.
Shot 3: a tracking shot behind the boat leaving the harbor toward a golden sunrise. Sound: waves,
creaking wood, distant gulls, a rising string melody. Photorealism, 35mm film

How to use Seedance in AI2Face

Seedance is available in the AI2Face generator: clips with sound up to 30 seconds, and work with references. The cost is shown before you start.

Two Google models sit in the same generator for other jobs:

• Veo 3.1 — clips of 4–8 seconds up to 4K, with speech and effects, photo-to-video, up to 3 references, and a first and last frame.
• Gemini Omni 1.1 Flash — clips up to 10 seconds, image references, and a first and last frame.

If you need 4K, build a long story from several 8-second Veo 3.1 clips with one character from a reference. The cost of both Google models is shown before you start.

Limitations

• Real faces. For a portrait of a real person used as a reference, ByteDance requires identity verification or legal permission.
• Complex physics and several characters. ByteDance itself says that the physical plausibility of complex motion, and the stability of scenes where several characters interact, still need work.
• Detail and text. In the Seedance 2.0 evaluation, the developer names unstable detail, rare sound distortions and inaccurate text in the frame.
• Availability. Officially the models run in the Chinese apps Jimeng and Doubao Pro and through an API for developers. At release, the API for 2.5 was marked “coming soon”.

Conclusions

Seedance 2.5 is the model for long story clips: 30 seconds with sound in one generation, and dozens of references, which none of the competitors can do yet. Seedance 2.0 remains a working model for clips up to 15 seconds, with editing. Seedance is available in AI2Face. If you need 4K or a precise transition between two frames, take Veo 3.1 in the same generator. A long story on Veo is built from several 8-second scenes with one character.