Seedance 2.5
ByteDance production-grade video model: native 30-second clips, large reference sets and precise localized edits
What it's the best tool for
- Native clips of up to 30 seconds in a single generation, no stitching of short takes
- Synced audio produced inside the same generation
- Four modes: text-to-video, image-to-video, video-to-video and editing
- Large reference sets: up to 30 images, 10 videos and 10 audio files, referenced from the prompt
- Localized edits of one area of the frame without rebuilding the rest of the scene
- First-frame and last-frame control, multilingual prompts up to 10000 characters
When to reach for something else
- Maximum resolution is 720p, there is no 1080p or 4K option
- Negative prompts are not supported
- Fixed seed is not supported, an exact repeat of a result is not achievable
- 720p costs roughly twice as much as 480p, and video-to-video costs more than generating from scratch
- Long clips take noticeably longer to render than short ones
How Seedance 2.5 responds
Four scenarios where it pays for itself
More about Seedance 2.5
Seedance 2.5: 30-Second AI Video with Audio
Seedance 2.5 is a video model from ByteDance, released on 7 August 2026. Its defining feature is length: a clip of up to 30 seconds is produced natively, in a single generation, rather than stitched together from short takes. Lighting, characters and the logic of motion hold across the whole scene, and audio is synced to the picture from the start.
Four working modes
The model covers text-to-video, image-to-video, video-to-video and editing of existing footage. Localized editing deserves a separate mention: you can change one area of the frame — an object, a caption, a detail of clothing — without rebuilding the rest of the scene. For iteration this matters, because you do not have to regenerate the entire clip and lose shots that already worked.
Large reference sets
Seedance 2.5 accepts substantial inputs: up to 30 images, up to 10 videos and up to 10 audio files. You can point to them directly in the prompt text — where the character comes from, where the visual style comes from, where the voice or the musical bed comes from. That approach suits series work, since one reference library keeps a dozen clips visually consistent. First-frame and last-frame control is available as well, which helps join neighbouring scenes.
Formats and prompts
Output is available at 480p and 720p across six aspect ratios, and a single request returns up to four variants. Prompts are multilingual and long, up to 10000 characters, so a scene can be described in detail: shot size, camera movement, dialogue, atmosphere and character behaviour.
Limitations
The resolution ceiling is 720p, with no 1080p or 4K option. Negative prompts and fixed seeds are not supported. Rendering at 720p costs roughly twice as much as 480p, and video-to-video costs more than generating from scratch. Long clips take noticeably more time to render than short ones, which is the trade-off for a native 30-second scene.
How to try it
The model runs on NetRoom online in the browser: no installation and no keys. The per-second rate is shown in a separate block on this page.
What changed Seedance 2.5
- + Added the Seedance 2.5 (ByteDance) video model to the generation catalog.
Try Seedance 2.5
right now
Free access to basic models. No card, no obligations.