Seedance 2.5 — the complete guide
Dreamina Seedance 2.5 is ByteDance's next-generation AI video model — 30-second one-take generation, 50 multimodal references, and native video editing built into the API. Here's what's actually new, what it costs, and how to use it.
What's new vs. Seedance 2.0
| Capability | Seedance 2.0 | Seedance 2.5 |
|---|---|---|
| Max single-take duration | 15 seconds | 30 seconds |
| Reference assets per request | 15 (9 images + 3 videos + 3 audio) | 50 (30 images + 10 videos + 10 audio) |
| Native video editing | Limited | Dedicated task type — add / remove / replace |
| Native video extension | Limited | Dedicated task type — extend forward/backward |
| Audio-only reference input | Not supported | Supported |
| Resolution | 480p – 4K | 480p, 720p only |
| Output formats | mp4 | mp4, mov (higher color fidelity) |
| Languages | 4 languages | 11 languages |
Key capabilities
30-second one-take generation
Where Seedance 2.0 caps out at 15 seconds, 2.5 can generate a coherent 30-second clip in a single pass — scene changes, tempo shifts, and a full story arc handled without stitching separate clips together afterward.
Native video editing & extension
This is the standout feature. Instead of regenerating a clip from scratch, Seedance 2.5 can take an existing video and:
- Edit it — add, remove, or replace a specific object or element from a text description
- Extend it — continue the clip forward or backward in time, optionally guided by additional reference footage
Both task types are auto-detected from your prompt wording (e.g. "add a hat to the character" triggers editing) combined with a video being present in the request.
50 multimodal references
Up to 30 reference images, 10 reference videos, and 10 reference audio clips can be combined in one generation — more than triple Seedance 2.0's limit — letting you assemble complex, multi-source scenes (character look + location + motion reference + soundtrack) in a single request.
Pricing
Seedance 2.5 is billed by tokens through BytePlus's ModelArk platform, at a lower rate when a video is included as input than for text/image-only requests:
| Input type | Rate |
|---|---|
| With video input (editing, extension, reference-to-video with a video) | $0.0064 / 1K tokens |
| Without video input (text-to-video, image-only reference) | $0.0107 / 1K tokens |
Token count follows width × height × frame rate × duration ÷ 1024 — so cost scales with resolution and length, same as the rest of the Seedance family.
Use Seedance 2.5 without touching the API
AI Cineworks has Seedance 2.5 wired in as a selectable model across Text-to-Video, Image-to-Video, and OnSet (reference-to-video) — including a one-click "Smart Edit" on any generated video that uses 2.5's native editing capability automatically.
Frequently asked questions
Up to 30 seconds in a single generation, with a complete story arc and scene changes handled in one pass — double the 15-second limit of Seedance 2.0.
480p and 720p. Unlike Seedance 2.0, it does not currently support 1080p or 4K output.
Yes — it has a native video editing task type that can add, remove, or replace elements in an existing video from a text prompt, plus a separate video extension task type to continue a clip forward or backward.
Up to 50 total: 30 reference images, 10 reference videos, and 10 reference audio clips in a single generation request.
Yes — Seedance 2.5 is a selectable model in AI Cineworks' Text-to-Video, Image-to-Video, and OnSet (reference-to-video) tabs.