ByteDance ships Seedance 2.5, a 30-second video generation model
ByteDance's Seed team officially launched Seedance 2.5 on 2026-07-31, a new-generation video creation model built on the multimodal audio-video joint-generation architecture introduced in Seedance 2.0. The team says user expectations have shifted from generating a single clip to completing a full creative work, and frames the new model around foundational generation and reference-based generation, with gains in long-form storytelling, multimodal reference, and editing.
The single-pass generation limit doubles from 15 to 30 seconds, and the model supports multiple rounds of extension so a user can chain clips into a multi-minute video while keeping characters, environments, and narrative pacing consistent. ByteDance also says shot transitions and scene changes are stronger in longer sequences, with better image, audio, and motion quality. The announcement gives worked examples: a one-take backstage-to-stage concert sequence, a Peking opera scene where the camera circles a performer's flowing sleeves, and a chase scene extended across multiple 30-second rounds using a reference video.
On referencing, Seedance 2.5 now accepts up to 30 images, 10 video clips, and 10 audio clips as inputs in a single pass, up from what earlier versions supported. The model reads composition, scene, style, characters, and props across all the supplied material and applies them to the generated video, including in multi-character or group scenes where it aims to keep each subject's appearance and voice stable. ByteDance highlights three specific reference modes: clay-render referencing, where a textureless 3D model supplies spatial structure, camera angles, and blocking; motion referencing; and creative referencing. For clay render specifically, the model also derives realistic lighting, such as light direction, color temperature, and shadow, from the render's spatial information.
On editing, Seedance 2.5 adds timestamp-level control over audio and video content for targeted edits, plus enhanced green-screen, camera-perspective, and reference-based editing aimed at professional use in film and advertising. ByteDance also says the model reduces uncontrolled artifacts in subtitles and background music and generally cuts down on the artificial look common in AI-generated video, through changes to object textures, skin and eye rendering, lighting, and color saturation. The source text is cut off mid-sentence at the start of a further editing-features passage, so any additional claims made there are not captured here.
Seedance 2.5 is rolling out today on Jimeng AI and Doubao Pro, reached through each platform's video-generation menu by selecting Seedance 2.5. API access via BytePlus ModelArk is described only as coming soon, with no date given. The announcement discloses no pricing for the model or for the upcoming API access, and includes no benchmark results or third-party comparisons against other video-generation models.
Key facts
- Seedance 2.5 extends single-pass video generation from 15 to 30 seconds and supports multi-round extensions for multi-minute output.
- The model accepts up to 30 images, 10 video clips, and 10 audio clips as reference material in a single pass.
- New editing controls include timestamp-level targeting plus improved green-screen, camera-perspective, and reference-based editing.
- Seedance 2.5 is rolling out today on Jimeng AI and Doubao Pro; API access via BytePlus ModelArk is coming soon with no date given.
- No pricing is disclosed for the model or for the upcoming API access, and no independent benchmarks are provided.
Why it matters
The announcement frames Seedance 2.5 around a shift from generating a single clip to completing a full creative work: doubling the single-pass limit to 30 seconds, adding multi-round extension for multi-minute output, and pushing image, audio, and motion quality specifically to cut down on the artificial look common in AI-generated video. That combination, longer coherent takes plus more controllable referencing and editing, targets production workflows rather than one-off clip generation.
Who it affects
Video creators and marketers already on Jimeng AI or Doubao Pro get the new model directly through each platform's video-generation menu. Professional users in film and advertising are the explicit target for the upgraded editing tools (timestamp-level control, green screen, camera perspective, reference-based editing). Developers who want programmatic access have to wait, since API access via BytePlus ModelArk is only described as coming soon.
How to use it
Seedance 2.5 is reached via Jimeng Web, Video Generation, then selecting Seedance 2.5, or via Doubao Pro through the same path. API access is coming through BytePlus ModelArk but is not yet available and carries no announced date. The announcement gives no pricing for either the platform access available today or the future API.
How solid is it
This is ByteDance's own product announcement, self-describing its results with worked example prompts and an accompanying showcase film rather than third-party testing. No benchmark scores or comparisons against competing video-generation models are included, so the quality and consistency claims come only from the vendor's own framing and examples.
Risks and caveats
All performance claims are vendor-stated and illustrated with curated example prompts, not independent benchmarks. Pricing for both the current platform access and the forthcoming API is undisclosed, and the API's timeline is only "coming soon" with no date attached. The source text is also truncated mid-sentence in its final editing-features section, so some described capabilities beyond what is quoted here may not be fully captured.
“Seedance 2.5 centers on foundational generation and reference-based generation, delivering major breakthroughs in long-form storytelling, multimodal reference, and editing.”
— the Seedance 2.5 announcement