ByteDance has officially introduced Seedance 2.5, its new generation of video creation model. The announcement, published on July 31, 2026 on the Seed team blog, highlights a shift in what users expect: it’s not enough to generate a clip, you need to complete a creative work.
According to seed.bytedance.com, the model maintains the unified architecture of joint audio and video generation from Seedance 2.0, but now focuses on fundamental generation and reference-based generation. The result is advances in long-form storytelling, multimodal reference, and editing.
30-second clips and multi-round extension
Seedance 2.5 can generate high-quality audio and video clips up to 30 seconds in a single pass. Additionally, it supports multiple rounds of extension, allowing the user to add new scenes to the existing video.
Improved scene transitions and shot changes ensure greater continuity in long videos. Image, audio, and motion quality also saw notable gains, resulting in a more natural and polished look than typical AI-generated videos.
With this, it is possible to produce content of several minutes with consistent audiovisual language, telling a complete story in a single take.
Enhanced multimodal references
The model accepts up to 30 images, 10 video clips, and 10 audio clips as reference materials in a single pass. This represents a complete upgrade from the previous version.
Reference capabilities have been strengthened, including clay rendering, motion, and creative references. The goal is for the model to better understand the creator’s intent and realize complex ideas involving multiple subjects, scenes, and shot changes.
More precise and stable editing
Seedance 2.5 offers timestamp-level control for targeted audio and video editing. This significantly improves the efficiency and controllability of the creative process.
Advanced editing features such as green screen, camera perspective, and reference-based editing have also been enhanced. The idea is to meet the rigorous demands of complex professional fields such as cinema and advertising.
Long-form storytelling in a single pass
Single-pass video generation has been extended from 15 to 30 seconds. Within this interval, the model organizes multiple logically connected shots, allowing a story to develop with presentation, development, turning point, and resolution.
An example cited by the Seed team is a singer at a show: the model depicts the interaction with the crew in the dressing room, the walk down the corridor, the meeting with the dancers, and the ascent to the stage, rather than just showing the moment of entry.
The multi-round extension capability maintains consistency of main characters, environments, and narrative rhythm. This reduces the effort of splitting clips, reassembling, and correcting transitions.
Cinematic visual quality
The model also tackles the artificial look common in AI-generated videos. Object textures, skin and eye characteristics, lighting, and color saturation have been systematically optimized.
Uncontrolled occurrences in subtitles and background music have been minimized. The final result approaches the cinematic quality of real footage.
Availability
Seedance 2.5 is now available on Jimeng AI, Doubao Pro, and other platforms. API access will arrive soon via BytePlus ModelArk.
The Seed team invites users to test the model and share feedback. The project page is available on the official Seed website.