Seedance 2.5 Launches: ByteDance's Video Model Now Generates 30-Second Clips in a Single Pass
On July 31, ByteDance's Seed team released Seedance 2.5, the next-generation video creation model. The update is positioned around one-take creation and flexible referencing, with three areas of focus: long-form storytelling, multimodal reference, and editing.
30 seconds per generation, with multi-round extensions
Seedance 2.5 doubles the single-pass generation length from 15 seconds (in 2.0) to 30 seconds. Within 30 seconds, the model can organize multiple logically connected shots, letting a story unfold through setup, development, turning points, and resolution — rather than simply extending a single moment. In one official example, a clip of a singer's stage performance depicts the full sequence: interacting with staff in the dressing room, walking through the backstage corridor, meeting the dancers, and stepping onto stage together.
Multi-round extension is another key feature. Users can append subsequent shots to existing video outputs, and the model maintains consistency of main characters, environment, and narrative pacing, eventually producing multi-minute content with a unified audiovisual language. ByteDance says shot transitions and scene changes were optimized, keeping subjects stable across cuts with synced audio and visuals.
To address the "oily" artificial look common in AI-generated video, the model was systematically tuned on object textures, skin and eye features, lighting, and color saturation. Uncontrolled subtitles and background music are also reduced.
Multimodal referencing: 30 images, 10 video clips, 10 audio clips
Seedance 2.5 accepts up to 30 images, 10 video clips, and 10 audio clips as reference materials in a single pass. The model understands composition, scenes, styles, characters, and props across the materials and applies them as instructed. In multi-character shots and group scenes, it can preserve the appearances and voices of multiple characters while keeping each subject stable.
Clay render referencing is a highlight of this upgrade. Users can build the spatial structure, character poses, motion paths, and camera angles of a scene using textureless 3D models; the model then generates video based on that structure, so complex shots land closer to the creator's intent. Lighting control also benefits from the clay render's spatial information, producing physically plausible light direction, color temperature, intensity, and shadows.
Timestamp-level editing
During generation, users can control the story, camera perspective, movement, and rhythm of specific time ranges via prompts. After generation, targeted modifications to characters, actions, or plot elements within specific clips are supported while maintaining continuity. Green screen editing, camera perspective editing, and reference-based editing were all strengthened. In green screen scenarios, the model renders how the subject responds to the new environment's physical rules — clothing flutter, hair state, gait rhythm, and lighting interaction.
Platform rollout and industry scenarios
Seedance 2.5 is rolling out on Jimeng AI and Doubao Pro, with API access coming soon via BytePlus ModelArk. In education, it has entered real learning settings — Doubao Learning's "Doubao Classroom" uses the model to turn historical context behind lessons into vivid visuals. In industrial manufacturing, embodied intelligence, and autonomous driving, the model generates high-quality synthetic video data to train robots' perception and manipulation skills, and simulates long-tail scenarios such as extreme weather and complex traffic conditions.
The Hacker News thread for Seedance 2.5 has drawn 140+ points. Commenters generally praise the generation quality — one called it "very close to climbing out the other side of the uncanny valley" — while others repeat the usual complaint about closed weights and no open source. It's a familiar debate in the AI video space: quality and controllability keep rising, but whether the community accepts an API-only model without weights is another question.