ByteDance Releases Next-Generation Video Creation Model Seedance 2.5

The Zhitong Finance App learned that on July 31, ByteDance officially released Seedance 2.5, a next-generation video creation model. ByteDance said that after the release of Seedance 2.0, users' expectations for the video creation model are moving from “generating a clip” to “completing a piece of work.” Seedance 2.5 continues Seedance 2.0's unified multi-modal audio and video co-generation architecture, focusing on breaking through long storytelling capabilities, multi-modal reference capabilities, and editing capabilities. Focusing on real industrial scenarios, creations have more room for imagination and further unleash productivity.

According to ByteDance, the core highlights of Seedance 2.5 are as follows:

The single generation time is up to 30 seconds, and supports multiple rounds of extension: Seedance 2.5 can generate 30 seconds of high-quality video clips in a single run and supports multiple rounds of extension. At the same time, the model optimizes lens connections and scene transitions to maintain better continuity in long videos. The model also greatly optimizes picture quality, sound quality, and motion quality, and effectively reduces the “greasiness” common in video generation. Users can produce several minutes of consistent content with a unified audiovisual language and high quality, and present wonderful stories all at once.

The multi-modal reference capability has been fully upgraded: Seedance 2.5 allows users to input up to 30 images, 10 videos, and 10 audio segments at a time as reference material. The model also enhances various reference capabilities such as white model reference, motion reference, and creative reference, so that the model can better understand creative intentions and realize complex ideas with multiple subjects, multiple scenes, and multiple camera changes.

More accurate and stable editing capabilities: Seedance 2.5 supports precise timestamp control, can target video content, and improve creative efficiency and control. At the same time, the model enhances various editing capabilities such as green screen editing, perspective editing, and reference editing to meet the creative needs of more professional and complex scenes such as film, television, and advertising.

With long narratives, multi-modal references, and improved editing capabilities, models can not only generate longer videos at once, but also better understand creative intent and complete the entire process from idea to video implementation in a more controlled manner. Next, we'd like to invite you to watch a creative short film, all independently generated by Seedance 2.5.

30-second long story, supports multiple rounds of extension, and presents a wonderful story at once

Seedance 2.5 increases the time it takes to generate a single video from 15 seconds to 30 seconds, while further improving storytelling in long videos. The model can organize multiple logically related shots within 30 seconds, so that the story unfolds gradually from paving the way, progression, and transition to the end, rather than simply continuing a single picture. For example, create a clip of the singer taking a mirror to the end of the stage. The model played out the complete story of the singer interacting with the staff in the powder room, meeting the companion dancer through the backstage hallway and performing together, not just the singer.

With the model's ability to extend multiple rounds, users can naturally connect subsequent footage to existing video results, maintain the continuity of the main characteristics, scene environment, and narrative rhythm during the extension process, and finally generate several minutes of coherent content with a unified audiovisual language, reducing the cost of splitting the footage, repeated splicing, and processing and connecting.

In terms of image presentation, the model can better achieve natural mirror connection. The subject remains stable even after multiple camera cuts, and the audio and video are always aligned, making the long video generation results more consistent. For example, in a Peking Opera performance, the camera moves gracefully around the mirror along with the protagonist's water sleeve, and the presentation of the subject and background is always consistent. The water sleeve swings in the air to form a natural arc, which follows the laws of physics.

Furthermore, in response to the common “greasiness” problem in video generation, Seedance 2.5 systematically optimizes details such as object material, character skin and eye appearance, light and shadow, and image saturation. The model also reduced the uncontrolled subtitle and background music, making the film closer to a real film.

Multi-modal reference capabilities have been fully upgraded, and complex creative tasks are more manageable

Seedance 2.5 further enhances multi-modal reference capabilities, allowing users to input up to 30 images, 10 videos, and 10 audio segments at a time as reference material. More quantity and different types of reference material can better restore the ideas in the user's mind and generate complex video content with more subjects, richer scenes, and more flexible lens changes.

The model can comprehensively understand elements such as picture composition, scene, style, characters, and props in different materials, and use these reference materials for video generation according to instructions. In complex scenes such as multiple people in the same frame and group portrait narratives, it is also possible to simultaneously restore the images and sounds of multiple characters and maintain the stability of the characteristics of each subject.

In addition, the model has also improved various reference capabilities such as white model reference, motion reference, and creative reference, making the subject, movement, and motion in the picture more controllable. For example, in the white model reference, users can use an untextured 3D model to create the spatial structure, subject position, motion trajectory, and lens position of the image, and then let the model reference this set of structures to generate videos, making the composition and scheduling of complex lenses closer to expectations. At the same time, the model also enhances lighting control capabilities, and uses spatial information from the white mold to generate lighting effects that conform to real physical laws, including light source direction, color temperature, intensity, shadow projection, etc., to make the final image more natural.

More accurate and stable editing capabilities, effectively improving creative efficiency

In video generation creation, users usually need to control the rhythm during production and continue to polish the details after generation. How many seconds an action appears, when the camera changes, or whether the characters or actions in a clip need to be adjusted are all related to the final effect of the film. More accurate and stable editing capabilities allow users to accurately restore ideas, improve creative efficiency, and reduce the cost of repeated generation.

Seedance 2.5 supports accurate editing of audio and video content through time stamps. At the time of generation, users control the story, image perspective, motion, and overall rhythm of a certain period of time through prompts, making it closer to the creative intention. After generation, users can also make targeted modifications to the characters, movements, sounds, or plot in the specified clip, while maintaining the continuity and realism before and after the revisions.

The model also further enhanced various editing capabilities such as green screen editing, perspective and motion picture editing, and reference editing to meet more professional and complex creative scenarios such as film, television, and advertising. For example, in terms of green screen editing, models can not only replace different scenes and tell different stories while keeping the subject unchanged, but can also better render physical effects on the character's body based on the physical rules of different scenes, including the direction in which clothes flutter, hair state, gait rhythm, light and shadow, etc., to make the subject and scene more harmonious.

Dive deeper into a wider range of industrial scenarios and continue to explore application values

As model capabilities improve, Seedance 2.5 is also deepening into a wider range of industrial scenarios such as education and industry. In the field of education, models have begun to enter real learning scenarios. For example, Seedance 2.5 can transform the historical background, characters, and plot behind the text into video content, and transform learning content from static explanations into more vivid and immersive images. At the same time, models can help teachers produce teaching videos more efficiently and transform abstract content such as scientific principles, historical events, and experimental processes into dynamic presentations, which not only lowers the production threshold for teaching materials, but also supports flexible content customization.

In industries such as industrial manufacturing, physical intelligence, and autonomous driving, Seedance 2.5 has also begun to enter more specific production processes. The model can generate high-quality synthetic video data to help train the robot's perception and operation ability; it can also be used for industrial simulation, process training, and equipment demonstration. In autonomous driving scenarios, models can also simulate low-frequency scenarios such as extreme weather and complex road conditions, providing more samples for system testing and training.

Summary and outlook

ByteDance said that Seedance 2.5 further enhances the ability to understand and present the real world, moving video generation from fragment-level output to a more complete creative process. We have also seen that there is still room for improvement in the physical rationality of complex movements and the stability of extremely multi-subject interaction scenarios.

In the future, the Seedance team will continue to explore longer and more coherent narratives, better understand user generation and editing experiences, and further enhance the model's understanding of the laws of the physical world. We hope that the Seedance series models will be more vivid, more controllable, and more understanding of creators, helping more users express their creativity while continuing to explore and serve a wider range of industry needs.