Vidu S2 Operational

Vidu S2 is a real-time video model developed by ShengShu Technology. Available to try upon release, it includes two core models. S2-Editing supports real-time style changes, outfit swaps, character replacement, and background replacement for video streams. Frame alignment technology prevents visual artifacts during high-speed motion. S2-Avatar enables real-time interaction at 720P and 25-42FPS, allowing reference images and outfit changes to be inserted at any time, with full-body motion tracking. It achieves the best results in all nine categories of StreamAV-Bench. Its original SRF training eliminates drift in long-form generation, while its Spatial video exploration targets VR applications.

S2-Editing real-time editing: Change the style of a playing video stream, enable virtual try-ons, replace characters, or change backgrounds in real time with just one text instruction and an optional reference image.
S2-Avatar real-time interaction: Generate high-definition characters in real time at 720P and 25-42FPS. Insert new reference objects at any time and have characters "catch" them while maintaining state memory.
One-click real-time outfit changes: Upload one image and instantly change a character's outfit without interrupting the video stream, while preserving fluid motion and emotional continuity.
Full-body motion tracking: Supports complete body performances based on instructions such as solo dance and 2D/3D animation, overcoming the limitations of traditional "talking head" digital humans.
Vidu S2-Spatial video: Converts monocular real-time video into synchronized left- and right-eye views, providing a spatial real-time experience for immersive devices such as VR.

Frame-aligned sparse attention: The core technology behind S2-Editing. During real-time editing, every frame of the new video is precisely locked to and aligned with the same moment in the original video, ensuring that motion trajectories remain intact and frames do not tear after outfit or character changes. This solves the stability challenges of real-time image editing in high-speed video.
SRF (Self-Replay Forcing) training: Developed to eliminate "crashes" and "drift" in long-form streaming generation, this technique makes the model confront its own generation history and self-correct errors. It effectively gives the model "self-correcting memory," allowing it to perform consistently in continuously changing scenes.
Dynamic reference mechanism: S2-Avatar lets users add new reference images at any point during video generation. The model responds instantly and integrates new elements into the scene while retaining state information, enabling coherent interactions with clear cause-and-effect continuity.
Spatial video conversion: An algorithm generates synchronized left- and right-eye views from monocular real-time video. By simulating binocular parallax with a single camera input, it preserves spatial consistency along with temporal continuity, paving the way for real-time VR experiences.

Web experience: Visit https://vidu.com/vidu-stream to try it in real time.
API integration: Developers can consult the documentation at platform.vidu.com/vidu-stream/doc and integrate S2 capabilities.
Real-time editing: Enter a text instruction and upload an optional reference image while a video stream is playing to change its style, outfit, character, or background.
Real-time interaction: Talk with S2-Avatar to drive character performances, and upload images at any time to show new objects or change outfits with one click.

Truly real time: 720P, 25-42FPS streaming output, with editing and interaction that do not interrupt the video stream.
Comprehensive editing: Four major real-time editing directions, including style, clothing, characters, and backgrounds, cover key commercial scenarios such as e-commerce, livestreaming, and marketing.
Dynamic references and state memory: Insert new elements at any time and let characters "remember" the context, enabling coherent, engaging interactions beyond the limitations of "talking head" digital humans.
Strong long-form stability: SRF self-correction training eliminates crashes and drift in streaming generation, allowing characters to perform consistently for extended periods.
Leading benchmark performance: S2-Avatar ranks first in all nine StreamAV-Bench categories, while Editing leads in key dimensions including instruction following, appearance fidelity, and motion preservation.
Low barrier, high returns: Available to try upon release, it can turn one original video into multiple stylized assets, significantly reducing set design, filming, and post-production costs.

E-commerce livestreaming: When fans request perfumes or clothing, characters can present the products in real time or change outfits with one click, recreating an authentic shopping-host experience.
Virtual try-on: Consumers can see how clothing follows body movements before purchasing, reducing decision costs and return rates.
IP commercialization: Brand virtual idols and game protagonists can "drop into" livestreams and sell products in place of real hosts, maximizing fan immersion.
Cross-border e-commerce and scenario marketing: Replace backgrounds with one click, switching from selling outdoor jackets in an instant to snowy mountains, or placing cosmetics in a Paris fashion show. Set-design costs approach zero.
Scaled asset production: Generate cyberpunk, ink-wash, anime, and other stylistic assets in bulk from one live-action original, supporting overseas short dramas, game user acquisition, and major brand campaigns.