Shengshu launches Vidu S2 real-time video model

Shengshu Technology announced Vidu S2, the successor to Vidu S1, which generated video continuously in response to user input instead of rendering fixed clips. The company calls S2 "the world's leading" model for real-time interaction and editing. The announcement excerpt contains no benchmark figures to back that claim.

S2 has two parts. Vidu S2-Avatar handles spoken conversation and lets users direct full-body actions, including dance, by voice. It outputs at 720P and above, up from S1's 540P limit. Users can upload new reference images while the stream is running. A character can pick up and show an object, put on a referenced garment or move into a new scene. A VLM-based agent system interprets these mixed text, speech and image inputs. The model also produces pre-rendered avatar videos offline.

Vidu S2-Editing works on live video streams. It can apply the style of a reference image, change a subject's clothing, swap the person on screen for a referenced subject, or replace the background.

Shengshu says it upgraded data, architecture, training, inference and serving. To improve motion, it added solo dance footage and 2D and 3D animation to the training data. It kept clips with smooth camera movement and corrected them with a background-stabilization step. It also rewrote captions as timestamped action sequences, so the model learns the order and outcome of actions. The company also describes exploratory work on real-time spatial video.

Jintao Zhang leads the S-series work. He is a doctoral researcher advised by Professor Jun Zhu and heads streaming video generation and inference at Shengshu. An enterprise API and a technical report are available. Pricing, latency figures and evaluation results were not in the material reviewed.