Vidu S1

Vidu S1 is a video generation model built to run live rather than to render clips offline. It generates video in real time and accepts spoken instructions as its control signal, so a person can talk to it while it keeps producing frames^[1]. Shengshu released the model on 3 July 2026 at the AI forum of the Global Digital Economy Conference in Beijing^[2].

A character from a single image

To start a session, a user uploads one image and the system turns it into an interactive character. No per-character modelling or fine-tuning is involved, so the same pipeline can be pointed at any face or figure a user brings^[2]. From there the model does not stop at a fixed clip length: according to the company it keeps generating indefinitely, without the blurring or drift that normally accumulates in long rollouts^[1]. Output is 540p video at up to 42 frames per second, produced on ordinary consumer GPUs^[1]. Shengshu describes S1 as the first generative model to support interaction over unlimited video length^[2].

Autoregressive diffusion

The company calls the technical route autoregressive diffusion: each new segment is predicted from the frames already generated, together with the current voice command and the running dialogue context^[2]. Both halves of the name describe familiar ideas. An autoregressive model produces a sequence one piece at a time and feeds each finished piece back in as input for the next step, which is how language models write text word by word. A diffusion model generates by starting from noise and removing it step by step until a clean sample appears. Put together, the model denoises the next stretch of video while conditioning on everything it has already produced and on what the user has just said. Because the condition includes the past frames and the live instruction, the same mechanism that makes the picture continuous also makes it answer to the voice.

Where the speed comes from

Two Shengshu components carry the system: the TurboDiffusion inference framework and the TurboServe serving engine^[1][2]. The acceleration itself comes from several techniques: few-step generation, low-bit SageAttention, a sparse attention scheme abbreviated SLA, and SpargeAttention^[2]. Few-step generation means reaching a usable frame in a handful of denoising passes instead of the dozens a standard diffusion model would need. Sparse attention, in general, means skipping the parts of the attention computation whose weights are negligible and calculating only the relations that matter, which saves computation while leaving the visible result largely unchanged.

Where the company aims to put it

Shengshu positions S1 for settings where someone is on the other side of the video in real time: video calls, interactive livestreaming, game NPCs, brand avatars, customer service and extended reality (XR)^[2]. In each of these the model is not delivering a finished film but standing in for a participant who has to answer immediately and keep answering for as long as the session lasts.

Where it sits

Most video generation models are aimed at a bounded output: a prompt goes in and a clip of fixed length comes out. S1's departure is the combination of live speed, voice control and open-ended duration, which turns generation from a rendering job into something closer to a conversation. The release paper lists Jintao Zhang, Kai Jiang, Jintao Chen and Xu Wang among the first authors^[1].