Shengshu releases the Vidu S1 real-time interactive video model

Background

Most video generation models work offline: a user gives a text prompt or an image, waits while the model computes, and receives a short clip. That suits making footage, but not a conversation, because a conversation needs the picture to change while the other person is still speaking. Making an avatar talk is a separate line of work: conventionally it starts with a 3D model, then rigging, then driving, and each character is handled on its own, which is slow and expensive. Vidu S1 is presented as a way to fold both into one step.

The release

On 3 July 2026 Shengshu announced Vidu S1 in a keynote given by founder Zhu Jun^[1]. The company framed the release as a shift from generating one clip offline to interacting continuously^[1]. The accompanying paper appeared on arXiv the same day under the number 2607.03118^[2]. At the same event, the Beijing Software and Information Service Industry Association named Shengshu a model enterprise for new models and new applications^[1].

What it does in use

The setting is a video call: the model takes voice input and changes the character's expression, gesture and body movement in response^[1]. What drives the picture is the speech itself rather than a pre-written script or frame-by-frame adjustment. It runs at 540p and 25 frames per second, with a peak of 42 frames per second^[1]. 540p sits below the 1080p of common high-definition video, and 25 frames per second is close to the frame rate of ordinary video playback; the numbers point to a picture meant to keep flowing rather than to be perfect in any single frame.

Creating a character is meant to be cheap: a single uploaded image is enough, with no modelling, rigging or separate training^[1]. Shengshu also says the picture stays stable even after hours of continuous generation^[1]. Drift and distortion over long runs are a familiar difficulty for generative models, so stability matters here as much as resolution.

Why it matters

In the company's framing, the change is from making a single offline clip to dealing with the person on screen in real time^[1]. If that holds, these models move from supplying material for post-production into settings that need an immediate reaction, where the output is a conversation rather than a render. It also lowers the entry cost for a character: an image, rather than a 3D asset and a rig, becomes the starting point^[1].

Vidu S1 therefore belongs to a line of work that treats video generation as an interface rather than a renderer — a model you talk to, running at conversational rates, rather than one you submit a job to and collect from later^[1].