Vidu S2

Two models, not one

Vidu S2 is not a single model. The paper, posted to arXiv on 10 September 2026, presents two: S2-Avatar and S2-Editing.^[1] The split follows the two things people most often want from a live video system — a character on screen that responds to them in the moment, and an existing video whose appearance can be changed as it plays. Both are described as real-time, and both continue the line that began with Vidu S1.

From S1 to S2

The project's repository lists two references: the S1 paper (2607.03118) and the S2 paper (2609.11638).^[2] The distance between them shows up most plainly in resolution. Vidu S1's avatar model produced 540p video; S2-Avatar produces 720p in real time.^[1] Higher resolution in a streaming setting matters because the output is meant to be watched live, not reviewed afterwards and re-rendered.

S2-Avatar: a character that keeps up

S2-Avatar generates video of a digital character at 720p in real time.^[1] Two capabilities separate it from a conventional talking-head generator. The first is that it accepts new reference images mid-stream, so a character or a scene can be updated during a session without stopping and starting over.^[1] The second is instruction following, which the authors describe as stronger than in S1 and which now covers actions such as dancing.^[1] In practice this means the session is not limited to a face and a voice: the user can direct what the character does, and can change what the character looks like, while the stream is running.

S2-Editing: transforming video as it arrives

S2-Editing works on video that already exists. It transforms an incoming video stream and covers four kinds of operation: style rendering, clothing replacement, character replacement and background replacement.^[1] The common thread is that the underlying motion and timing come from the source footage, while appearance is supplied by the model. Because the input is a stream rather than a finished clip, the effect is available while the video is being watched or transmitted, rather than after a separate rendering pass.

Stereo output

Both models also explore spatial video generation, producing stereo views intended for immersive displays.^[1] That means the same generation process can supply the slightly different images each eye needs, instead of leaving stereo to a conversion step downstream.

Results and availability

The authors report that S2 outperforms all baselines in their experiments.^[1] That claim is theirs and is stated for their own evaluation setup. For anyone who wants to judge the behaviour directly, a playable demo runs at vidu.com/vidu-stream.^[1]

Where it sits

The design choice worth noticing is the division of labour. Rather than stretching one model across every task, S2 separates the generation of an interactive character from the editing of an existing stream, and then gives both a shared interest in spatial output. Read against S1, the paper is a step from producing short video clips towards running a continuous video session that can be redirected while it is under way — the difference between rendering a result and holding a conversation with one.