Vidu S1: the real-time interactive video model, from a companion operator's view

Almost every video model you have read about here does the same job: take a prompt or an image, render one clip of a fixed length, done. Vidu S1 does a different job. It generates video autoregressively, frame after frame, in response to voice input, so the session has no fixed end. You speak, the avatar reacts, you keep speaking, it keeps going. ShengShu frames this as a move from single-clip generation to continuous real-time interaction, and that framing is fair.

The control loop is the interesting part. According to ShengShu, the model interprets the semantic meaning and emotional context of what you say and drives facial expressions, gestures, and full-body motion from that, rather than playing back preset animations. In principle that means the avatar responds to intent, not just to keywords. We have not verified how well that holds up in a real conversation, and we flag that below.

It is worth being clear that this is the clearest example so far of a broader trend toward interactive, streamable video. Runway has signalled the same direction with its world-model work. Vidu S1 is the one you can try today, which is why it is the one worth writing up.

The core flow is short: upload a single image, pick or supply a voice, and you have an interactive character. The source can be a photograph of a real person, an illustrated or anime character, or a pet, and the avatar it produces will move its whole body, change expression, and gesture as it talks.

The intended uses cluster around presence: a companion you can hold a live conversation with, a virtual host or presenter, a character for a livestream, a talking version of an existing brand mascot or fictional character. Because the session is unbounded, it is aimed at sustained interaction rather than at producing a finished, exportable film.

Two routes exist right now. There is a public demo where you can try the experience directly, and an API beta for developers who want to build on top of it. Which one fits depends on whether you want to feel it or ship it.

This is not a high-resolution model, and it is not trying to be. Output is 540p, 960 by 540, at 25 frames per second and up to 42. Next to the 1080p and 4K single-clip models, that is visibly lower resolution, and if your goal is a crisp hero shot for a landing page, this is the wrong tool.

But resolution is the wrong axis to judge it on. The job here is latency and continuity: keeping a live avatar responsive in real time is a much harder engineering problem than rendering one clip slowly at high quality, and 540p is the price that buys real-time interaction on hardware people own. Judge it as a live presence, not as a clip generator, and the number reads differently.

The consumer-GPU point is the one we would not gloss over. ShengShu says S1 runs on consumer-grade GPUs, which puts a live interactive avatar within reach of a local setup rather than a data-centre rental. If you care about running things yourself, that matters, and it is the same instinct we cover in our Z-Image guide for local, no-account image generation.

For an AI companion product, the gap has never really been text. It has been presence: a face that reacts while you talk, in real time, without a render queue in the middle. A model that turns one image and a voice into a live, voice-steered avatar is aiming straight at that gap, which is why we are paying attention rather than filing it under video-gen novelty.

The honest caveat from our side of the fence is that a companion is conversation and presence, not spectacle. What decides whether something like this is usable is not the demo reel, it is whether the reactions feel timed and intentional across a long session, whether the latency stays low, and whether the emotional read holds when the input is messy human speech. Those are exactly the things a launch announcement cannot tell you, and that we have not tested.

There is also the plain fact that live interactive video is heavier to run than a chat model. The consumer-GPU claim is encouraging, but a public demo under load and a production deployment serving many concurrent sessions are different animals. We would price and stress-test that carefully before building on it.