Vidu S1 claims 42 FPS voice-controlled video on consumer GPUs

A new arxiv preprint from the Vidu S1 authors drops a claim that, if it survives independent testing, matters for anyone building interactive AI media. Vidu S1 is described as a real-time interactive video generation model that takes voice commands as its steering input, outputs 540p at up to 42 FPS, and does it on what the abstract calls regular consumer GPUs. The abstract also says the system supports infinite-length real-time generation without blurring, drift, or visual distortion, which is the piece that would be most striking if it holds up outside the authors' own bench.

The user flow, per the paper, is that you upload a custom image of a real person, an anime character, or a pet, pick a voice tone, and then drive the resulting character in real time through voice instructions. Under the hood the authors credit two systems they call TurboDiffusion and TurboServe. The abstract does not spell out which consumer GPU class they mean, what the test metrics actually are, or which competing systems Vidu S1 was measured against. The claim is simply that it achieves the best performance across all test metrics while meeting real-time inference requirements.

Why any of this is interesting: real-time, coherent, long-horizon video generation has been one of the harder unsolved problems in the space, and the usual workaround has been to stitch short clips together with the seams showing. If a model can genuinely run infinite-length at 42 FPS on hardware a solo creator can afford, the shape of the interactive-media stack changes, because live AI characters stop being a datacenter feature and become something a small studio or a streamer could ship.

The honest caveat is that this is an abstract-stage paper making an unusually broad marketing-flavored claim, the specific hardware class is not disclosed, and there is no independent third-party evaluation yet. The custom-image upload path for real people also opens the impersonation and deepfake question that any live-avatar system now has to answer, and the paper does not.

What is worth watching is the playable online demo the authors point to. That is the fastest way outsiders will get to test whether the infinite-length, no-drift claim survives contact with adversarial prompts and long sessions, which is where prior long-video systems have quietly fallen over.