Vidu S2 research explores live AI-generated and edited spatial video
A September 10 preprint from researchers at Tsinghua University and Shengshu Technology describes Vidu S2 and an experimental route to stereoscopic video for VR headsets. This is research reporting—not confirmation of a universally compatible consumer product.
Confirmed facts
Interactive characters and stream editing
The authors describe two models. S2-Avatar generates interactive character video and accepts new visual references during a stream; the paper reports 720p output. S2-Editing changes incoming video, including appearance, clothing, characters and backgrounds. These capabilities and comparative performance results are the researchers' findings, not independent HoloNexa tests.
How the spatial extension works
For ordinary single-view video, the proposed pipeline estimates depth and synthesizes synchronized left- and right-eye images. Existing stereoscopic footage can instead have both views edited together before being separated for display. The authors describe sending the paired output to VR headsets.
Synthesized depth is not the same as capturing a complete three-dimensional scene. The paper discusses filling newly exposed image regions and stabilizing depth over time; these steps address artifacts introduced by the conversion.
Important research limitations
The authors identify resolution and end-to-end latency as remaining deployment challenges, particularly when headset motion must match camera passthrough. Panoramic scenes that viewers can explore by turning their heads are presented as future work, not an established capability of the current fixed-view approach.
The cited source is an arXiv preprint. It does not establish peer-reviewed validation, a headset compatibility list, retail pricing or a release date for the spatial extension.
Why it matters · Analysis
Live stereoscopic generation could become useful for interactive characters and visual experiences. Practical adoption depends on comfortable image quality, consistent depth and sufficiently low latency; a promising demonstration alone does not settle those requirements.