AlayaVista: Streaming World Modeling from Panoramic States to Perspective Video
AlayaVista splits the job: keep the world as a panoramic latent state, then render only the camera view needed.
Starting from one perspective image, the system builds a 360-degree scene prior and evolves it under camera control. A viewport renderer turns that panoramic state into perspective video latents, while a refiner adds detail, reduces artifacts, and performs super-resolution. The authors say the design is adapted for streaming through chunk-autoregressive generation and few-step distillation. They also introduce MUGEN, a 1,318-hour real-world panoramic video dataset with 4K-or-higher footage and semantic and geometric annotations. HF Daily Papers' note
Starting from one perspective image, the system builds a 360-degree scene prior and evolves it under camera control. A viewport renderer turns that panoramic state into perspective video latents, while a refiner adds detail, reduces artifacts, and performs super-resolution. The authors say the design is adapted for streaming through chunk-autoregressive generation and few-step distillation. They also introduce MUGEN, a 1,318-hour real-world panoramic video dataset with 4K-or-higher footage and semantic and geometric annotations. HF Daily Papers' note
score 5