AlayaWorld: Interactive Long-Horizon World Modeling - Full Technical Report (v1.1)
AlayaWorld v1.1 keeps the same core model, but rebuilds how it feeds visual and spatial conditions into long video generation.
The report says the architecture, chunk-wise autoregressive scheme, and training data are unchanged from the earlier release. The revision replaces depth-warped spatial memory with a streaming 3D point-cache renderer. It also moves visual conditions into the same causal-VAE latent space as generated video, with matching temporal structure. The listed changes include motion-aware latent conditioning, causal encoding of re-rendered memory, hard memory dropout, a unified VAE protocol, and removing the camera AdaLN branch so viewpoint control comes through the spatial condition. ArXiv · AI/CL/LG's note
The report says the architecture, chunk-wise autoregressive scheme, and training data are unchanged from the earlier release. The revision replaces depth-warped spatial memory with a streaming 3D point-cache renderer. It also moves visual conditions into the same causal-VAE latent space as generated video, with matching temporal structure. The listed changes include motion-aware latent conditioning, causal encoding of re-rendered memory, hard memory dropout, a unified VAE protocol, and removing the camera AdaLN branch so viewpoint control comes through the spatial condition. ArXiv · AI/CL/LG's note
score 6