Alaya-EVOKE presents a method for generating long interactive worlds with linear scaling

A paper titled Alaya-EVOKE has been published on arXiv, in which the authors tackle the task of generating long interactive worlds. They move the scene state to an external bank indexed by camera and rebuild the teacher for long-horizon supervision: sparse attention with chunk-based grouping, sampling of distant frames, and linear global state. This yields linear growth in memory and compute as session length increases. The authors have not released the code. More details — on the paper's page.

What it means

The work pertains to interactive world generation, not image-to-3D generation. There is no direct overlap with the tools in our reference: OpenLRM, Hunyuan3D 2.1, and TRELLIS solve the image-to-3d task — creating a static model from a single image, without sessions, world memory, or interactive scene continuation. That said, the method touches on a problem common to generative 3D systems: maintaining context as the generation horizon grows. None of our tool fact sheets describe such mechanisms — none of them mentions an external scene state store or long-horizon supervision. For OpenLRM and TRELLIS, context is limited to a single input image; for Hunyuan3D 2.1, to an image and a text description. For now, this is a research work without code, and nothing changes for users of our tools. If the approach gets an open-source implementation, it could influence future world and video generators, but not current image-to-3d pipelines.