世界模型與影片生成
AlayaWorld v1.1 Switches to a Streaming 3D Point Cache, Raising Its Long-Horizon Video Consistency Score to 89.5
AlayaWorld has rewritten the spatial memory and conditioning pipeline of its interactive world model so that camera controls, historical frames, and generated content use aligned causal VAE representations. A new report claims the highest overall consistency score on WBench, but the updated weights and inference code have yet to be released.
Alaya Lab released the AlayaWorld v1.1 technical report on August 13. The model backbone, training data, and segment-by-segment autoregressive generation method remain unchanged; the main updates concern how conditioning signals are represented and injected. The previous version warped depth maps into spatial memory and passed camera parameters through a separate AdaLN branch. In the new version, after each video segment is generated, per-pixel 3D points are registered in a persistent cache and then re-rendered from the planned viewpoint for the next segment. Encoded by the same causal VAE, the rendered output directly serves as both geometric memory and viewpoint control.
The team also replaced static first-frame conditioning with dynamic nine-frame conditioning, aligned the historical window in pixel space, and changed the handling of memory tokens from “zeroing them out” to removing them entirely. These changes are intended to eliminate discrepancies in temporal boundaries, sequence lengths, and latent statistics between training and segment-by-segment inference. Tokens corresponding to invalid regions in the 3D render are also removed to maintain FlashAttention compatibility.
Across WBench’s 158 navigation cases, AlayaWorld achieved an average consistency score of 89.5, outperforming eight comparison systems. Its background, perspective, and geometric consistency scores reached 94.1, 86.6, and 94.1, respectively. However, it scored only 51.6 on scene setup and 65.1 on causal fidelity, indicating that retaining spatial information in generated frames does not necessarily mean the model understands environmental semantics or physical causality.
Deployers still cannot treat the paper’s reported results as directly reproducible. The public repository continues to describe the previous AdaLN and Depth-Anything-3 pipeline, while the “improved weights” remain listed as forthcoming. The training code and portions of the data have likewise not been released. The next key milestone will be whether the new weights, the ViGeo point-cache implementation, and a matching WBench evaluation configuration are released together.