Oxford, Monash, Bristol teams unveil Oneira, an interactive video world model
Teams from the University of Oxford, Monash University and the University of Bristol (among others) propose Oneira, an interactive video world model that separates "what happens in the world" from "how the picture is rendered": a Coding Agent maintains a World State Table recording each object's semantic label, 3D bounding box and interaction state, an Engine renders that table into a simplified condition video, and a video generation model fills in real appearance and motion detail. The authors also built InteractionBench, with 100 interaction cases and 50 memory cases, and compared Oneira against MiniMax-H3 Ref2VA, LingBot-World-V2, YUME 1.5 and AlayaWorld. They report Oneira leads the strongest baseline by 29, 19 and 29 percentage points on whether an interaction occurs, timing accuracy and correct target object; after the camera leaves and returns, PSNR and SSIM improve by 93% and 83% and state retention rises 18%, with VBench video quality at or near the best.
If the separate-world-state approach reproduces and spreads, generative world models could move from mere navigation and viewing toward precise, persistent interaction with specified objects.
关注AI的 2026-10-13 12:00 北京
在交互准确性和长时一致性上明显优于现有方法。
最近，来自牛津大学&莫纳什大学&布里斯托大学的团队提出了交互式视频世界模型 Oneira，将这类模型从开放式的生成走向开放世界的交互。
现有的视频世界模型已经能够随着镜头的移动不断生成新的场景，但智能体在其中通常只能导航，或进行一些简单的交互。Oneira 通过 Coding Agent 维护的 World State Table 记录场景中的物体及其状态。无论是输入图像中已有的物体，还是镜头移动时新生成的物体，Oneira 都能精确地改变它们，并让改变后的状态在后续视频中持续保留。
#Oneira#University of Oxford#Monash University#University of Bristol#World State Table#InteractionBench#MiniMax-H3 Ref2VA#LingBot-World-V2#YUME 1.5#AlayaWorld