Researchers have published a new paper, World in World, describing a training-free interface that brings flexible camera and time control to frozen autoregressive video world models. According to the paper, such models already support interactive, long-horizon exploration, but controlling them flexibly remains difficult.
The core challenge is that exploring a source video from new viewpoints requires the generated rollout to stay synchronised with the recorded event, place observed content in the requested view, plausibly complete newly exposed regions, and recover previously generated appearance on revisits. Existing approaches typically rely on task-specific modules or additional training.
World in World instead operates purely at inference time. It converts heterogeneous control evidence into camera- and time-labelled clean visual states, which are then read through the native self-attention of a frozen causal video model. The evidence includes source-video observations, target-view scene projections, geometry renderings that guide completion of newly exposed subject regions, and retrieved generated states beyond the rolling cache. Each evidence source carries token-level support and its own availability schedule.
Two mechanisms drive the interface. A correspondence router combines persistent point identities with geometry to establish token correspondences, guiding supported queries toward matching source-video tokens. Evidence-wise attention CFG (EWA) then independently regulates each auxiliary channel's additional contribution using attention responses from the same denoising forward pass.
The paper reports that the shared interface supports camera-controlled rerendering, long-horizon revisiting, and human-motion transfer with the same frozen backbone. Evaluation focuses on camera-controlled video rerendering under diverse viewpoint changes, assessing perceptual quality, temporal consistency, and camera-following accuracy.