World in World: Explore the World with World Models
A frozen causal video model is given new camera and timing evidence at inference time, without extra training.
World in World turns source-video observations, projected target views, geometry renderings, and retrieved past generated states into labelled visual evidence the model can attend to. The paper says this helps keep new-view rollouts synchronized with the recorded event while filling newly exposed regions and restoring prior appearance on revisits. The same interface is used for camera-controlled rerendering, long-horizon revisiting, and human-motion transfer. Evaluations focus on rerendering quality, temporal consistency, and camera-following accuracy. HF Daily Papers' note
World in World turns source-video observations, projected target views, geometry renderings, and retrieved past generated states into labelled visual evidence the model can attend to. The paper says this helps keep new-view rollouts synchronized with the recorded event while filling newly exposed regions and restoring prior appearance on revisits. The same interface is used for camera-controlled rerendering, long-horizon revisiting, and human-motion transfer. Evaluations focus on rerendering quality, temporal consistency, and camera-following accuracy. HF Daily Papers' note
score 5