InfiniteNature-Zero: Learning Perpetual View Generation of Natural Scenes from Single Images

We present a method for learning to generate unbounded flythrough videos of natural scenes starting from single view. This capability is learned collection photographs, without requiring camera poses or even multiple views each scene. To achieve this, we propose novel self-supervised view generation training paradigm where sample and render virtual trajectories, including cyclic paths, allowing our model learn stable views. At test time, despite never having seen video, approach can take image long trajectories comprised hundreds new with realistic diverse content. compare recent state-of-the-art supervised methods that require posed multi-view demonstrate superior performance synthesis quality. Our project webpage, video results, at https://infinite-nature-zero.github.io .