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

  • 2022-07-22 16:41:06
  • Zhengqi Li, Qianqian Wang, Noah Snavely, Angjoo Kanazawa
  • 58

Abstract

We present a method for learning to generate unbounded flythrough videos ofnatural scenes starting from a single view, where this capability is learnedfrom a collection of single photographs, without requiring camera poses or evenmultiple views of each scene. To achieve this, we propose a novelself-supervised view generation training paradigm, where we sample andrendering virtual camera trajectories, including cyclic ones, allowing ourmodel to learn stable view generation from a collection of single views. Attest time, despite never seeing a video during training, our approach can takea single image and generate long camera trajectories comprised of hundreds ofnew views with realistic and diverse content. We compare our approach withrecent state-of-the-art supervised view generation methods that require posedmulti-view videos and demonstrate superior performance and synthesis quality.

 

Quick Read (beta)

loading the full paper ...