1: \begin{abstract}

2: 

3:   The advent of generative radiance fields

4:   has significantly promoted the development of 3D-aware image synthesis.

5:   %

6:   The cumulative rendering process in radiance fields makes training these generative models much easier since gradients are distributed over the entire volume, but leads to diffused object surfaces.

7:   %

8:   In the meantime,

9:   compared to radiance fields occupancy representations could inherently ensure deterministic surfaces.

10:   However, if we directly apply occupancy representations to generative models,

11:   during training they will only receive sparse gradients located on object surfaces

12:   and eventually suffer from the convergence problem.

13:   %

14:   In this paper, we propose \textbf{Generative Occupancy Fields (GOF)},

15:   a novel model based on generative radiance fields that can learn compact object surfaces without impeding its training convergence.

16:   %

17:   The key insight of GOF is a dedicated transition from the cumulative rendering in radiance fields to rendering with only the surface points

18:   as the learned surface gets more and more accurate.

19:   %

20:   In this way, GOF combines the merits of two representations in a unified framework.

21:   %

22:   In practice,

23:   the training-time transition of \emph{start from radiance fields and march to occupancy representations} is achieved in GOF

24:   by gradually shrinking the sampling region in its rendering process

25:   from the entire volume to a minimal neighboring region around the surface.

26:   %

27:   Through comprehensive experiments on multiple datasets,

28:   we demonstrate that GOF can synthesize high-quality images with 3D consistency

29:   and simultaneously learn compact and smooth object surfaces.

30:   Our code is available at \url{https://github.com/SheldonTsui/GOF_NeurIPS2021}.

31:   

32: \end{abstract}

33: