Some forms of novel visual media enable the viewer to explore a 3D scene from\narbitrary viewpoints, by interpolating between a discrete set of original\nviews. Compared to 2D imagery, these types of applications require much larger\namounts of storage space, which we seek to reduce. Existing approaches for\ncompressing 3D scenes are based on a separation of compression and rendering:\neach of the original views is compressed using traditional 2D image formats;\nthe receiver decompresses the views and then performs the rendering. We unify\nthese steps by directly compressing an implicit representation of the scene, a\nfunction that maps spatial coordinates to a radiance vector field, which can\nthen be queried to render arbitrary viewpoints. The function is implemented as\na neural network and jointly trained for reconstruction as well as\ncompressibility, in an end-to-end manner, with the use of an entropy penalty on\nthe parameters. Our method significantly outperforms a state-of-the-art\nconventional approach for scene compression, achieving simultaneously higher\nquality reconstructions and lower bitrates. Furthermore, we show that the\nperformance at lower bitrates can be improved by jointly representing multiple\nscenes using a soft form of parameter sharing.\n