We present a novel view synthesis method based upon latent voxel embeddings of an object, which encode both shape and appearance information and are learned without explicit 3D occupancy supervision. Our method uses an encoder-decoder architecture to learn such deep volumetric representations from a set of images taken at multiple viewpoints. Compared with DeepVoxels, our DeepVoxels++ applies a series of enhancements: a) a patch-based image feature extraction and neural rendering scheme that halves the 2D U-Net parameter numbers, and enables neural rendering at high resolution; b) learned view-dependent feature transformation kernels to explicitly model perspective transformations induced by viewpoint changes; c) a recurrent-concurrent aggregation technique to alleviate single-view update bias of the voxel embeddings recurrent learning process. Combined with d) a simple yet effective implementation trick of frustum representation sufficient sampling, we achieve significantly improved visual quality over the prior deep voxelbased methods (33% SSIM error reduction and 22% PSNR improvement) on 360◦ novel-view synthesis benchmarks of diffuse objects.