Generative Occupancy Fields for 3D Surface-Aware Image Synthesis (Supplementary Material) Here we provide implementation details, additional results on CARLA dataset, and proof of the

Our models are trained on 8 TIT AN XP GPUs on all datasets. CelebA and Cats takes about 26 hours, 66 hours and 12 hours respectively. Table 2: The setting of several hyperparameters to be adjusted during training.Training Stage (iterations) BFM CelebA Cats batch size resolution lr ( G By contrast, our method can not only synthesize realistic images but also learn good shapes. In the experiments, we discover the trade-off between the FID score and shapes. However, the corresponding shapes will degenerate under this circumstance. It's also an interesting problem to be Our work aims at generating images in a 3D consistent manner and simultaneously learn compact and smooth object surfaces.