This is the code for paper "Likelihood-Based Generative Radiance Field with Latent Space Energy-Based Model for 3D-Aware Disentangled Image Representation".
The Carla_MCMC_with_pose is for MCMC inference with ground truth pose, Carla_amortize_inference is for amortized inference with ground truth pose, Carla_amortized_without_pose is for armortized inference without ground truth pose.
Some codes adapted from this awesome pytorch implementation of NeRF
For implementing the experiments for training without ground truth camera pose, we use this fabulous implementation of S-VAE.