Many deep generative models have been proposed to reconstruct pseudo-healthy images for anomaly detection. Among these models, the variational autoencoder (VAE) has emerged as both simple and efficient. While significant progress has been made in refining the VAE within the field of computer vision, these advancements have not been extensively applied to medical imaging applications. We present a benchmark that assesses the ability of multiple VAEs to reconstruct pseudo-healthy neuroimages for anomaly detection in the context of dementia. We first propose a rigorous methodology to define the optimal architecture of the vanilla VAE and select the best hyper-parameters of the VAE variants. Relying on a simulation-based evaluation framework, we thoroughly assess the ability of 20 VAE models to reconstruct pseudo-healthy images for the detection of dementia-related anomalies in 3D brain FDG PET and compare their performance. This benchmark demonstrated that the majority of the VAE models tested were able to reconstruct images of good quality and generate healthy looking images from simulated images presenting anomalies. Even if no model clearly outperformed all the others, the benchmark allowed identifying a few models that perform slightly better than the vanilla VAE. It further showed that many VAE-based models can generalize to the detection of anomalies of various intensities, shapes and locations in 3D brain FDG PET.