Learning a generative model from partial data (data with missingness) is a\nchallenging area of machine learning research. We study a specific\nimplementation of the Auto-Encoding Variational Bayes (AEVB) algorithm, named\nin this paper as a Variational Auto-Decoder (VAD). VAD is a generic framework\nwhich uses Variational Bayes and Markov Chain Monte Carlo (MCMC) methods to\nlearn a generative model from partial data. The main distinction between VAD\nand Variational Auto-Encoder (VAE) is the encoder component, as VAD does not\nhave one. Using a proposed efficient inference method from a multivariate\nGaussian approximate posterior, VAD models allow inference to be performed via\nsimple gradient ascent rather than MCMC sampling from a probabilistic decoder.\nThis technique reduces the inference computational cost, allows for using more\ncomplex optimization techniques during latent space inference (which are shown\nto be crucial due to a high degree of freedom in the VAD latent space), and\nkeeps the framework simple to implement. Through extensive experiments over\nseveral datasets and different missing ratios, we show that encoders cannot\nefficiently marginalize the input volatility caused by imputed missing values.\nWe study multimodal datasets in this paper, which is a particular area of\nimpact for VAD models.\n