The scarcity of terrestrial LiDAR (TLS) data is an opportunity for data synthesization techniques. Generative deep learning techniques such as style transfer or super resolution can be used to improve and expand existing datasets. Forestry is an excellent candidate for these techniques because of the need for higher quality subcanopy measurements. In this study we demonstrate a conditional variational autoencoder (cVAE) that applies style transfer to infill TLS fidelity voxel grids from ALS voxel grids. This generator decreases the ALS vegetation underestimation from an average of 50% to 20% of terrestrial measurements. It performs best for surface fuel and ladder fuels (of interest in fire modeling) and for trees around 15 m in height.