“Psychedelic Distortion” (2023)
Hierarchical VAE achieves fidelity to input images while enabling diverse image generation by replacing the distribution for sampling latent variables in particular layers with prior distribution in the reconstruction process.
Using Very Deep VAE (one of the hierarchical VAEs) trained on facial image dataset FFHQ, we discovered that applying partial reconstruction to slightly noised facial images leads to artistic distortions in the generated images.
Lab Link: https://www.kawa-lab.org/