Diffusion bridges vector quantized Variational AutoEncoders
This work addresses the problem of slow and complex prior training in VQ-VAE models for generative tasks, offering an incremental improvement by enabling end-to-end training.
The paper tackles the slow generation and complex prior training in Vector Quantized-Variational AutoEncoders (VQ-VAE) by proposing a model that trains the prior and encoder/decoder simultaneously using a diffusion bridge, achieving competitive results on mini-Imagenet and CIFAR datasets with efficient optimization and sampling.
Vector Quantized-Variational AutoEncoders (VQ-VAE) are generative models based on discrete latent representations of the data, where inputs are mapped to a finite set of learned embeddings.To generate new samples, an autoregressive prior distribution over the discrete states must be trained separately. This prior is generally very complex and leads to slow generation. In this work, we propose a new model to train the prior and the encoder/decoder networks simultaneously. We build a diffusion bridge between a continuous coded vector and a non-informative prior distribution. The latent discrete states are then given as random functions of these continuous vectors. We show that our model is competitive with the autoregressive prior on the mini-Imagenet and CIFAR dataset and is efficient in both optimization and sampling. Our framework also extends the standard VQ-VAE and enables end-to-end training.