Unsupervised Controllable Generation of Diffusion Models with Latent Variables in VAEs

Minju Kim, Seonggyeom Kim, Dong-Kyu Chae

Abstract: This study introduces a method for controlling image generation in Diffusion Models using the disentangled latent variables of Beta-VAE and Factor-VAE, variations of the Variational Autoencoder. By integrating these disentangled latent variables into the well-known Denoising Diffusion Probabilistic Models (DDPM), the proposed method enhances image generation both qualitatively and quantitatively compared to the existing VAE variations. Furthermore, it allows for adjusting the latent variables, providing a novel way of manipulating image output in diffusion models. This approach is versatile, applicable to various existing disentanglement VAEs, and offers a new direction for unsupervised control in image generation.