<ul><li>Discrete diffusion models have shown promise for modeling complex discrete data, with masked diffusion models offering a balance between quality and generation speed.</li><li>Variational Autoencoding Discrete Diffusion (VADD) is proposed as a framework that enhances discrete diffusion by capturing correlations among dimensions using latent variable modeling.</li><li>VADD includes an auxiliary recognition model for stable training through variational lower bounds maximization and amortized inference over the training set.</li><li>Empirical results show that VADD outperforms masked diffusion model baselines in terms of sample quality, especially with fewer denoising steps.</li></ul>

Variational Autoencoding Discrete Diffusion with Enhanced Dimensional Correlations Modeling

Discover more