This dissertation addresses generative modeling as the problem of density estimation for sparse, multimodal high-dimensional data distributions. Real-world data such as images and text occupy only a small subset of the ambient space and often exhibit ...
This dissertation addresses generative modeling as the problem of density estimation for sparse, multimodal high-dimensional data distributions. Real-world data such as images and text occupy only a small subset of the ambient space and often exhibit multiple separated modes, making direct high-dimensional density estimation challenging. While diffusion-based models currently provide state-of-the-art sample quality and mode coverage, they rely on score-based training and iterative reverse-process sampling, which incurs substantial computational cost. Motivated by this limitation, we propose a two-stage generative framework that aims to achieve competitive generation quality with significantly reduced sampling cost.
In Stage 1, we train a variational autoencoder (VAE) to obtain a continuous latent representation that preserves the structural characteristics of observed data. Importantly, this stage is not intended to enforce a fixed simple prior as the final generative prior; instead, it is designed to construct a latent space that is suitable for stable decoding and meaningful sampling. In Stage 2, we explicitly estimate the latent posterior density (i.e., the aggregated posterior) using a cluster-based probabilistic density model, such as a full-covariance Gaussian mixture model (GMM). This step directly addresses the question of where and how frequently to sample in latent space, thereby modeling multimodality in a principled manner. Generation is then performed by sampling from the estimated latent density and passing the sample through the deterministic decoder once, enabling one-pass generation without iterative score-based sampling. Moreover, conditioning on mixture components naturally supports mode-wise generation and interpretability.
We evaluate the proposed approach primarily on CelebA-HQ 256×256 and report results using multiple complementary metrics, including FID, Inception Score, Density/Coverage, and CLIP-MMD. Empirically, our method achieves competitive distributional similarity compared to latent diffusion variants while substantially improving sampling efficiency due to its one-pass decoding. We further demonstrate the practicality of latent density estimation via an efficient GPU-based GMM-EM implementation, and provide qualitative analyses such as cluster-wise sample inspection and covariance/PCA-based latent manipulation, highlighting interpretability and controllability at the mode level.
In summary, this dissertation presents a structured latent-space approach to sparse, multimodal density estimation for generative modeling. By separating continuous latent representation learning from explicit latent density estimation, the proposed framework enables accurate latent distribution modeling without score-based sampling, offering a favorable trade-off between generation quality and computational cost.