Abstract

Generative models aim to learn the underlying data distribution and synthesize realistic samples such as images, videos, and audio. Deep learning has brought remarkable success to this field, enabling models to capture complex, high-dimensional structures and produce highly realistic outputs. However, a fundamental challenge remains — achieving both computational efficiency and high-quality generation simultaneously. Among modern approaches, diffusion-based generative models stand out for their exceptional sample fidelity but require hundreds of iterative denoising steps, making them computationally expensive and slow. To improve efficiency, recent work trains diffusion models in the low-dimensional latent space of a pretrained Variational Autoencoder (VAE), where the model generates latent representations that are later decoded into full-resolution images using a VAE decoder. This latent diffusion paradigm substantially reduces memory and compute cost; however, two major challenges remain for high-quality generation: (1) high generation quality demands high reconstruction capability of the pretrained VAE, and (2) the iterative denoising process is still slow and computationally heavy. If the number of denoising steps is reduced during generation to improve efficiency, it results in low sample fidelity. This dissertation aims to achieve high efficiency while maintaining high sample-fidelity in diffusion generative models, and addresses these challenges from two complementary directions. First, we propose a Bayesian nonparametric framework that jointly learns both the latent representations and the optimal encoder–decoder architecture of VAEs, allowing the network to automatically adapt its complexity based on the data, and improving its reconstruction performance. This framework is further extended to Graph Neural Networks (GNNs) to enable adaptive neighborhood learning, enhancing their flexibility and expressivity. Our adaptive framework has applications in biomedical domains, where we use it for biomedical interaction prediction tasks, demonstrating its effectiveness in modeling complex biological relationships. To address the second challenge, we introduce a novel viewpoint for a few-step generation in diffusion-based models, and propose an algorithm to improve the existing few-step generation models. Specifically, we provide a theoretical interpretation to distillation-based few-step generative models and based upon it propose a trajectory-level objective to distill knowledge from base model to few-step models. Training with the proposed objective significantly improves few-step generation performance compared to existing point-level distillation objectives. Together, these contributions advance the design of generative models toward a unified goal of achieving high generation quality and computational efficiency, paving the way for scalable and reliable generative modeling.

Publication Date

6-2026

Document Type

Dissertation

Student Type

Graduate

Degree Name

Computing and Information Sciences (Ph.D.)

Department, Program, or Center

Computing and Information Sciences Ph.D, Department of

College

Golisano College of Computing and Information Sciences

Advisor

Rui Li

Advisor/Committee Member

Linwei Wang

Advisor/Committee Member

Haibo Yang

Campus

RIT – Main Campus

Download

Share

COinS