Diffusion Models

Our ongoing research explores diffusion models across computer vision, medical imaging, and computational biology. First, we investigate methods to improve diffusion models for high-fidelity image generation, with particular emphasis on better capturing data distributions, semantic structure, and sample quality. Second, we develop diffusion-based historical image colorization techniques to reconstruct plausible and visually consistent colors in archival grayscale photographs while preserving important structural and contextual information. Third, we study diffusion-based medical image augmentation, including the generation of realistic synthetic medical images to increase data diversity and improve the robustness and generalization of downstream diagnostic models. Finally, we investigate diffusion-based gene expression imputation, where generative models are used to estimate missing or sparsely observed gene-expression values while preserving biologically meaningful relationships. Collectively, these projects aim to advance diffusion modeling both methodologically and through impactful applications in heritage preservation, healthcare, and computational biology.