Advised by Prof. Mingyuan Zhou, Department of Information, Risk, and Operations Management (IROM), McCombs School of Business.

Designed an end-to-end data curation and preprocessing pipeline for video super-resolution (VSR), including a temporally consistent simulator of real-world video degradation. Evaluated existing VSR baselines and developed a diffusion model based on CogVideoX 1.5.

Designed an automatic short-form video localization pipeline that adapts textual and graphical elements per locale using SOTA detection, OCR, inpainting, and segmentation models. Built a complete landscape-to-portrait conversion workflow with KMeans + Gaussian-process smoothing for stable subject tracking.

Identified four major limitations of existing visual in-context learning methods and proposed iPromptDiff — an SD-based architecture that decouples content vs. task and routes visual perception through text embeddings, beating baselines in-domain and OOD even when text prompts are missing.

Studied non-Gaussian iterative corruption / recovery from a Bregman-divergence perspective. Proposed binomial / Poisson hierarchical VAE frameworks for sparse, skewed, heavy-tailed data. Built and open-sourced a PyTorch codebase covering DDPM, DDIM, and classifier-free guidance — 230★+ on GitHub.

Introduced an information-theoretic surrogate loss (ASK) for DkNN classification with matching attack and defense algorithms achieving SOTA results. Co-developed an immune-inspired adversarial framework (RAILS) capable of defending against unseen attacks.