I am a fourth-year Computer Science Ph.D. student at the University of Maryland, advised by Soheil Feizi. I am currently a Research Intern at Meta, where I work on autoregressive video diffusion models, and have previously interned at Adobe, Netflix, ByteDance TikTok, and Pinterest.
My research broadly focuses on understanding and interpreting image and video generation and editing models, identifying their limitations, and developing methods that provide users with more reliable and fine-grained control over model outputs. I am particularly interested in controllable image editing, compositional generation, localizing and editing knowledge in text-to-image generative models, and autoregressive video diffusion models.
My past research has also spanned Deep Learning Robustness, 3D Vision, and applications of Machine Learning to neuroscience, including improving seizure detection systems.
Email / Google Scholar / Linkedin / Github / CV
Our paper introduces StructFlow, a flow matching framework that encodes spatial locality directly into the source noise distribution, producing transport paths aligned with image structure and enabling fine-grained local editing, robust structure preservation, and smooth semantic interpolation.
Our paper introduces SliderEdit, a framework that disentangles multi-part editing instructions and exposes each as a continuous slider, enabling fine-grained, interpretable, and compositional control in instruction-based image editing.
Our paper introduces AgentComp, an agentic orchestration framework that autonomously builds compositional training data and trains text-to-image models to better distinguish compositionally similar prompts and images, resulting in stronger and more reliable compositional generation without sacrificing image quality.
Our paper presents a method to localize where knowledge is encoded within Diffusion Transformers (DiT), enabling more interpretable and efficient model editing for personalization and knowledge unlearning.
Our paper demonstrates that text-to-image generative models often fail at accurately composing attributes and relationships due to sub-optimal text conditioning by the CLIP text-encoder, and we show that significant compositional improvements can be achieved by fine-tuning a simple linear projection on CLIP's representation space.
Our approach enhances seizure detection in epilepsy patients by converting raw EEG data into informative embeddings, achieving state-of-the-art classification results with an event-based sensitivity of 100% and specificity of 99% on the CHB-MIT scalp EEG dataset.
This research offers a data-centric approach to enhance machine learning model robustness against imperceptible adversarial perturbations by detecting and removing challenging samples during training, resulting in improved adversarial training.
PhytoOracle (PO) introduces modular, scalable pipelines for processing large volumes of phenomics data, improving efficiency, enabling data fusion, and supporting multi-system trait extraction, with broad applicability across species and various data sources, including drone data.
Our study identifies vulnerabilities in current out-of-distribution (OOD) detection methods and presents the Adversarially Trained Discriminator (ATD) as a robust alternative with improved performance.