Bang-Dang Pham

Ph.D. Student @ UW-Madison  ยท  Working on |

I am a second-year CS Ph.D. student at the University of Wisconsin–Madison. Previously, I was an AI Research Resident supervised by Prof. Minh Hoai, Prof. Cuong Pham, and Dr. Tuan Anh Tran at Qualcomm AI Research (formerly VinAI Research). In Summer 2025, I interned at Sony R&D Lab, working on gallery-based image restoration.

I received my B.Sc. in Computer Science from Ho Chi Minh City University of Science (HCMUS) under the supervision of Dr. Ngoc-Thao Nguyen and Prof. Minh-Triet Tran.

My research interests lie in Generative Models, particularly Video diffusion models, 3D/4D Generative frameworks, and Image Restoration.

News

Feb 2026 🏆 BluRef accepted to CVPR 2026 - See you guys in Denver, CO 🇺🇸
May 2025 💼 I'm thrilled to join Sony R&D Lab as a Research Intern, where I'll be working on gallery-based image restoration.
Aug 2024 🎓 Excited to begin my PhD journey at University of Wisconsin–Madison.
Feb 2024 🏆 Blur2Blur accepted to CVPR 2024 - See you guys in Seattle, WA 🇺🇸
Feb 2023 🏆 HyperCUT accepted to CVPR 2023 - See you guys in Vancouver, BC 🇨🇦
Dec 2022 🏅 Received Excellent B.Sc. in Computer Science (top 1%, GPA: 3.9/4.0) from HCMUS.
Jul 2022 🔬 Joined Qualcomm AI Research as a Research Resident.

Experience

Applied Scientist Intern  •  Amazon Incoming
Working on World modeling and Video generation.
Mentor: Liu He and Huidong Liu
May 2026 – Present
Research Intern  •  Sony R&D
Developed methods for Gallery-based image restoration, with a focus on reference-based pipelines and real-world deployment.
May 2025 – Aug 2025
AI Research Resident  •  Qualcomm AI Research (formerly VinAI Research)
Worked on Unsupervised Image Deblurring and Generative Models.
Jul 2022 – Aug 2024

Highlighted Research

BluRef teaser
Bang-Dang Pham, Anh Tran, Cuong Pham, Minh Hoai
CVPR 2026 arXiv

Introduces a novel unpaired reference-based deblurring framework for camera-specific image restoration, eliminating the need for controlled data collection or paired supervision. The method leverages an iterative dense-matching training strategy to construct high-quality pseudo ground truth, enabling robust generalization across diverse real-world blur scenarios.

Blur2Blur pipeline
CVPR 2024

Introduces a novel framework to train camera-specific image deblurring algorithms by transforming challenging real blurry images into known blur kernels using only unpaired data, simplifying the deblurring process and demonstrating superior performance on benchmarks.

HyperCUT teaser

Tackles image-to-video deblurring, resolving sequence order ambiguity using a self-supervised method. Also introduces a diverse real blur-to-video image dataset (RB2V) covering domains like faces, hands, and streets.

Awards and Honors

Reviewer Activities