Duc Vu

πŸ‘‹ Hi! I'm Duc. Nice to meet y'all! I am a first-year PhD student at Rutgers University under the supervision of Distinguished Prof. Dimitris N. Metaxas. I was a predoctoral researcher at Qualcomm AI Research πŸ“‘, where I am fortunate to be supervised by Dr. Anh Tran. Prior to this, I received my B.S. in Data Science & Statistics and B.A. in Economics from πŸŽ“ Miami University - Oxford in the U.S. (2024).

πŸ”  Open to Summer 2027 Research Internships or research collaborations. Let’s create impactful work together!  πŸš€

Research

My research focuses on video generation and editing and post-training for image/video models. I am particularly interested in modeling physical dynamics and object interactions to generate realistic, controllable videos. My previous work explored diffusion distillation and efficient one-step generation for applications such as video enhancement and image inpainting. Across these areas, I aim to improve generation quality and controllability while reducing computational cost, making generative AI more scalable, efficient, and accessible.

Physics-Aware Video Generation/Editing Post-Training One/Few-Step Diffusion Image/Video Inpainting

Selected Publications

* denotes equal contribution

Cross-Space Distillation Preview
Cross-Space Distillation: Teaching One-Step Students with Modern Diffusion Teachers.
ECCV 2026

We formalize Cross-Space Distillation and introduce Bridge, a lightweight latent-space interface that makes standard one-step distillation possible across mismatched resolutions, VAEs, architectures, and diffusion/flow paradigms.

Anti-I2V: Safeguarding your photos from malicious image-to-video generation.

A memory-efficient adversarial attack against diverse image-to-video diffusion models, enabled by robust dual-space perturbation optimization.

InverFill: One-Step Inversion for Enhanced Few-Step Diffusion Inpainting.

A highly efficient one-step inversion diffusion network for high-quality few-step image inpainting.

Improved Training Technique for Shortcut Models.
NeurIPS 2025

iSM resolves five major shortcut-model flaws with dynamic guidance, wavelet loss, sOT, and Twin EMA, yielding markedly better image generation.

SwiftBrush v2: Make Your One-step Diffusion Model Better Than Its Teacher.

An improved SwiftBrush version that makes the one-step diffusion student beats its multi-step teacher.

EFHQ: Multi-purpose ExtremePose-Face-HQ dataset.

A high-quality dataset centered on extreme pose faces, supporting face synthesis, reenactment, recognition benchmarking, and more.