Portrait of Duy A. Nguyen

Duy A. Nguyen

PhD Student ยท Computer Science ยท UIUC


I build multimodal and generative models that stay reliable under real-world conditions: missing inputs, corrupted data, crowded scenes.

Seeking Summer 2027 research internships

Hi, I'm Duy!


I'm a third-year PhD student in Computer Science at the University of Illinois Urbana-Champaign, advised by Prof. Minh N. Do and co-advised by Prof. Khoa D. Doan (VinUniversity).

I grew up in Hanoi, where one of my first research projects was forecasting river levels from noisy sensor data. It left me with a question I keep coming back to: what happens to a model when the data stops being clean? Today I chase it in multimodal models that survive missing inputs, in medical VLMs that know when not to answer, and in image generators that try to keep every person recognizable, even in a crowd of 20.

I love the full loop of research: a fresh idea, the experiments that prove it wrong, and the day it finally works. If that sounds like your team, let's talk!

News


  • 2026 SURE accepted at ACCV 2026. See you in Osaka!
  • 2026 Troupe, crowd-scale multi-identity image generation, is under submission.
  • 2026 When Should Medical VLMs Answer?, on medical VLM reliability under deployment perturbations, is under review.
  • 2026 Answer, Clarify, or Abstain at the ICML 2026 Workshop on Efficient Multimodal Question Answering.
  • 2025 Robult at IJCAI 2025 and MiniELM at ACL Findings 2025.
  • May 2025 Joined Cantina Labs as a Research Scientist Intern.

What I work on


Robust multimodal learning

Models that keep working when a modality is missing or labels are scarce, and that know when to be uncertain. IJCAI 2025, ACCV 2026.

Missing modalitiesUncertainty

Trustworthy vision-language models

Stress-testing medical VLMs on realistic, clinician-validated corruptions, and teaching them when to answer, ask, or abstain.

VLMsRobustnessCalibration

Efficient, controllable generation

Distilling large diffusion and language models into fast ones, and steering them with RL: few-step image and video diffusion, crowd-scale multi-identity generation, LLM query rewriting.

DiffusionLLMsRL

All research โ†’

Selected publications


All publications โ†’

Experience


2025 โ€“ now ยท Cantina Labs

Research Scientist Intern

Speeding up in-house image and video diffusion models through distillation while preserving quality; identity-preserving and audio-driven lip-sync video generation.

2023 โ€“ now ยท University of Illinois Urbana-Champaign

PhD Researcher

Advised by Prof. Minh N. Do and co-advised by Prof. Khoa D. Doan. Robust multimodal learning, uncertainty estimation and reliable vision-language models.

Full background โ†’

All models are wrong, but some are useful.

George E. P. Box, statistician

Get in touch


I'm looking for Summer 2027 research internships in multimodal learning, generative models and trustworthy AI. Email is the best way to reach me, and I'm always happy to chat about research.