Jeongsoo Park

jp2678@cornell.edu

I am a fourth-year CS PhD student at Cornell Tech, advised by Prof. Andrew Owens.

I am broadly interested in computer vision and machine learning. My current research interests include multimodal generation, vision-language models, visual forensics, and efficient deep learning.

I moved with my advisor from the University of Michigan to Cornell Tech in Fall 2025. Previously, I received my M.S. in ECE from Michigan, advised by Prof. Justin Johnson, and my B.S. from Sungkyunkwan University, advised by Prof. Jong Hwan Ko.

Research Interests

I am interested in generating, understanding, and authenticating multimedia content, and making these systems more efficient.

Multimodal generation

Generating images, videos, and audio-visual content with meaningful control.

Vision-language models

Connecting visual understanding with language and reasoning.

Visual forensics

Understanding the origins and authenticity of visual content.

Efficient deep learning

Reducing the cost of training, inference, and data processing.

Publications

Community Forensics: Using Thousands of Generators to Train Fake Image Detectors

Jeongsoo Park, Andrew Owens
CVPR 2025
Project page / Dataset / Code / arXiv

We train fake image detectors using images from thousands of various generators to study generalization in detecting generated images.

RGB no more: Minimally-decoded JPEG Vision Transformers

Jeongsoo Park, Justin Johnson
CVPR 2023
Project page / Code / arXiv

We train ViTs directly from JPEG encoded features and accelerate train and eval by up to 39.2% and 17.9%.

Overview of Auto-Tiler: Variable-Dimension Autoencoder with Tiling for Compressing Intermediate Feature Space of Deep Neural Networks for Internet of Things

Auto-Tiler: Variable-Dimension Autoencoder with Tiling for Compressing Intermediate Feature Space of Deep Neural Networks for Internet of Things

Jeongsoo Park, Jungrae Kim, Jong Hwan Ko
Sensors, 2021
Paper

We use autoencoders and show up to 67% higher accuracy and 81% reduced latency during over-the-network inference versus the image or video codec-based compression.

Research Experience