MS · Computer Vision & Image Processing Lab · Sogang University
Graduate researcher advised by Prof. Unsang Park. I work on vision–language models and diffusion-based generation — from training-free control of diffusion attention to how we evaluate multimodal systems.
Vision–Language Model / Diffusion Models / Multimodal
I'm a master's student in the Computer Vision and Image Processing Lab (CVIP) at Sogang University, advised by Prof. Unsang Park. Before joining CVIP, I was an undergraduate researcher at the AI Accelerator Lab at Hallym University and a research intern at ETRI.
My current research treats cross-attention in diffusion models as a signal that can be analyzed and edited at inference time, and asks how vision–language models should be measured — designing diagnostics that separate what a model knows from how a benchmark is built.
Earlier work spanned sleep-stage classification, multi-modal physiological signals, and applied competitions in detection, segmentation, and OCR.
Under review (ICASSP 2027). A factorial study of audio–visual temporal binding that separates speech-time references from waveform access — showing that explicit timing improves accuracy on CMTB, while the additional effect of audio varies across models and evaluation settings.
Under review (ICASSP 2027). A matched-prompt and real-cascade study of upstream sound-event tags — showing that forwarding detector claims can reduce downstream accuracy and that message packaging and declared confidence change how strongly multimodal models follow those claims.
Under review (NAACL 2027). An audit of how visual evidence construction changes VLM resistance to supplied textual claims — combining human evidence-sufficiency checks and rendering controls to show that preserved target pixels and similar neutral accuracy do not guarantee comparable robustness scores.
arXiv preprint. An evaluation framework for long-horizon, fixed-camera video generation that distinguishes background drift from natural motion — reporting static fidelity and flow persistence separately, with drift leakage as interpretive context rather than an additional measurement axis.
arXiv preprint. A memory-and-training framework for long-horizon, fixed-camera nature video generation that balances spatial persistence with motion continuity — sustaining plausible fluid dynamics such as water, fire, and smoke over multi-minute autoregressive rollouts.
Accepted at CIKM 2026. A matched original / permutation / hard diagnostic protocol probing how vision–language models respond to same-type, visually plausible near-miss distractors — showing that evidence-first prompting is a model-dependent intervention rather than a universal fix.
Manuscript in preparation for WACV 2027. An inference-time method that modulates cross-attention logits in the Fourier domain to control generation without retraining, while largely preserving semantic alignment.
Expert Systems with Applications (ESWA), 2026
Int'l Conf. on ICT Convergence (ICTC), 2024 · ETRI Human Understanding AI Paper Challenge
arXiv preprint · SOTA on SleepEDF-78 and SHHS
Annual Symposium of KIPS (ASK), 2024 · in Korean
Joint Conf. on Communications and Information (JCCI), 2023 · in Korean
Transformer-based, PSG-free prediction of obstructive sleep apnea severity from multi-view facial images. Basis for the ESWA publication.
Class-level forgetting by shifting the decision boundary of a trained network — motivated by deepfake and data-removal settings.
github ↗Object detection + OCR pipeline that reads nutrition labels and logs personalized intake. Four-person team capstone project.
github ↗Route-adjustment system that recommends safer ship paths from ocean-state conditions to help prevent maritime accidents.
github ↗