I work on visual generative models, vision-language models, and domain transfer — with a focus on reward optimization and alignment for compositional text-to-image generation.
I am a master's student in Artificial Intelligence at Korea University and a researcher in UTL Lab. My research interests include image generative models, text-to-image diffusion models, vision-language models, and domain transfer. I enjoy building reward-driven training methods that make visual models follow complex prompts more faithfully.
Text-to-image generation, diffusion models, VLMs, domain generalization, and unsupervised domain adaptation.
Multi-concept reward optimization and reinforcement learning for compositional generation.
Python, PyTorch, C, computer vision, machine learning, deep learning, and reinforcement learning.
Master's Degree in Artificial Intelligence · Seoul, Korea
UTL Lab Researcher · Advisor: Prof. Donghyun Kim
Bachelor's Degree in Mobile System Engineering · Yongin, Korea
HAIL Lab Undergraduate Research Intern · Advisor: Prof. Dongjae Kim
Proposed CMO to resolve attribute omissions in text-to-image models via correlation-based reward reweighting.
Achieved state-of-the-art compositional generation on multi-concept benchmarks using SD3.5 and FLUX.1-dev.
Proposed MiPO to improve diffusion classifiers' minority perception via GRPO-based policy optimization.
Boosted zero-shot classification across five benchmarks using SD 1.5/2.0 without extra image data.
Demonstrated that large-scale pre-trained models significantly outperform existing domain transfer methods.
Revealed through class prototype analysis and pruning that the advantage comes from high cross-dataset class similarity.
Designed a Vision Transformer-based reinforcement learning agent and compared CNN-DQN and ViT-DQN in the CartPole environment.
Excellence Prize · Samsung Research
Developed an efficient and lightweight VLM for general VQA tasks.
Excellence Prize · Hongik University, Samsung Medical Center, Kyung Hee University Dental Hospital
Classified oral health risk using CT image data in a closed environment.
Excellence Prize · National Radio Research Agency
Built classification and regression models using space radio signals.
Excellence Prize · Dankook University
Predicted bike rental demand using feature engineering and machine learning models.
Feel free to reach out for research discussions, collaboration, or project questions.