Taegyeong Lee

AI Researcher · Generative AI, LLMs, RAG

I'm passionate about novel research that explores generating images or videos from audio and text — integrating various modalities. I enjoy conducting research that is simple yet effective, leveraging multimodal and generative models to make a strong impact in the real world.

Currently working as an AI researcher on the FnGuide Inc, focusing on LLMs and RAG. Previously, I earned my Master's degree from UNIST AIGS. I interned at ETRI and completed the Software Maestro 8th. I also served as a software developer in the Promotion Data Management Division at the Republic of Korea Army Headquarters. I hold a Bachelor of Computer Engineering from Pukyong National University.

Email Scholar Github

Research

My current primary research interests include Generative AI, LLMs, and RAG.

DEO: Training-Free Direct Embedding Optimization for Negation-Aware Retrieval

Taegyeong Lee^*, Jiwon Park^*, Seunghyun Hwang^*, JooYoung Jang^†

Underreview

arXiv

We propose DEO, a fine-tuning-free method that optimizes embeddings for negation-aware retrieval, improving performance across text and multimodal tasks.

Exploring Diffusion Models for Generative Forecasting of Financial Charts

Taegyeong Lee, Jiwon Park, Kyunga Bang, Seunghyun Hwang, Ung-Jin Jang

Preprint

github arXiv

We propose a novel stock price prediction approach that treats time-series data as images and leverages text-to-image generative models to generate and evaluate future chart patterns.

QGuard: Question-based Zero-shot Guard for Multi-modal LLM Safety

Taegyeong Lee, Jeonghwa Yoo, Hyoungseo Cho, Soo Yong Kim and Yunho Maeng

ACL 2025 Workshop (The 9th Workshop on Online Abuse and Harms)

github arXiv

This paper proposes a simple yet effective question prompting method to block harmful prompts, including multi-modal ones, in a zero-shot and robust manner.

Multi-aspect Knowledge Distillation with Large Language Model

Taegyeong Lee^*, Jinsik Bang^*, Soyeong Kwon, Taehwan Kim

CVPR 2025 Workshop (The 12th Workshop on Fine-Grained Visual Categorization)

github arXiv

We introduce a multi-aspect knowledge distillation method using MLLMs to enhance vision models by learning both visual and abstract aspects, improving performance across tasks.

Generating Realistic Images from In-the-wild Sounds

Taegyeong Lee, Jeonghun Kang, Hyeonyu Kim, Taehwan Kim

ICCV 2023

github arXiv

We propose a diffusion-based model that generates images from wild sounds using audio captioning, attention mechanisms, and CLIP-based optimization, achieving superior results.

Generating Emotional Face Images using Audio Information for Sensory Substitution

Taegyeong Lee, Hyerin Uhm, Chi Yoon Jeong, Chae-Kyu Kim

Journal of Korea Multimedia Society, 2023 (Outstanding Paper Award)

We propose a method to generate images optimized for sound intensity, enhancing V2A models for improved face image generation.

An enhanced model of Face Reenactment and Transformation Models based on Head pose vector

Taegyeong Lee, Gyubin Park, HyeJin Seo, Su-Hwa Jo, Chae-Kyu Kim

Conference of Korea Multimedia Society, 2021 (AI Capstone Design Grand Prize)

We propose a face synthesis method using dense landmarks for accurate head pose estimation, yielding more natural results than existing methods.

Taegyeong Lee

Research

Academic Activities / Awards