SangHyeok Lee

I am an undergraduate student in Information and Communication Engineering at Inha University. Currently, I am an AI Engineer Intern in the Physical Intelligence Lab at LG AI Research.

Previously, I worked on diffusion-based generative models at the Generative Computing Lab, advised by Prof. Namhyuk Ahn, and completed the Naver Boostcamp AI Tech program with a focus on computer vision.

My research interests include multimodal learning, world models, and diffusion models.

Email CV Scholar GitHub LinkedIn

SangHyeok Lee

News

Publications

StyleComposer qualitative results
Under Review

StyleComposer: Training-Free Multi-Reference Style Composition

Sanghyeok Lee, Jihye Kang, Namhyuk Ahn

TL;DR: A training-free method that composes color, texture, and spatial structure from separate references and provides independent control over each attribute.

Project Page arXiv

Compositional Image Synthesis qualitative results
ICASSP 2026

Compositional Image Synthesis with Inference-Time Scaling

Minsuk Ji*, Sanghyeok Lee*, Namhyuk Ahn (* equal contribution)

TL;DR: A training-free framework that improves compositional text-to-image generation by grounding LLM-synthesized layouts and reranking candidates with an object-centric VLM judge at inference time.

Project Page arXiv Code

Education

Experience

Awards & Honors

Projects

CCTV-based cargo object analysis pipeline
CCTV-based Cargo Analysis 2nd

End-to-end parcel counting and size estimation from conveyor-belt CCTV. The pipeline combines multi-view RF-DETR detection, OC-SORT tracking, rail-referenced geometry and depth, and a six-model CNN ensemble without requiring camera intrinsics.

code
Financial AI Challenge RAG system
Financial Security QA System 1st

Advanced RAG system for financial security QA built on KT Mi:dm 11.5B — FAISS retrieval, reranking, and self-consistency validation. Winner of the Financial Security Institute's Financial AI Challenge.

code interview
SummarAI paper chatbot
SummarAI: Paper Chatbot

End-to-end PDF understanding pipeline (formulas, tables, figures) with an advanced RAG stack — HyDE, query refinement, and reranking — scoring 0.75 on the Allganize Korean RAG benchmark.

code
Hand bone semantic segmentation
Hand Bone Segmentation 1st

Medical X-ray semantic segmentation with UNet++ and custom encoder backbones, boosting the Dice score by 4.5%p with a soft-ensemble strategy. Ranked 1st of 23 teams.

code
Recyclable waste object detection
Recyclable Waste Detection 1st

Detection pipeline built on MMDetection (DINO, CO-DETR, Cascade R-CNN), improving mAP50 from 49% to 76% with SR augmentation, TTA, and WBF ensembling. Ranked 1st of 24 teams.

code