SangHyeok Lee

I am an undergraduate student studying Information and Communication Engineering at Inha University and an undergraduate researcher at the Multisensory Intelligence Lab at UNIST.

My research focuses on world models, diffusion models, and representation learning.

SangHyeok Lee

News

Publications

Under Review StyleComposer qualitative results

StyleComposer: Training-Free Multi-Reference Style Composition

Sanghyeok Lee, Jihye Kang, Namhyuk Ahn

Compose color, texture, and spatial structure from separate style references, with independent control and no training.

ICASSP 2026 Compositional Image Synthesis qualitative results

Compositional Image Synthesis with Inference-Time Scaling

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

Improve compositional image generation with LLM-generated layouts and an object-centric VLM judge at inference time.

Experience

Education

Awards & Honors

More awards & honors (6)
  • 2025.08 3rd · SW-Centered University Digital Competition, AI DivisionDaconCodeSlides
  • 2025.08 3rd · INHA AI ChallengeInha Univ · DaconCode
  • 2025.06 Silver · Image Matching ChallengeKaggleCodeKaggle
  • 2025.06 4th · Bias-A-Thon: The Bias Response ChallengeSKKU CMI · DaconCodeSlides
  • 2024.07 1st · INHA SW NET-Zero Collaborative HackathonInha UnivCode
  • 2025 Scholar · Geumgang Hong Soon-Tak Scholarship FoundationAcademic Excellence

Selected Projects

Robot World Model video comparison

Robot World Model

Action-conditioned robot video generation with Cosmos-Predict2.5, LoRA, and inverse-dynamics supervision.

CCTV-based cargo object analysis pipeline

CCTV-based Cargo Analysis

Parcel counting and size estimation from conveyor-belt CCTV, combining detection, tracking, and depth.

Financial AI Challenge RAG system

Financial Security QA System

A financial security RAG system with retrieval, reranking, and self-consistency validation.

More projects (3)

SummarAI: Paper Chatbot

PDF understanding and retrieval-augmented question answering over formulas, tables, and figures.

Hand Bone Segmentation

Medical X-ray segmentation with UNet++ and soft ensembling; ranked first of 23 teams.

Recyclable Waste Detection

Waste detection with augmentation and model ensembling; ranked first of 24 teams.