Minsuk Ji

I am a M.S. student in Artificial Intelligence at Inha University, affiliated with the Generative Computing Lab, advised by Prof. Namhyuk Ahn. I received my B.S. in Electronic Engineering from Inha University in Feb 2025.

Previously, I worked as a Computer Vision Developer intern at Light Vision (Jun 2024 – Dec 2024).

I am interested in image generation, generative AI, and large language models (LLMs), and especially in building diverse, practical applications powered by these AI technologies.

I also keep a notes site where I write about what I'm studying and the projects I'm working on.

Email Email (Inha) Scholar GitHub LinkedIn

Minsuk Ji

News

Publications

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

Agentic amodal completion pipeline overview
KIISE Transactions on Computing Practices

A Reasoning-Based Amodal Completion Method for Occluded Objects Using Vision-Language Models

Minsuk Ji, Namhyuk Ahn† († corresponding author)

TL;DR: A vision-language model drives every stage of occluded-object completion β€” dynamically discovering scene entities, judging and adaptively expanding boundary-truncated regions, and writing appearance-aware inpainting prompts β€” reducing detection failures and improving reconstruction quality over fixed-pipeline baselines.

Code

Education

Experience

Projects

Unmanned systems RAG chatbot answering a question with a cited source
Hybrid RAG Chatbot for Unmanned Systems Specs

A retrieval-augmented QA chatbot over Jane's Unmanned Systems spec documents (350 documents split into 1,849 section-level chunks). BM25 keyword search and BGE-M3 dense retrieval are fused with Reciprocal Rank Fusion, and Qwen3.5-9B generates answers with source citations, so questions can be asked in Korean over English documents. On a rule-generated benchmark of 2,299 QA pairs across 11 question types, the hybrid retriever reached Hit@1 88.3% and Hit@5 99.3% (vs. 26.7% and 80.8% for BM25 alone), and adding the retrieved context raised answer Token F1 from 8.8% to 48.3%. (AI Convergence Project course, 2026)

Compositional Image Synthesis qualitative results
ReFocus: Compositional Image Synthesis

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 code
Agentic amodal completion pipeline overview
Agentic Amodal Completion

A VLM-driven pipeline that reasons through entity discovery, occlusion-order filtering, adaptive boundary expansion, and appearance-aware prompting to reconstruct occluded objects. Published in KIISE Transactions on Computing Practices.

code
CCTV-based crowd counting placeholder
CCTV-based Crowd Counting for Emergency Situations

A real-time model that analyzes CCTV footage from an unmanned parking control system to detect crowd density and count people during accidents or emergencies. (Light Vision)

Weather classification placeholder
Weather Classification for Parking Lot Monitoring

A deep learning model that quickly recognizes and automatically classifies adverse weather conditions, such as heavy rain and heavy snow, from parking lot CCTV footage. (Light Vision)

Water segmentation placeholder
Water Segmentation for Flood Damage Prevention

A segmentation system that precisely detects water regions in CCTV footage to distinguish normal conditions from flooding, for disaster preparedness. (Light Vision)