Seonho (Griffin Sunho) Lee

Profile Image

I am Seonho Lee (Griffin Sunho Lee in the US). I am working as an Applied AI Researcher at Krafton AI. I got a M.S. degree from KAIST AI in Artificial Intelligence and had worked as an AI graduate researcher at the Computer Vision and Machine Learning Lab at KAIST, advised by Prof. Hyunjung Shim.

Particularly, I am researching on the following topics:

1. Generative AI

2. Vision-Language Understanding

CV E-mail Github LinkedIn Google Scholar

Email: glanceyes (at) kaist (dot) com

News

Mar. 2026: I begin to work as an Applied AI Researcher at Krafton AI.

Jan. 2026 Our What “Not” to Detect paper got accepted to ICLR 2026!

Oct. 2025: Two papers PartCATSeg and 3D-Aware VLM Finetuning got selected as QIFK(Qualcomm Innovation Fellowship Korea) 2025 Finalist.

Aug. 2025: Our 3D-Aware VLM Finetuning paper got accepted to EMNLP 2025 Findings!

Jun. 2025: I got ML engineering internship at Snap Inc. in Santa Monica, CA.

Jun. 2025: I got awarded Korean Presidential Science Scholarship for Graduate Students.

May. 2025: Our ScribbleDiff paper got accepted to ICIP 2025!

Feb. 2025: Our PartCATSeg paper got accepted to CVPR 2025!

Jan. 2025: Our DreamCatalyst paper got accepted to ICLR 2025!

More News

Sep. 2024: Our PartCLIPSeg paper got accepted to NeurIPS 2024!

Sep. 2023: I got accepted to CVML Lab as a M.S. graduate student in KAIST AI.

Selected Publications

*: equal contribution. †: corresponding author.

3D-Aware Vision-Language Models Fine-Tuning with Geometric Distillation

3D-Aware-VLM

Seonho Lee*, Jiho Choi*, Inha Kang, Jiwook Kim, Junsung Park, Hyunjung Shim†
Keywords: 3D-Aware VLM Finetuning

EMNLP 2025 Findings, Paper, Codes

Fine-Grained Image-Text Correspondence with Cost Aggregation for Open-Vocabulary Part Segmentation

PartCATSeg

Jiho Choi*, Seonho Lee*, Seungho Lee, Minhyun Lee, Hyunjung Shim†
Keywords: Open-Vocabulary Part Segmentation

CVPR 2025, Paper, Codes

Scribble-Guided Diffusion for Training-free Text-to-Image Generation

Scribble-Guided Diffusion

Seonho Lee*, Jiho Choi*, Seohyun Lim, Jiwook Kim, Hyunjung Shim†
Keywords: Conditional Image Generation

ICIP 2025, Paper, Codes

DreamCatalyst: Fast and High-Quality 3D Editing via Controlling Editability and Identity Preservation

DreamCatalyst

Jiwook Kim*, Seonho Lee*, Jaeyo Shin, Jiho Choi, Hyunjung Shim†
Keywords: 3D Editing, Score Distillation
Selected as a Daily Paper in HuggingFace

ICLR 2025, Paper, Project Page, Codes

Understanding Multi-Granularity for Open-Vocabulary Part Segmentation

PartCLIPSeg

Jiho Choi*, Seonho Lee*, Seungho Lee, Minhyun Lee, Hyunjung Shim†
Keywords: Open-Vocabulary Part Segmentation

NeurIPS 2024, Paper, Codes

Dense Reward for Multi-View 3D Reasoning with Global Maps and Local Views

DR3D

Jiho Choi*, Seonho Lee*, Seojeong Park, Hyunjung Shim†
Keywords: Multi-View 3D Reasoning, Reinforcement Learning

Under Review

→ More Publications

Work Experience

Krafton - AI Research Div., Applied AI Researcher

Mar. 2026 ~ Present , Seoul, Republic of Korea
Working in AI Art Gen Team at Krafton AI

Snap Inc. - Generative ML team (VideoCraft), ML Engineer Intern

Jun. 2025 ~ Sep. 2025, Santa Monica, CA, USA
Led a project for establishing cross-reference dataset preprocessing pipeline for personalized video generation
Developed new model architecture with multi-subject adapters for VideoAlchemist 2.0

Projects

Raon-VisionEncoder: A SigLIP2-class Vision Encoder Trained Entirely on Open Data, Krafton AI

Developed a vision encoder with comparable performance to SigLIP2-NaFlex using only open data
Built VLM training pipeline integrating the vision encoder with a language model for downstream VQA evaluation and training optimization

Blog Post

VideoAlchemist 2.0: Multi-Subject Personalized Video Generation, Snap Inc.

Developed a personalized video generation model supporting multiple subjects with fine-grained temporal control
Built cross-reference dataset pipeline and multi-subject adapter architecture for personalized video generation (contributed as a foundation for AlcheMinT)