M.S. Student, KAIST AI · BISPL · Advisor: Prof. Jong Chul Ye
I am an M.S. student at KAIST AI (BISPL) advised by Prof. Jong Chul Ye. I work on 3D vision and generative modeling — camera-controlled video generation, diffusion-based image enhancement, and vision-language-action models — with an eye toward robotics and embodied AI.
latest
Two more preprints are on arXiv, both under review: FastOPD, an on-policy distillation framework for lightweight VLA deployment, and DriftOPD, sequence-level reverse-KL distillation for one-step VLA policies.
Three new preprints are on arXiv, all under review: Dynamic Predictive Planning for dynamic manipulation, Action Upcycling for policy acceleration, and Adjoint Guidance Flow for VLA critic guidance.
CRePE was accepted as a poster at NeurIPS 2026.
Starting my M.S. at the Kim Jaechul Graduate School of AI, KAIST, joining BISPL under Prof. Jong Chul Ye.
CRePE was accepted to the ECCV 2026 Workshop on 3D in the Era of World Models, in Malmö this September.
Received my B.S. in Computer Science and Engineering from Korea University — early graduation in 3.5 years.
Preprint on DMD-augmented unpaired neural Schrödinger bridges for ultra-low-field MRI enhancement is on arXiv.
Our ultra-low-field brain MRI enhancement method was presented at the MICCAI ULF-EnC Challenge Workshop.
Received the Academic Excellence Award from Korea University.
background
I am a first-year M.S. student at the Kim Jaechul Graduate School of AI, KAIST, advised by Prof. Jong Chul Ye at BISPL. I received my B.S. in Computer Science and Engineering from Korea University in August 2026, graduating early in 3.5 years.
Before joining BISPL full-time, I interned there for over a year working on ultra-low-field MRI enhancement and camera-controlled video generation, and earlier studied Kubernetes scheduling at the Distributed and Cloud Computing Lab at Korea University.
Geometry-aware representations and camera-controlled video generation as a step toward world action models.
Vision-language-action policies and simulation (MuJoCo, Isaac Sim, LIBERO) for embodied agents.
Distribution matching, Schrödinger bridges, and diffusion models for image and video synthesis.
Ultra-low-field MRI enhancement and medical foundation models that work under real scanner constraints.
research
fig. 1 — roboticsDynamic Predictive Planning treats manipulation of moving targets as counterfactual planning: the world-action model's own rollout estimates when an interaction will happen and where the target will be, and a counterfactual observation lets the policy invoke a skill it already has instead of improvising a recovery. Runs in real time on a single consumer GPU with no training on dynamic data.
fig. 2 — video generationA positional encoding that models curved rays so a single video diffusion model can be controlled by pinhole, fisheye and panoramic cameras alike. An earlier version appeared at the ECCV 2026 Workshop on 3D in the Era of World Models.
timeline
M.S. in Artificial Intelligence · BISPL · Advisor: Prof. Jong Chul Ye
Seoul, Korea · Concentration: 3D Vision and Robotics
B.S. in Computer Science and Engineering · Advisor: Prof. Jaehoon Lee
Seoul, Korea · GPA 4.3 / 4.5 · Early graduation in 3.5 years · Academic Excellence Award
Advisor: Prof. Jong Chul Ye
Seoul, Korea
Advisor: Prof. Jong Chul Ye
Seoul, Korea
Advisor: Prof. Heonchang Yu
Seoul, Korea
projects
Counterfactual planning for moving targets: the world-action model's own rollout predicts when contact happens and where the target will be, so the policy reuses a skill it already has. Real time on one consumer GPU.
Curved Ray Expectation Positional Encoding so one video model follows pinhole, fisheye and panoramic camera trajectories.
Diffusion, residual-artifact suppression and unpaired Schrödinger-bridge models that lift 64 mT scans toward 3 T quality.
Large-scale pretraining and adaptation of a foundation model for medical imaging tasks.