Taesung Kwon

Postdoctoral Researcher, KAIST AI.

About Me

Hi, I am Taesung, an InnoCORE Postdoctoral Fellow at KAIST. I recently completed my Ph.D. at KAIST, where I was co-advised by Prof. Jong Chul Ye and Prof. Mooseok Jang. During my Ph.D., I also gained valuable industry experience as a research intern at Disney Research in Zurich. My research primarily focuses on improving the conditional generative process by controlling diffusion dynamics and designing efficient architectures. In parallel, I investigate how temporal correlations can be exploited to enhance generative modeling and solve inverse problems, particularly in tasks like video restoration.

My recent research spans three key areas, including but not limited to:

  • Generative Models: Diffusion models for video inverse problems, video frame interpolation, and image generation/editing.

  • Efficient Generative Modeling: Designing efficient architectures and training/sampling frameworks for diffusion models.

  • Physics-Based Vision and Medical Imaging: Applications to physical sciences (e.g., inverse scattering in optics) and medical imaging.

Education

KAIST

Ph.D. in Bio and Brain Engineering

Advisors: Jong Chul Ye and Mooseok Jang

2022 - 2026

KAIST

M.S. in Bio and Brain Engineering

Advisor: Jong Chul Ye

2020 - 2022

KAIST

B.S. in Bio and Brain Engineering

Advisor: Yoonkey Nam

2015 - 2020

Experience

InnoCORE Postdoc Fellow

LLM Innovation Research Center

Postdoc Advisor: Jong Chul Ye

2026.03 - Current

Disney Research

Research Internship @ Zürich

Mentor: Vinicius Azevedo

2025.05 - 2025.08

ETH Zürich

Visiting Researcher

Joint research program with Disney Research

2025.05 - 2025.08

Research

Reviving ConvNeXt for Efficient Convolutional Diffusion Models

Taesung Kwon, L. Bianchi, L. Wittke, F. Watine, F. Carrara, J. C. Ye, R. Weber, V. C. Azevedo

CVPR 2026

Fully convolutional diffusion model based on ConvNeXt.

Zero4D: Training-Free 4D Video Generation From Single Video Using Off-the-Shelf Video Diffusion Model

https://zero4dvid.github.io/

J. Park, Taesung Kwon, J. C. Ye

CVPR 2026 (Findings)

Zero-shot 4D video generation using off-the-shelf video diffusion model.

Alignment-Guided Score Matching for Text-to-Image Alignment in Diffusion Models

J. Lee*, Y. Hong*, Taesung Kwon, J. C. Ye

To appear

Text-image alignment-guided score matching.

VISION-XL: High Definition Video Inverse Problem Solver using Latent Diffusion Models

https://vision-xl.github.io/

Taesung Kwon, J. C. Ye

ICCV 2025

Solving HD video inverse problems using only latent diffusion models. Supporting wide-range ratio using SDXL.

ViBiDSampler: Enhancing Video Interpolation Using Bidirectional Diffusion Sampler

https://vibidsampler.github.io/

S. Yang*, Taesung Kwon*, J. C. Ye (*co-first)

ICLR 2025

State-of-the-art performance, efficiently generating high-quality, smooth videos between keyframes.

Solving Video Inverse Problems Using Image Diffusion Models

https://svi-diffusion.github.io/

Taesung Kwon, J. C. Ye

ICLR 2025

Solving video inverse problems using only image diffusion models, with batch-consistent sampling stretagy.

Video Diffusion Posterior Sampling for Seeing Beyond Dynamic Scattering Layers

https://github.com/star-kwon/VDPS

Taesung Kwon*, G. Song*, Y. Kim, J. Kim, J. C. Ye, M. Jang (*co-first)

IEEE TPAMI

Video reconstruction through dynamic scattering layer using video diffusion models.

Highly Personalized Text Embedding for Image Manipulation by Stable Diffusion

https://hiper0.github.io/

I. Han*, S. Yang*, Taesung Kwon, J. C. Ye

arXiv, 2023

Image manipulation using personalized text embedding from Stable Diffusion.

DiffusionCLIP: Text-Guided Diffusion Models for Robust Image Manipulation

https://github.com/gwang-kim/DiffusionCLIP

G. Kim, Taesung Kwon, J. C. Ye

CVPR 2022

Robust text-driven image manipulation using text-to-image diffusion models.

Noise Distribution Adaptive Self-Supervised Image Denoising using Tweedie Distribution and Score Matching

https://github.com/cubeyoung/NoiseAdaptive2Score

K. Kim, Taesung Kwon, J. C. Ye

CVPR 2022

Noise distribution and level adaptive denoising framework using Tweedie distribution and score matching.

Cycle-free CycleGAN using Invertible Generator for Unsupervised Low-Dose CT Denoising

https://github.com/star-kwon/TCI_CyclefreeCycleGAN

Taesung Kwon, J. C. Ye

IEEE TCI

Efficient CycleGAN framework for low-dose CT denoising.

Workshop

NTIRE 2022 Spectral Recovery Challenge and Data Set

CVPRW 2022

Recovering hyperspectral information from JPEG-compressed RGB images.

NTIRE 2022 Challenge on Night Photography Rendering

CVPRW 2022

Recovering the visual appearance of night photography.

Patents

  • Method and Apparatus for Low-Dose X-Ray Computed Tomography Image Processing Based on Efficient Unsupervised Learning Using Invertible Neural Network

J. C. Ye, Taesung Kwon

U.S. Patent, Granted, No. 12,346,997, 2025

Korean Patent, Granted, No. 10-2643601, 2024

  • Method and Apparatus for Generating Intermediate Video Frames via Bidirectional Sampling

J. C. Ye, S. Yang, Taesung Kwon

Korean Patent, Filed, No. 10-2025-0114505, 2025

  • Method and Apparatus for Solving Video Inverse Problem using Image Diffusion Models

J. C. Ye, Taesung Kwon

Korean Patent, Filed, No. 10-2025-0171861, 2025

Awards and Honors

  • KAIST Scholarship, KAIST, 2022-2025
  • Korean Government Scholarship, KAIST, 2020-2021

Media

  • TPAMI Research featured on YTN News (a major Korean news channel), 2025 youtube

Services

  • Conference reviewers: CVPR, ICLR, ICCV, ECCV, ICML
  • Journal reviewers: IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Computational Imaging