avatar

Dongjin Kim

Ph.D. Student
CS, Hanyang University
dongjinkim (at) hanyang.ac.kr


πŸ‘¨β€πŸŽ“ About Me

I'm a Ph.D. student in Computer Science at Hanyang University, South Korea, where I am fortunate to be advised by Prof. Tae Hyun Kim.

My research focuses on developing reliable and efficient restoration systems for real-world images and videos. Broadly, my work lies at the intersection of computer vision, machine learning, and generative modeling.

I design both dataset synthesis frameworks and restoration models for tasks such as video stabilization, image denoising, and super-resolution. A key emphasis of my research is ensuring robustness under practical constraints, including test-time adaptation, continuous degradations, diverse unseen noise patterns.

πŸ“’ News

  • 2026/02 β€” PNG: Diffusion-Based sRGB Real Noise Generation via Prompt-Driven Noise Representation Learning
    accepted to CVPR 2026πŸŽ‰.
  • 2025/11 β€” DegFlow: Continuous Degradation Modeling via Latent Flow Matching for Real-World Super-Resolution
    accepted to AAAI 2026πŸŽ‰.
  • 2025/06 β€” IDF: Iterative Dynamic Filtering for Generalizable Denoising accepted to ICCV 2025πŸŽ‰.
  • 2024/02 β€” MetaStab: Harnessing Meta-Learning for Full-Frame Video Stabilization accepted to CVPR 2024πŸŽ‰.
  • 2024/01 β€” NAFlow: sRGB Real Noise Modeling with Normalizing Flows accepted to ICLR 2024πŸŽ‰.
  • 2023/07 β€” Task Agnostic Restoration of Natural Video Dynamics accepted to ICCV 2023πŸŽ‰.
  • 2023/04 β€” InterFlow: Controllable Degradation for Real-World Super-Resolution accepted to ICML 2023πŸŽ‰.

πŸ“š Publications

    * Equal contribution. †Corresponding author.

  1. CVPR 2026
    Diffusion-Based sRGB Real Noise Generation via Prompt-Driven Noise Representation Learning
    Jaekyun Ko*, Dongjin Kim*, Soomin Lee, Guanghui Wang, Tae Hyun Kim†
    Paper Code

    tl;dr: PNG synthesizes realistic sRGB noise directly from input images without requiring camera metadata, enabling real-world sRGB denoising more practical and broadly applicable across devices.

    Real-World Denoising Dataset Generation Degradation Modeling One-step Diffusion

  2. AAAI 2026
    Continuous Degradation Modeling via Latent Flow Matching for Real-World Super-Resolution
    Hyeonjae Kim*, Dongjin Kim*, Eugene Jin, Tae Hyun Kim†
    Paper Code

    tl;dr: DegFlow learns a continuous real-world degradation manifold from sparse HR–LR pairs and uses it to generate realistic LR images at any scale, enabling stronger fixed-scale and arbitrary-scale SR models.

    Real-World Super-Resolution Dataset Generation Degradation Modeling Flow Matching

  3. ICCV 2025
    IDF: Iterative Dynamic Filtering Networks for Generalizable Image Denoising
    Dongjin Kim*, Jaekyun Ko*, Muhammad Kashif Ali, Tae Hyun Kim†
    Paper Code Project Page

    tl;dr: We design lightweight, iterative pixel-wise filters that adapt to local content and generalize to unseen real noise while keeping runtime low.

    Generalizable Denoising Dynamic Filtering Light-Weight Efficiency

  4. CVPR 2024
    Harnessing Meta-Learning for Improving Full-Frame Video Stabilization
    Muhammad Kashif Ali, Eun Woo Im, Dongjin Kim, Tae Hyun Kim†
    Paper Code

    tl;dr: We use meta-learning to adapt full-frame video stabilization across scenes and cameras, improving robustness without task-specific tuning.

    Video Stabilization Meta-Learning Temporal Consistency Robustness

  5. ICLR 2024
    sRGB Real Noise Modeling via Noise-Aware Sampling with Normalizing Flows
    Dongjin Kim*, Donggoo Jung*, Sungyong Baik, Tae Hyun Kim†
    Paper Code

    tl;dr: We model real sRGB noise with normalizing flows and noise-aware sampling to generate realistic noise for training and evaluation.

    Dataset Generation Denoising Normalizing Flows Noise-Aware Sampling

  6. ICCV 2023
    Task Agnostic Restoration of Natural Video Dynamics
    Muhammad Kashif Ali, Dongjin Kim, Tae Hyun Kim†
    Paper Code

    tl;dr: We restore natural video dynamics in a task-agnostic way, improving stability and temporal coherence without task-specific labels.

    Video Restoration Stabilization Task-Agnostic

  7. ICML 2023
    Learning Controllable Degradation for Real-World Super-Resolution via Constrained Flows
    Seobin Park*, Dongjin Kim*, Sungyong Baik, Tae Hyun Kim†
    Paper Code

    tl;dr: We learn realistic and controllable degradations with constrained flows, enabling stronger real-world super-resolution.

    Real-World SR Degradation Modeling Normalizing Flows Controllable



🀝 Services

Conference Reviewers

Journal Reviewers