Dong-Hwan Jang

RewardFlow: Generate Images by Optimizing What You Reward

Onkar Kishor Susladkar, Dong-Hwan Jang, Tushar Prakash, Adheesh Sunil Juvekar, Vedant Shah, Ayush Barik, Nabeel Bashir, Muntasir Wahed, Ritish Shrirao, Ismini Lourentzou
CVPR, 2026

RewardFlow is a zero-shot, training-free framework for text-guided image editing and generation using reward-guided Langevin dynamics. We steer pretrained diffusion and flow-matching models at inference with hierarchically designed coarse-to-fine differentiable rewards (e.g., a VQA-based reward for semantic supervision and a SAM-guided reward for localized edits), controlled by a prompt-aware adaptive policy.

PyraTok: Language-Aligned Pyramidal Tokenizer for Video Understanding and Generation

Onkar Kishor Susladkar, Tushar Prakash, Adheesh Sunil Juvekar, Kiet A. Nguyen, Dong-Hwan Jang, Inderjit S. Dhillon, Ismini Lourentzou
CVPR, 2026

PyraTok is a language-aligned pyramidal tokenizer for video understanding and generation that learns discrete latents across multiple spatial-temporal resolutions. We introduce Language-aligned Pyramidal Quantization (LaPQ), discretizing encoder features at several depths with a shared large binary codebook. PyraTok achieves state-of-the-art video reconstruction, text-to-video quality, and zero-shot performance on segmentation, action localization, and understanding, scaling to 4K/8K resolutions.

Model Stock: All we need is just a few fine-tuned models

Dong-Hwan Jang, Sangdoo Yun, Dongyoon Han
ECCV, 2024 (Oral, Top 2.3% among submitted papers).

arxiv / github / article (marktechpost)

Our analysis reveals that fine-tuned weights reside on a thin shell in the weight space, maintaining nearly constant layer-wise angles and norms. By leveraging the geometric link between center proximity and performance, we propose an efficient merging method that approximates the weight center with only two models—achieving superior ID/OOD performance at 24x less cost than Model Soup.

Merge and Bound: Direct Manipulations on Weights for Class Incremental Learning

Taehoon Kim, Dong-Hwan Jang, Bohyung Han
CVPR Workshop on Continual Learning in Computer Vision, 2024

We introduce Merge-and-Bound (M&B), an innovative approach for Class Incremental Learning that optimizes model weights through two merging techniques: inter-task and intra-task weight merging, alongside a bounded update to prevent catastrophic forgetting. Without altering architectures or objectives, M&B integrates into existing methods, showing superior performance on CIL benchmarks against top competitors.

Pooling Revisited: Your Receptive Field is Suboptimal

Dong-Hwan Jang, Sanghyeok Chu, Joonhyuk Kim, Bohyung Han
CVPR, 2022

We propose DynOPool, a learnable pooling layer that finds the optimal scale factors and receptive fields of intermediate feature maps. While fixed pooling operators often result in suboptimal receptive fields by ignoring dataset-specific characteristics, DynOPool enables data-driven resolution scaling, achieving up to a 5.6%p accuracy gain on fine-grained recognition (FGVC-Aircraft) and approximately 33% fewer MACs on ImageNet with comparable accuracy. It adapts feature map sizes and shapes for enhanced accuracy and efficiency in tasks like image classification and semantic segmentation.

DS4C Patient Policy Province Dataset: a Comprehensive COVID-19 Dataset for Causal and Epidemiological Analysis

Jimi Kim*, Seojin Jang*, Joong Kun Lee*, Dong-Hwan Jang* (* equal contributions)
NeurIPS Workshop on Causal Discovery & Causality-Inspired Machine Learning, 2020

project page / article (korean) (auto-translated)

We present DS4C South Korea Patient, Policy, and Provincial data (DS4C-PPP dataset). The dataset contains comprehensive data that could be used for causal analysis, such as per-patient symptom onset and confirmed date, travel frequency, hospital accessibility, and 61 preventative policies enacted in South Korea.

DIFF: Deblurring Implicit Feature Function

We propose spatially-variant motion deblur network based on the implicit neural representation. A spatially-variant deblurring network takes deformed features and their offsets as inputs.

U.S. Patent Application Number: 17/973,809 (in progress)

Dynamic Spatially-Adaptive Modulation for image Dehazing

We propose a novel framework of the dynamic spatially-adaptive modulation for image dehazing. The proposed algorithm introduces a selection module that conditionally determine the necessary modulation pathways in a bottom-up manner by providing a loss function optimizing both accuracy and efficiency.

DepthFinder: Universal Image Restoration based on Adaptive Inference

We adaptively find the appropriate number of the residual blocks according to the severity and distortion type of the input in universal image restoration task.