Sangeek Hyun

I am a Ph.D. candidate in the Visual Computing Lab (VCLab) at Sungkyunkwan University, supervised by Prof. Jae-Pil Heo. I received my Master's and Bachelor's degrees from Sungkyunkwan University. My research interests include various tasks in machine learning and computer vision, with a particular focus on generative models and video understanding.

Email  /  Google Scholar  /  Github  /  LinkedIn

Research

Recently my interest has been in 3D generative models using Generative Adversarial Networks and gaussian splatting. I am also interested in various generation tasks using large-scale diffusion models.

Pure Transformer GANs can be scaled up and beats diffusion/flow models on an one-step conditional generation on ImageNet-256 only within 40 epochs.

This paper introduces a fast and effective VAR-based method for subject-driven image generation, using selective and scale-wise weighted tuning to overcome fine-tuning challenges and outperform diffusion models.

This paper proposes AESOP, a simple yet effective loss that replaces pixel-wise Lp loss with a distance in the autoencoder output space, enabling better reconstruction in perceptual super-resolution without sacrificing visual quality.

First 3D GANs utilize gaussian splatting without any structural priors, achieving faster rendering speed at high-resolution data compared to NeRFs.

Training-free style transfer utilizes large-scale diffusion models by manipulating the attention features.

GAN compression technique by pruning the diversity-aware channels in StyleGAN architecture.

Task-disruptive Background Suppression mitigates the negative impact of dissimilar or target-similar support backgrounds, improving the accuracy of segmenting novel target objects in query images.

CG-DETR improves temporal grounding by using adaptive cross-attention and clip-word correlation to accurately identify video highlights corresponding to textual descriptions.

Propose a frequency-based motion representation for video GANs, enabling speed-aware motion generation, which improves video quality and editing capability.

Disentangled representation learning for unsupervised neural quantization addresses deep learning quantizers' limitations in leveraging residual vector space, enhancing search efficiency and quality.

Query-Dependent DETR improves video moment retrieval and highlight detection by enhancing query-video relevance and using negative pairs to refine saliency prediction.

The Local Attention Pyramid (LAP) module addresses class-wise visual quality imbalance in GAN-generated scene images by enhancing attention to diverse object classes, particularly small and less frequent ones.

Self-supervised approaches with dual discriminators improve video GANs by ensuring appearance consistency and motion coherency through contrastive learning and temporal structure puzzles.

VarSR leverages latent distributions to address the many-to-one nature of single image super-resolution, generating diverse high-resolution images from low-resolution inputs.

Website template.