Research Questions

Research Statement

"우리는 인간의 인식을 넘어, 우리가 보고 느끼는 세상을 더욱 정교하게 모델링하고 이해하는 것을 목표로 합니다. 이를 통해 우리가 세상을 경험하고 해석하는 방식에 혁신적인 변화를 일으키고자 합니다. 또한 우리는 생성 모델이 세계를 이해하는 시스템을 구축하기 위한 가장 강력한 솔루션이자 효과적인 접근법이라고 믿습니다. 이 때, 단순히 높은 성능을 넘어, 설명 가능하고, 효율적이며, 직관적이고, 사용하기 쉬운 이미징 알고리즘 개발을 추구합니다."

Topics & Keywords of Interest

Below is a list of project topics and keywords that reflect our research interests. 

• Note 1: Topics marked with (!!!) indicate that we care currently looking for a new intern or student to work on these. For 2025 and 2026, we are offering openings in three topics. 

• Note 2: The provided examples are intended to help clarify and convey the concept, not limited to our research specifically.

Computational Imaging 

• Natural image restorations (e.g., super-resolution, denoising, deblurring for image, video, etc.)

• Bio/Medical/Satellite imaging

Efficient & Sustainable Learning 

• (!!) Federated learning for generative models (e.g., PRISM)

• Lighter & faster generative models (e.g., Diffusion2GAN)

• Data-efficient learning (e.g., few-shot/pre-training; privacy-free synthetic data) for generative models

Physical AI: Generative model as a representation learner

• (!!!) Multimodal (>2) representation learning (e.g., imagebind) + Robotics

• Generative classifiers (e.g., unsupervised segmentation, hierarchical grouping, etc)

Deep generative models for complex data (beyond 2D) with better control

• (!!) (Long) Video understanding & generation (e.g., powerful video backbone, world simulator, digital human, etc.) 

• 3D/4D generation and beyond (e.g., text-to-4D)

• Layout/layout-conditioned generation (e.g., graphic design, storyboard-to-movie, etc.)

• Generative models for non-image data (e.g., scene-text, tabular, time-series, etc.)

Explainable & Reliable Generation 

• Evaluating generative models (e.g., fidelity vs diversity, memorization, physical plausibility)

• Deepfake detection and prevention

Bridging between signal processing and deep learning communities

• Providing a design principle for deep learning architectures (e.g., Is the current attention mechanism optimal for manipulation?)