Seoul National University’s College of Engineering announced that Hyunsoo Lee, an undergraduate student in the Department of Electrical and Computer Engineering, has presented research in generative visual computing, spanning image editing and 3D content generation, at leading international conferences in machine learning and computer vision, including NeurIPS, CVPR, and ECCV. These conferences represent premier venues in their respective fields, making Lee’s sustained record of publication as an undergraduate particularly noteworthy.
Generative visual computing extends beyond enabling AI model to understand real-world visual information; it seeks to generate new visual content from user-provided text, images, and partial observations. Although generative AI, including diffusion models, has advanced rapidly in recent years, precisely aligning outputs with user intent and maintaining structural consistency across multiple generations remain fundamental challenges.
Lee’s research leverages the prior knowledge encoded in pretrained generative models to produce outputs that satisfy user intent and observational constraints while remaining consistent across related generations. His work spans text-driven and high-resolution image editing, diffusion synchronization, and 3D content generation.
Lee began by studying text-conditioned image editing. In “Conditional Score Guidance for Text-Driven Image-to-Image Translation,” for which he served as a co-first author, the researchers introduced a conditional score guidance method that modifies text-specified regions while preserving the structure of the source image. The paper was accepted by NeurIPS 2023, one of the leading conferences in machine learning.
This line of research was subsequently extended to high-resolution image editing. “Low-Resolution Editing is All You Need for High-Resolution Editing,” accepted by CVPR 2026, uses a low-resolution edit as a semantic reference and transfers fine-grained details from the source image to the high-resolution output through synchronized patch optimization. The method enables editing at resolutions above 1K while preserving both global semantics and local detaill, which have been difficult for conventional generative models to handle.
Beyond the generation of individual outputs, Lee investigated the principles of diffusion synchronization, which coordinates multiple generations so that they share a consistent underlying structure. “SyncSDE: A Probabilistic Framework for Diffusion Synchronization,” presented at CVPR 2025, identifies from a probabilistic perspective where correlations should be introduced among diffusion trajectories and provides a unified framework for collaborative generation across images, human motion, and 3D content. By moving beyond the independent generation of each output, the work establishes a theoretical foundation for producing coherent sets of interrelated content.
“Accelerated Likelihood Maximization for Diffusion-based Versatile Content Generation,” accepted by ECCV 2026, efficiently optimizes unobserved regions during sampling, allowing a single method to address distinct tasks such as 3D mesh texturing and human motion completion. The approach reduces the need to train a separate model for each task and demonstrates how pretrained generative models can be adapted flexibly across a broad range of visual content generation problems.
Lee has also extended generative methods to the 3D structure of the world. “Image-Guided Geometric Stylization of 3D Meshes,” presented at CVPR 2026 as a co-first author, goes beyond transferring the appearance of a reference image onto a 3D object’s texture: it modifies the object’s geometry itself to reflect the target style. In “Point2Pose: A Generative Framework for 3D Human Pose Estimation with Multi-View Point Cloud Dataset,” presented at WACV 2026, the researchers reformulate 3D human pose estimation from point clouds as a conditional generation problem and introduce a generative approach to 3D human pose estimation.
Lee has also contributed to research on applying generative models to scientific domain. The ICML 2026 paper “Calibrated Test-Time Guidance for Bayesian Inference,” on which Lee is a co-author, shows that existing guidance methods can produce miscalibrated approximations to Bayesian posterior distributions and proposes consistent estimators for posterior sampling. The work extends guidances beyond content creation to problems in scientific and statistical inference.
Lee is majoring in Electrical and Computer Engineering and pursuing a double major in Mathematical Sciences at Seoul National University. He previously conducted research in Professor Bohyung Han’s Computer Vision Laboratory and Professor Young Min Kim’s 3D Vision Laboratory, both within SNU’s Department of Electrical and Computer Engineering. As an exchange student at the University of California, Irvine, he also conducted research with Professor Stephan Mandt. He is currently studying representation learning mechanisms inspired by memory at SNU’s Machine Perception and Reasoning Laboratory.
“Harnessing the capabilities of generative models while respecting both user intent and the structure of the source image remains an important research challenge,” said Bohyung Han, professor in SNU’s Department of Electrical and Computer Engineering. “With a particular focus on image editing, Hyunsoo Lee has consistently explored methods for achieving fine-grained control over generated outputs.”
“Connecting visual information from images to geometric deformation and human pose estimation expands the scope of generative AI,” said Young Min Kim, professor in SNU’s Department of Electrical and Computer Engineering. “Drawing on his understanding of both recent advances in generative models and their mathematical foundations, Hyunsoo Lee continues to pursue challenging problems in new domains.”
“My research has been guided by a central question: How can we reliably generate outputs that satisfy user intent and observational constraints while drawing on the knowledge encoded in generative models?” Lee said. “Going forward, I hope to connect generative modeling with representation learning to develop intelligent systems that can adapt to diverse scenarios while making their underlying mechanisms more interpretable.”
Lee’s research may ultimately contribute to digital content creation, virtual reality, and human-motion analysis, as well as scientific and engineering domains that demand both precise generation and reliable inference.
□ Introduction to the SNU College of Engineering
Seoul National University (SNU) founded in 1946 is the first national university in South Korea. The College of Engineering at SNU has worked tirelessly to achieve its goal of ‘fostering leaders for global industry and society.’ In 12 departments, 323 internationally recognized full-time professors lead the development of cutting-edge technology in South Korea and serving as a driving force for international development.
Computational simulation/modeling
Not applicable