|
MinKyu Lee 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 low-level vision, image restoration and generative modeling. 2minkyulee@gmail.com (Primary Contact) / Google Scholar / Github / LinkedIn / CV |
|
Publications |
SeaCache: Spectral-Evolution-Aware Cache for Accelerating Diffusion Models
Jiwoo Chung, Sangeek Hyun, , Byeongju Han, Geonho Cha, Dongyoon Wee, Youngjun Hong, Jae-Pil Heo† CVPR, 2026 Paper (PDF) / arXiv / Code A training-free, plug-and-play caching schedule that applies a timestep-dependent Spectral-Evolution-Aware (SEA) filter before measuring feature distances, so reuse decisions track content (signal) rather than high-frequency noise, improving latency–quality trade-offs. |
Progressive Distractor Filtering for Robust 3D Gaussian Splatting
Kangmin Seo, , Tae-Young Kim, ByeongCheol Lee, JoonSeoung An, Jae-Pil Heo† CVPR Findings, 2026 Paper (PDF) / arXiv / Code A two-phase 3DGS framework: progressive filtering iteratively finds and masks view-inconsistent distractors via rendered-vs-train discrepancies, then a reconstruction phase restores fine details from the purified representation. PDF-GS show robust, distractor-free reconstructions with no inference-time overhead. |
Analyzing the Training Dynamics of Image Restoration Transformers: A Revisit to Layer Normalization
, Sangeek Hyun, Woojin Jun, Hyunjun Kim, Jiwoo Chung, Jae-Pil Heo† ICLR, 2026 OpenReview / arXiv / Code This paper analyzes training instability in image restoration Transformers caused by conventional per-token LayerNorm, and proposes an IR-tailored normalization (i-LN) that stabilizes training and improves performance across restoration tasks. |
Diffusion Feature Field for Text-based 3D Editing with Gaussian Splatting
Eunseo Koh, Sangeek Hyun, , Jiwoo Chung, Kangmin Seo, Jae-Pil Heo† NeurIPS, 2025 OpenReview This paper proposes DFFSplat, which injects a 3D-consistent diffusion feature field into the 3D Gaussian Splatting editing pipeline to enforce multi-view consistency and mitigate view-inconsistent artifacts like the Janus problem in text-based 3D editing. |
Fine-Tuning Visual Autoregressive Models for Subject-Driven Generation
Jiwoo Chung, Sangeek Hyun, Hyunjun Kim, Eunseo Koh, , Jae-Pil Heo† ICCV, 2025 Qualcomm Innovation Fellowship Korea 2025 Finalist Paper / arXiv / Code / Project 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. |
Auto-Encoded Supervision for Perceptual Image Super-Resolution
, Sangeek Hyun, Woojin Jun, Jae-Pil Heo† CVPR, 2025 Paper / arXiv / Poster / Code 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. |
Noise-free Optimization in Early Training Steps for Image Super-Resolution
, Jae-Pil Heo† AAAI, 2024 Paper / arXiv / Code This paper shows that early-stage SISR training with pixel-wise losses is overly affected by "inherent noise" in HR targets, and proposes a noise-free optimization scheme that estimates the optimal HR centroid and optimizes toward it, improving training stability and final super-resolution quality. |
Under Review |
Scalable GANs with Transformers
Sangeek Hyun, , Jae-Pil Heo† Project Page / arXiv 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. |
WavNAF: Learning Wave Propagation Priors for Neural Acoustic Fields
Taeho Kim, Hyunjun Kim, , SuBeen Lee, Inkyu An, Jae-Pil Heo† This paper proposes WavNAF, which derives acoustic parameter grids from NeRF-based scene geometry, uses FDTD-simulated pressure maps as wave-propagation priors, and learns a neural acoustic scaling module to efficiently reconstruct full-scale room impulse responses, improving room acoustics modeling over prior methods. |
Improving Sparse-View 3DGS Generalization via Flat Minima Optimization
Kangmin Seo, Sangeek Hyun, , Jae-Pil Heo† This paper improves sparse-view 3D Gaussian Splatting by adapting flat-minima optimization with scale-adaptive perturbations, stochastic per-Gaussian perturbation, and periodic parameter reinitialization, leading to sharper and more robust novel-view reconstructions without architectural changes. |
|
|