Research Highlights
Peer-Reviewed Conference Papers 12 total • 9 as (co-)first author (2 also co-corresponding) • 3 as co-author
- 2026 (6): ICLR (Oral, Top 1.13%) (co-first), ACCV (Oral) (first author; also Spotlight Presentation at NeurIPS 2026 Workshop ASCI), ICML (co-author), ECCV ×2 (co-first & co-corresponding), MICCAI (co-author)
- 2025 (2): ICCV (co-first), ICCV (co-author)
- 2024 (2): ECCV (first author), ECCV (co-first)
- 2023 (1): ICCV (first author)
- 2021 (1): ICIP (first author)
Papers Under Review 5 total • 2 as (co-)first author (1 also co-corresponding) • 2 as co-corresponding author • 1 as co-author
Academic Service 11 conferences • 2 journals • continuous since 2022
- Outstanding Reviewer, CVPR 2025 (top 5.6% of 12,593 reviewers)
- Reviewer for NeurIPS, ICLR, ICML, AAAI, CVPR, ICCV, ECCV, and IJCV
Patents & Industry Experience
- 9 years of industrial AI R&D at OGQ (2017 – Present), currently Principal AI Researcher
- 5 registered and 22 filed Korean patents, including two from the Seoul National University R&DB Foundation (2026)
Biography
My research works toward Physical AI: vision systems that perceive the physical world, let people interact with what is perceived, and remain deployable under real-world constraints. I pursue three connected axes, treating them not as separate topics but as the perception, generation, and deployment layers of a single pipeline: segmentation, extending from images to video and 3D; interactive segmentation and editing; and multimodal distillation and quantization. Across all three, my approach is to exploit the implicit knowledge already embedded in large-scale foundation models, using generative priors and lightweight interaction to remove the annotation and optimization burdens that make vision systems expensive to build and maintain.
I have been with OGQ since 2017, currently serving as the Principal AI Researcher and leading industrial AI research and development to maximize tangible real-world impact. In parallel, I have pursued my research under the guidance of Prof. Kyungsu Kim since 2021; following his appointment at Seoul National University, I joined the SNU AIBL Lab as an Affiliated Researcher in September 2024, where I mentor graduate and undergraduate researchers and lead core projects as a co-corresponding author. Integrating industrial R&D leadership with academic rigor, I aim to develop machine learning systems that are both scientifically grounded and practically deployable.
Research Interests
Toward Physical AI: perception, interaction, and efficient deployment of vision systems
- Core & Biomedical Vision AI2025 – Present
- Multimodal Distillation & Quantization2025 – Present
- Interactive Segmentation & Editing2023 – Present
- Image, Video & 3D Perception2021 – Present
Appointments
Education
Publications: Conference Papers
12 peer-reviewed papers: 9 as (co-)first author (2 also co-corresponding), 3 as co-author. Venues: ICLR (1 Oral, Top 1.13%), ACCV (1 Oral), ICML (1), ECCV (4), ICCV (3), MICCAI (1), ICIP (1).
- [C1]Sanghyun Jo*, Ziseok Lee*, Wooyeol Lee, Jonghyun Choi, Jaesik Park†, Kyungsu Kim†.“TRACE: Your Diffusion Model is Secretly an Instance Edge Detector”ICLR 2026 [Oral, Top 1.13% (223 papers selected out of 19,735 submissions)].PDF OpenReview Code Project [Sanghyun Jo: The co-first author*]
- [C2]Sanghyun Jo, Seo Jin Lee, Seohyung Hong, Yoorim Gang, Hyeongsub Kim, Hyungseok Seo†, Kyungsu Kim†.“One Click per Cell Type Suffices: Training-free Group Interaction for Cell Instance Segmentation”ACCV 2026 [Oral]; also selected for [Spotlight Presentation, Top-10] at NeurIPS 2026 Workshop ASCI (non-archival).
- [C3]Ziseok Lee*†, Minyeong Hwang*, Wooyeol Lee, Sanghyun Jo, Jihyung Ko, Young Bin Park, Jae-Mun Choi, Eunho Yang†, Kyungsu Kim†. “On the Collapse of Generative Paths: A Criterion and Correction for Diffusion Steering”ICML 2026 [Acceptance Rate: 26.6% (6,352 accepted out of 23,918 submissions)].PDF OpenReview Code Project [Sanghyun Jo: The co-author]
- [C4]Sanghyun Jo*†, Wooyeol Lee*, Ziseok Lee*, Jonghyun Choi, Jaesik Park, Kyungsu Kim†.“ISAC: Training-Free Instance-to-Semantic Attention Control for Multi-Instance Generation”ECCV 2026 [Acceptance Rate: 27.5% (2,883 accepted out of 10,473 submissions)].
- [C5]Sanghyun Jo*†, Donghwan Lee*, Eunji Jung*, Seong Je Oh, Kyungsu Kim†. “EraseLoRA: MLLM-Driven Foreground Exclusion and Background Subtype Aggregation for Dataset-Free Object Removal”ECCV 2026 [Acceptance Rate: 27.5% (2,883 accepted out of 10,473 submissions)].
- [C6]Hyeongsub Kim, Sihyun Kim, Minyoung Cho, Sanghyun Jo, Minhyeong Lee, Kyungsu Kim†.“GeneRAG: A Retrieval-Augmented Framework for Spatially Resolved Gene Expression Prediction”MICCAI 2026 [Acceptance Rate: 26.3% (1,211 accepted out of 4,601 submissions)].
- [C7]Sanghyun Jo*, Seo Jin Lee*, Seungwoo Lee, Seohyung Hong, Hyungseok Seo†, Kyungsu Kim†.“COIN: Confidence Score-Guided Distillation for Annotation-Free Cell Segmentation”ICCV 2025 [Acceptance Rate: 24.0% (2,698 accepted out of 11,239 submissions)].
- [C8]Joowon Kim*, Ziseok Lee*, Donghyeon Cho, Sanghyun Jo, Yeonsung Jung, Kyungsu Kim†, Eunho Yang†.“Early Timestep Zero-Shot Candidate Selection for Instruction-Guided Image Editing”ICCV 2025 [Acceptance Rate: 24.0% (2,698 accepted out of 11,239 submissions)].
- [C9]Sanghyun Jo, Fei Pan, In-Jae Yu, Kyungsu Kim†. “DHR: Dual Features-Driven Hierarchical Rebalancing in Inter- and Intra-Class Regions for Weakly-Supervised Semantic Segmentation”ECCV 2024 [Acceptance Rate: 27.9% (2,395 accepted out of 8,585 submissions)].
- [C10]Sanghyun Jo*, Soohyun Ryu*, Sungyub Kim, Eunho Yang, Kyungsu Kim†.“TTD: Text-Tag Self-Distillation Enhancing Image-Text Alignment in CLIP to Alleviate Single Tag Bias”ECCV 2024 [Acceptance Rate: 27.9% (2,395 accepted out of 8,585 submissions)].
- [C11]Sanghyun Jo, In-Jae Yu, Kyungsu Kim†. “MARS: Model-agnostic Biased Object Removal without Additional Supervision for Weakly-Supervised Semantic Segmentation”ICCV 2023 [Acceptance Rate: 26.7% (2,160 accepted out of 8,088 submissions)].
- [C12]Sanghyun Jo, In-Jae Yu†.“Puzzle-CAM: Improved Localization via Matching Partial and Full Features”ICIP 2021 [Acceptance Rate: 45.9% (791 accepted out of 1,722 submissions)].
Conference Papers Under Review
- [U1]
- [U2]Donghwan Lee*, Sanghyun Jo*, Sihyun Kim, Ziseok Lee, Kyungsu Kim†. “ReState: Noise-Level-Aware State Correction for Training-Free Object Removal in Few-Step Diffusion Models”, Under Review.PDF [Sanghyun Jo: The co-first author*]
- [U3]
- [U4]Jihyung Ko*, Eunji Jung*, Hyeongsub Kim, Ziseok Lee, Jae Won Cho, Sanghyun Jo†, Kyungsu Kim†. “PatchGate: Narrowing the Verbalization Gap with Intrinsic Object Inventories in Frozen Vision-Language Models”, Under Review.PDF [Sanghyun Jo: The co-corresponding†]
- [U5]Chae Yeon Lim*†, JaeHyeon Woo*, Sanghyun Jo, Kyungsu Kim†. “LiftConsensus: Lifting 2D Change Evidence into 3D via Multi-View Spatial Consensus for Retail Change Detection”, Under Review.[Sanghyun Jo: The co-author]
Preprint
- [1]
Awards & Honors
- Outstanding Reviewer, CVPR 2025Jun. 2025Selected as top 5.6% (711 of 12,593 reviewers).
-
1st place out of 1,153 participants. Developed case representation and multi-branch modeling for civil court ruling prediction.
-
5th place out of 1,747 participants. Built a robust vehicle detector under domain shift.
Academic Service
Reviewer for 11 international conferences and 2 journals since 2022, including NeurIPS, ICLR, ICML, AAAI, CVPR, ICCV, ECCV, and IJCV. Recognized as an Outstanding Reviewer at CVPR 2025 (top 5.6% of 12,593 reviewers).
- 2027: ICLR, AAAI
- 2024: ICPR, ICIP
- 2023: ICIP
- 2022: ICML, ICIP
Journals: IJCV (2024–2026), IEEE TCSVT (2022)
Invited Talks & Presentations
- Yonsei University, Seoul, KoreaJun. 2026Vision AI Already Knows: Understanding the World Without Labels
- ICLR 2026, Rio de Janeiro, BrazilApr. 2026TRACE Oral Presentation (Oral Session 5B: Video and Scene Generation)
- MODULABS, Seoul, KoreaDec. 2024DHR & TTD Recap: Sharing Our ECCV 2024 Work with Local Researchers
- Yonsei University, Seoul, KoreaDec. 2023Advanced AI Developments
- Google Developer Groups (GDG), Incheon, KoreaOct. 2023Strategies for Reducing Labeling Costs in Vision AI
- MODULABS, Seoul, KoreaNov. 2022Experience Sharing Session: Vision Research
- MODULABS, Seoul, KoreaOct. 2022RSEPM Paper Presentation
Patents
5 registered and 22 filed Korean patents covering image segmentation, image editing and synthesis, vision-language and large language models, de-identification, and gene expression prediction; 25 with OGQ as the applicant and 2 with the Seoul National University R&DB Foundation. Titles and inventor order follow the official records, except where noted. Linked titles open the public KIPRIS record; applications without a link are not yet published.
Registered (5)
- [P1]Sanghyun Jo, Kyungsu Kim, Seungon Bang. “Method for Background Composition of Error Object Area Performed on a Computing Device and Computing Device for Performing the Same”, Korean Patent No. 10-3048915, Nov. 2026 (Application No. 10-2024-0198644, Dec. 2024).
- [P2]Sanghyun Jo, Kyungsu Kim, Seungon Bang. “Dual Feature-Based Image Segmentation Method Performed by a Computing Device and Computing Device for Performing the Same”, Korean Patent No. 10-3036286, Oct. 2026 (Application No. 10-2024-0171260, Nov. 2024). The filed English title is shared with [P22]; the Korean title describes an image retrieval-based inpainting method.
- [P3]Kyungsu Kim, In-Jae Yu, Sanghyun Jo, Seungon Bang. “Method of Acquiring Object Segmentation Information Through Trained Neural Network and Server System Performing the Same Method”, Korean Patent No. 10-2916563, Jan. 2026 (Application No. 10-2022-0141458, Oct. 2022). Also filed as PCT/KR2023/005062 (WO 2024/090692).
- [P4]Seungon Bang, Sanghyun Jo. “Method for Safety Management by Object Segmentation Based on Deep-Learning and System for the Method”, Korean Patent No. 10-2831465, Jul. 2025 (Application No. 10-2021-0164131, Nov. 2021).
- [P5]Kyungsu Kim, In-Jae Yu, Sanghyun Jo, Seungon Bang. “Object De-Identification Method Through Segmentation and Server That Performs the Same Method”, Korean Patent No. 10-2769724, Feb. 2025 (Application No. 10-2023-0164620, Nov. 2023).
Filed (20, OGQ)
- [P6]Seungon Bang, Kyungsu Kim, Sanghyun Jo. “Method for Real-Time Data Retrieval and Question Answering Based on Function Calling”, Korean Patent Application No. 10-2026-0059440, Apr. 2026.
- [P7]Seungon Bang, Kyungsu Kim, Sanghyun Jo. “Method for Structuring Unstructured Data and Automatically Storing in Knowledge Graph Based on Large Language Model”, Korean Patent Application No. 10-2026-0059432, Apr. 2026.
- [P8]Seungon Bang, Kyungsu Kim, Sanghyun Jo, Seungjun Jo. “Vision Language Model-Based Image Description Generation Method Using Representative-Tag-Based Embedding Alignment and Information Compression”, Korean Patent Application No. 10-2025-0212590, Dec. 2025.
- [P9]Seungon Bang, Kyungsu Kim, Sanghyun Jo, Seungjun Jo. “Intelligent Video Security Monitoring Method and System Using Dual-Feature Hierarchical Refinement and Question-Answer-Based Reasoning”, Korean Patent Application No. 10-2025-0212568, Dec. 2025.
- [P10]Seungon Bang, Kyungsu Kim, Sanghyun Jo, Seungjun Jo. “Vision-Language Model Based Video Security Monitoring Method Performed by a Computing Device”, Korean Patent Application No. 10-2025-0212529, Dec. 2025.
- [P11]Seungon Bang, Kyungsu Kim, Sanghyun Jo, Seungjun Jo. “Segmentation-Based Object Removal and Virtual Advertisement Inpainting System and Method”, Korean Patent Application No. 10-2025-0184690, Nov. 2025.
- [P12]Seungon Bang, Kyungsu Kim, Sanghyun Jo, Seungjun Jo. “Method and System for Detecting Erroneous Objects and Extracting Error Frames Based on Tagged Images”, Korean Patent Application No. 10-2025-0184673, Nov. 2025.
- [P13]Seungon Bang, Kyungsu Kim, Sanghyun Jo, Seungjun Jo. “Weakly and Unsupervised Learning-Based Image Segmentation Method and System for PPL Compliance Analysis”, Korean Patent Application No. 10-2025-0184663, Nov. 2025.
- [P14]Seungon Bang, Kyungsu Kim, Sanghyun Jo, Seungjun Jo. “Method and System for Generating Advertising Images Using Instruction-Driven Diffusion Models With Early Seed Selection and Interface-Based User Input”, Korean Patent Application No. 10-2025-0184639, Nov. 2025.
- [P15]Seungon Bang, Kyungsu Kim, Sanghyun Jo, Seungjun Jo. “News Breaking Thumbnail and Video Generation System Using Representative-Tag-Based Object and Background Recommendation”, Korean Patent Application No. 10-2025-0184621, Nov. 2025.
- [P16]Seungon Bang, Kyungsu Kim, Sanghyun Jo. “Confidence Score-Guided Self-Distillation Method for Annotation-Free Instance Segmentation”, Korean Patent Application No. 10-2025-0162415, Oct. 2025.
- [P17]Seungon Bang, Kyungsu Kim, Sanghyun Jo. “Instruction-Guided Image Editing Method Using Early-Stopped Background Inconsistency Score Evaluation”, Korean Patent Application No. 10-2025-0162189, Oct. 2025.
- [P18]Seungon Bang, Kyungsu Kim, Sanghyun Jo. “Instance-Aware Edge Map Generation Without Pixel Annotations Using a Diffusion Model and Segmentation Mask Refinement Method Based Thereon”, Korean Patent Application No. 10-2025-0162165, Oct. 2025.
- [P19]Kyungsu Kim, In-Jae Yu, Sanghyun Jo, Seungon Bang. “Object De-Identification Method Through Segmentation and Server That Performs the Same Method”, Korean Patent Application No. 10-2025-0018458, Feb. 2025 (divisional of [P5]).
- [P20]Sanghyun Jo, Kyungsu Kim, Seungon Bang. “Method for Synthesizing Virtual Objects for Virtual Advertising Performed by a Computing Device and Computing Device for Performing the Same”, Korean Patent Application No. 10-2024-0198877, Dec. 2024.
- [P21]Sanghyun Jo, Kyungsu Kim, Seungon Bang. “Dual Feature-Based Image Segmentation Method Performed by a Computing Device and Computing Device for Performing the Same”, Korean Patent Application No. 10-2024-0171236, Nov. 2024. The filed English title is shared with [P22]; the Korean title describes an image segmentation method using tag-based activation maps.
- [P22]Sanghyun Jo, Kyungsu Kim, Seungon Bang. “Dual Feature-Based Image Segmentation Method Performed by a Computing Device and Computing Device for Performing the Same”, Korean Patent Application No. 10-2024-0171211, Nov. 2024.
- [P23]Sanghyun Jo, Kyungsu Kim, Seungon Bang. “Method for Image Tagging Based on Image Matching Performed by a Computing Device and Computing Device for Performing the Same”, Korean Patent Application No. 10-2024-0171187, Nov. 2024.
- [P24]Kyungsu Kim, In-Jae Yu, Sanghyun Jo, Seungon Bang. “Debiased Segmentation Method and Server That Performs the Same Method”, Korean Patent Application No. 10-2023-0164587, Nov. 2023.
- [P25]Kyungsu Kim, In-Jae Yu, Sanghyun Jo, Seungon Bang. “Obtaining Method of Object Mask Information for Public Safety and Server System Performing the Same”, Korean Patent Application No. 10-2022-0176164, Dec. 2022.
Filed (2, Seoul National University R&DB Foundation)
- [P26]Kyungsu Kim, Jihyung Ko, Eunji Jung, Sanghyun Jo, Hyeongsub Kim, Ziseok Lee. “Apparatus and Method for Mitigating Object Hallucination of Vision-Language Models for Images”, Korean Patent Application No. 10-2026-0204604, Oct. 2026, based on the PatchGate paper [U4].
- [P27]Hyeongsub Kim, Sihyun Kim, Minyoung Cho, Sanghyun Jo, Minhyeong Lee, Kyungsu Kim. Korean Patent Application No. 10-2026-0138802, Jul. 2026, based on the GeneRAG paper [C6]. Inventors are listed in the paper’s author order; the filed title and inventor order will be added once confirmed from the filing.