Jihwan Hong

News

[Sep. 26, 2026] My first-authored paper was accepted to NeurIPS 2026.
[Mar. 20, 2026] Our team achieved 3rd place in the PVUW MeViS-Audio Track at CVPR 2026.
[Mar. 9, 2026] Dynin-Omni, where I contributed to the video understanding component, was announced.
[Feb. 23, 2026] My first-authored paper was accepted to CVPR 2026.

Introduction

I am a student researching computer vision and machine learning at Seoul National University, in the Interdisciplinary Program in Artificial Intelligence.

My goal is to build systems that represent, understand, and generate complex visual worlds, and use these capabilities to solve real-world problems.

Education

M.S. in Artificial Intelligence (Interdisciplinary Program), Seoul National University (2025 – Present)
B.S. in Computer Science and Engineering, Korea University (2019 – 2025)

Research Experiences

Research Assistant, AIDAS Lab, Seoul National University, Korea (Aug. 2025 – Present)
Advisor: Prof. Jaeyoung Do
Research Intern, AIDAS Lab, Seoul National University, Korea (Jan. 2025 – Aug. 2025)
Advisor: Prof. Jaeyoung Do
Research Intern, MLV Lab, Korea University, Korea (Jun. 2023 – Dec. 2024; now at KAIST)
Advisor: Prof. Hyunwoo J. Kim
Research Intern, AI Lab, Korea University, Korea (Oct. 2020 – Jan. 2021)
Advisor: Prof. Dongsuk Yook

Publications

When Predicting Nothing Beats SAM 3: Revisiting Evaluation in Video Object Segmentation

Jihwan Hong, Woohyeon Park, Jaeik Kim, Jaeyoung Do^†

40th Annual Conference on Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track, 2026.

[Project] [Paper] [arXiv]

TL;DR

We identify a critical flaw in frame-wise J&F, the standard evaluation metric for video object segmentation: when target objects are rarely visible, predicting empty masks can outperform strong models such as SAM 3.
We introduce FaVOS to systematically evaluate this failure and propose volumetric J&F to address it.

Full figure comparing ground-truth video frames, Predict Nothing and SAM 3 masks and scores, explaining absence bias, and showing results on FaVOS-20.

2026

Dynin-Omni: Omnimodal Unified Large Diffusion Language Model

Jaeik Kim, Woojin Kim, Jihwan Hong, Yejoon Lee, Sieun Hyeon, Mintaek Lim, Yunseok Han, Dogeun Kim, Hoeun Lee, Hyunggeun Kim, Jaeyoung Do^†

arXiv preprint, arXiv:2604.00007, 2026.

[Project] [arXiv] [Code]

TL;DR

Dynin-Omni is an omnimodal discrete diffusion foundation model that unifies text, image, video, and speech understanding and generation through iterative refinement and bidirectional token modeling.

My role: I extended MMaDA's image understanding module to video, designing the architecture and training recipe for video understanding and video-to-speech generation.

2026

VIRST: Video-Instructed Reasoning assistant for SpatioTemporal Segmentation

Jihwan Hong, Jaeyoung Do†

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.

[Project] [Paper] [arXiv] [Code]

TL;DR

VIRST performs referring video object segmentation (RVOS), identifying and segmenting a target object throughout a video based on a text description.
By fusing visual features from the segmentation encoder and semantically aligned features into learnable tokens, alongside multiple anchor frames updated as the video progresses, VIRST achieves state-of-the-art results on all four evaluated RVOS benchmarks.

2026

Report of the 5th PVUW Challenge: Towards More Diverse Modalities in Pixel-Level Understanding

Chang Liu et al.

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Pixel-level Video Understanding in-the-Wild (PVUW), 2026.

[Paper] [arXiv]

2026