Hello, I'm Jiwon Kang

I am an M.S. student at KAIST AI, advised by Seungryong Kim at CVLAB.

Having started my research career in 3D vision as an undergraduate researcher, I have focused on Visual Generative AI during my M.S. studies and published four (co-)first-author papers at NeurIPS, ICLR, CVPR, and ECCV.

My current research interests include image/video diffusion, world (action) models, and multimodal generation across images, video, audio, and action.

I am actively seeking M.S.-level Research Engineer/Scientist positions starting in February 2027, as well as Ph.D. internship opportunities of 2027.

CV Scholar Github Linkedin Email

News

Experience

Kakao, Seoul
Research Intern
  • Developed a Whisper-based continuous speech tokenizer and an autoregressive diffusion TTS system for high-dimensional speech representations.
Mar – Jun 2026

Projects

Samsung Electronics, Seoul
Research Collaboration
  • Developed a diffusion-based, attribute-preserving face-swapping framework for background face anonymization (APPLE).
Aug – Oct 2025

Publications

* denotes equal contribution.  ·  † denotes corresponding author.

2026

RAIN-GS: Relaxing Accurate Initialization Constraint for 3D Gaussian Splatting

RAIN-GS: Relaxing Accurate Initialization Constraint for 3D Gaussian Splatting

Jaewoo Jung*, Jisang Han*, Honggyu An*, Jiwon Kang*, Seonghoon Park*, Seungryong Kim†
WACV 2026 350+ Stars
  • Relaxed the accurate SfM initialization requirement of 3D Gaussian Splatting, enabling training from sub-optimal or random point clouds.
Project Page PDF arXiv Code
Transferability Between Understanding and Generation in Unified Multimodal Models

Transferability Between Understanding and Generation in Unified Multimodal Models

Jiwon Kang*, Heeji Yoon*, Jaewoo Jung, Jaewon Min, Minkyeong Jeon, Biyeon Hwang, Sangwon Jung, Seungryong Kim†
ECCV
  • Investigated how understanding and generation capabilities (e.g., counting) transfer to one another within unified multimodal models.
  • Showed that transferability depends on the model architecture, and proposed a practical post-training strategy for unified multimodal models that leverages transferability.
Project Page PDF arXiv Code
APPLE: Attribute-Preserving Pseudo-Labeling for Diffusion-Based Face Swapping

APPLE: Attribute-Preserving Pseudo-Labeling for Diffusion-Based Face Swapping

Jiwon Kang*, Yeji Choi*, JoungBin Lee, Wooseok Jang, Jinhyeok Choi, Taekeun Kang, Yongjae Park, Myungin Kim, Seungryong Kim†
CVPR
  • Prior diffusion-based face swapping methods focused on changing facial identity, failing to preserve facial attributes.
  • Proposed a novel pseudo-labeling framework that preserves facial attributes while swapping identities, achieving Pareto-optimal results.
Project Page PDF arXiv Code
A Noise is Worth Diffusion Guidance

A Noise is Worth Diffusion Guidance

Donghoon Ahn*, Jiwon Kang*, Sanghyun Lee, Jaewon Min, Wooseok Jang, Minjae Kim, Hyungwon Cho, Sayak Paul, SeonHwa Kim, Eunju Cha†, Kyong Hwan Jin†, Seungryong Kim†
ICLR
  • Showed that classifier-free guidance can be replaced by refining the initial noise.
  • Enabled high-quality generation without a guidance pass, highlighting the importance of the initial noise in diffusion models.
Project Page PDF arXiv Code

2025

Where and How to Perturb: On the Design of Perturbation Guidance in Diffusion and Flow Models

Where and How to Perturb: On the Design of Perturbation Guidance in Diffusion and Flow Models

Donghoon Ahn*, Jiwon Kang*, Sanghyun Lee, Minjae Kim, Wooseok Jang, Jaewon Min, Sangwu Lee, Sayak Paul, Seungryong Kim†
NeurIPS
  • Systematically studied perturbation guidance in diffusion and flow models.
  • Showed that perturbing different layers/heads in a diffusion model exhibits different effects on the generated samples, which can be systematically combined to improve user-specified attributes via the proposed framework.
Project Page PDF arXiv Code
Self-Evolving Neural Radiance Fields

Self-Evolving Neural Radiance Fields

Jaewoo Jung*, Jisang Han*, Jiwon Kang*, Seongchan Kim, Min-Seop Kwak, Seungryong Kim†
ICCVW
  • Proposed a self-training framework that lets a NeRF improve itself from few input views via pseudo-label distillation.
Project Page PDF arXiv Code