Taekyung Ki

I am a researcher at KAIST MLAI, working on generative models and computer vision. From March 2022 to March 2025, I conducted research on video generation as part of my mandatory military service in South Korea. I began working in deep learning in February 2021. I received my M.S. in Mathematics in February 2021 and my B.S. in Mathematics in February 2019.

I am interested in the following research topics:

Generative models (score models, flow models, one-step models, etc.)
Interactive visual avatar agent
Video generation
Audio-visual and vision-language representation

If you have an interest in the above topics, feel free to drop me an e-mail.

Email / Scholar / Twitter / GitHub / Hugging Face / LinkedIn

Highlights

[June 2025] One paper accepted to ICCV 2025!
[June 2025] New preprint out. Frame Guidance; Training-free frame-level guidance method for large video diffusion models
[Mar. 2025] Joined KAIST MLAI as a researcher
[July 2024] One paper accepted to ECCV 2024
[Mar. 2022 - Mar. 2025] Mandatory military service for South Korea

Publications

*: Equal contribution, †: Corresponding author

	Frame Guidance: Training-Free Guidance for Frame-Level Control in Video Diffusion Models Sangwon Jang^, Taekyung Ki^, Jaehyeong Jo, Jaehong Yoon, Soo Ye Kim, Zhe Lin, Sung Ju Hwang^† arXiv, 2025 Project Page / arXiv / Code / Hugging Face Frame Guidance is a training-free guidance method for large-scale video diffusion models, which enables to generate frame-level controllable videos only using a single GPU.
	FLOAT: Generative Motion Latent Flow Matching for Audio-driven Talking Portrait Taekyung Ki, Dongchan Min, Gyeongsu Chae International Conference on Computer Vision (ICCV), 2025 Project Page / arXiv / Code / Hugging Face FLOAT is a flow matching-based audio-driven talking portrait generation method that can automatically enhance speech-driven emotional motion and edit head motion at test time using an implicitly learned motion orthonormal basis.
	Learning to Generate Conditional Tri-plane for 3D-aware Expression Controllable Portrait Animation Taekyung Ki, Dongchan Min, Gyeongsu Chae European Conference on Computer Vision (ECCV), 2024 Project Page / Paper / arXiv / Supp We propose (1) CLeBS, a contrastive pre-training framework for appearance-free facial expression hidden in 3DMM expression parameters, and (2) Export3D, a 3D-aware, expression-controllable portrait animation method that leverages CLeBS and NeRF.
	StyleLipSync: Style-based Personalized Lip-sync Video Generation Taekyung Ki^, Dongchan Min^ International Conference on Computer Vision (ICCV), 2023 Project Page / Paper / arXiv / Code / Supp StyleLipSync can generate person-agnostic audio-lip synchronized videos by leveraging the strong facial prior of style-based generator.
	Deep Scattering Network with Max-pooling Taekyung Ki, Youngmi Hur^† IEEE 2021 Data Compression Conference (DCC), 2021 Paper / Code We mathematically prove that the pooling operator is a crucial component for translation-invariant feature extraction in Scattering Network.

Awards and Honors

[Oct. 2021] 1st Prize Winner, NLP-based Math-Word Problem Track in 2021 AI Grand Challenge (AGC), funded by South Korea Ministry of Science and ICT.
[Feb. 2019] Graduated with top honors, ranked 1st in the Department of Mathematics.

Academic Services

Reviewer: CVPR 2025, ICCV 2025.