bumsookim.info

Bumsoo Kim
"Innovating entertainment market with machine learning"

I'm a Generative AI Researcher grounded at the intersection of Computer Graphics and Neural Networks.

I received my B.S. degree (double major in Art & Technology and Electronic and Electric Engineering) from Chung-Ang University, Korea, in 2023. Since March 2024. I have been pursuing M.S. degree in Applied Art & Technology at Chung-Ang University. I also plan to pursue a Ph.D. degree in a related field.

Email / CV / Failure CV / Short Bio / Linkedin / Google Scholar / Github

Experience

Boeing, Research and Technology, Korea - ML Research Intern, 2024
VIVE STUDIOS, R&D Center, Korea - Generative AI Research Intern, 2023
ETRI, Korea - Research Trainee, 2021

Paper Summary

NeurIPS'23 Workshop - 1 Paper (1st Author) : Oral (Highlight, Top25%) + Poster , on-site presented on New Orleans, USA
CVPR'24 Workshop - 1 Paper (1st Author) : Poster , on-site presented on Seattle, USA
WACV'25 - 1 Paper (Co-1st Author): Poster , on-site presented on Tucson, USA
AAAI'25 Workshop - 1 Paper (Co-1st Author): Oral (best paper candidate) + Poster , on-site presented on Philadelphia, USA

Ongoing Projects

Representative papers are highlighted

This is my MS 4th-semester project! (Gaussian Splatting)

TBD

Bumsoo Kim, Sanghyun Seo

(top-tier CG Conference), 2025.
(Will be Submitted)

(2.5D Representation)

This is my MS 3rd-semester project! (Diffusion + NeRFs)

Learning Pose-Guided Multi-Semantics for 3D Anamorphic Art

Bumsoo Kim, Sanghyun Seo

CVPR 2026, (top-tier CS/AI Conference), 2026.
(Submitted)

Diffusion-based textured 3D mesh generation system for 3D anamorphic art

This is my additional project. (Diffusion + Gaussian Splatting)

Sounds That Shape: Audio-Driven 3D Mesh Generation with Attribute-Decoupled Score Distiilation Sampling

Bumsoo Kim, Sanghyun Seo

ICASSP 2026, (top-tier CS/AI Conference), 2026.
(Under review)

[Project Page]

In this paper, we explore intriguing yet challenging task, audio-driven 3D generation system for mesh and UV texture. To sidestep the large cost of dataset requirements of audio-3D labels and new architectural design, we harness the SDS framework built on audio-to-image diffusion models. Moreover, we introduce attribute-decoupled guidance on 3D Gaussian splatting, whereby each property take the appropriate objective from separate gradients for more reasonable object shape. Our system enables to reconstructe explicit 3D object from a single audio cue and/or additional text.

This is my MS 2nd-semester project! (Gaussian Splatting)

Unveiling Importance of Trajectory in Few-Shot Gaussian Splatting

Bumsoo Kim, Abdul Muqeet, Yoonje Kang, Sanghyun Seo

(In-Progress)

In this paper, we reveal the importance of training view selection to optimize Gaussian blobs for better rendering quality in few-shot situation like 3-views. Instead of index-linearly selecting few training views as in existing works (e.g., FSGS, DNGaussian, CoR-GS), we introduce training view selection strategy based on uncertainty modeling using Baysian modeling.

Scent2Scene: A Participatory Olfacotry-to-Visual Experience Using Brain-Computer Interface and Generative AI

Rakbin Sung, Seongmi Woo, Dongmin Shin, Bumsoo Kim, Yonghoon Jung, Daewon Lee†, Sanghyun Seo†, Hye-Young Jo†, Yoonji Kim†

2025.
(Wait to submit)

In this paper, we examines how scene can be visualized as art using brain-computer interface and generative artificial intelligence.

This is my MS 1st-semester project! (Diffusion Models)

VideoTimeTravel: High-Fidelity Face Re-Aging Diffusion Models For Production Video

Bumsoo Kim, Sanghyun Seo

(Visual Computer, IF.2.9) 2025.
(Revision step)

[Project Page]

For video-level face re-aging with temporal consistency and high fidelity, we proposed a delta-based video face re-aging network by leveraging pre-trained diffusion models. It enables to conduct face re-aging on a variety of video conditions like CGI/VFX, NPR, movie trailer and so on. To avoid costly computing power of 3D convolutional kernel or temporal attention module, we bootstrap image-level diffusion model to video inference with delta strategy and optical flow estimation, demonstrating high fidelity editing performance in few minutes.

Video Face Re-Aging: Toward Temporally Consistent Face Re-Aging
Abdul Muqeet*, Kyuchul Lee*, Bumsoo Kim, Yohan Hong, Hyungrae Lee, Woonggon Kim, Kwang Hee Lee. (*equal contributors)

preprinted

[ArXiv], [Project Page]

For video-level face re-aging, we proposed a video-supervised learning approach with recursive scheme. In addition, we released video paired aging dataset with large facial motion and expression and two evaluation metric, TL/TG-Age.

Neural Mapper for Better Downstream Task with 2D Unfolded Texture

(In-Progress)

To encourage downstream models to learn unwrapped 2D textures without UV mapping information, an explicit flow-mapper is proposed to transform 2D textures into an optimal structure from the model perspective. This mapper helps the generative model in downstream tasks learn any textures well without performance degradation.

Eliminating the Screen Door Effect: A Cost-Effective LED Wall Solution for Virtual Production
Abdul Muqeet*, Bumsoo Kim*, Kyuchul Lee, Gwanghee Lee. (*equal contributors)

To mitigate the visual Screen Door Effect (SDE) problem that appears in images captured through LED walls in virtual production (VP) pipeline, we propose a learning-based SDE removal method with SDE supervision dataset.

Finished Projects

Representative papers are highlighted
Summary
- 6 SCIE
- 3 Top-tier Conf. Workshop
- 1 Top-tier Conf.

MultiFloodSynth: Multi-Annotated Flood Synthetic Dataset Generation
Yoonje Kang*, Yonghoon Jung*, Wonseop Shin*, Bumsoo Kim*, Sanghyun Seo. (*equal contributors)

The Association for the Advancement of Artificial Intelligence(AAAI) Workshop on Preparing Good Data for Generative AI(Good-Data), [Best Paper Condidate, Top30% (Oral Presentation)], (top-tier CS/AI Conference), 2025.

[AAAI 2025 Good-Data], [Paper]

We propose editable synthetic dataset generation pipeline for flood hazard detection situation and build faithful flood synthetic data, dubbed FloodSynthSet . For flood level recognition, we provide several control parameters which composite the final flood hazard environments in virtual simulator. Considering a variety of downstream task, our FloodSynthSet contains multiple annotation types e.g., normal map, 3D bounding box, camera 3D position, fine-grained/instance segmentation, and so on. In experiments, our synthetic dataset enhance the detection performance of flood-level than real-data training one.

Make VLM Recognize Visual Hallucination on Cartoon Character Image with Pose Information
Bumsoo Kim*, Wonseop Shin*, Kyuchul Lee, Yonghoon Jung, Sanghyun Seo. (*equal contributors)

Winter Conference on Applications of Computer Vision(WACV), (top-tier CS/AI Conference), Tucson, US, 2025.

[Project Page], [WACV 2025 Proceedings]

We proposed new approach to detect visual structural hallucination on cartoon character images which are generated by text-to-image model. With few samples, we harness the emerging capabilities of VLM, in-context visual learning with pose-guidance. In experiments, our proposition improved the detection performance on visual hallucination of cartoon character in two VLMs without weight update or separate training step.

Explicitly Color-Inspired Neural Style Transfer Using Patchified AdaIN
Bumsoo Kim, Wonseop Shin, Younghoon Jung, Youngsup Park, Sanghyun Seo.

Computer Modeling in Engineering & Sciences (SCIE, Q2), 2024

We proposed PatchizedAdaIN operation for color-aware arbitrary style transfer. It can preserve regional color distribution without any additional computational overhead.

ToonAging: Face Re-Aging upon Artistic Portrait Style Transfer
Bumsoo Kim, Abdul Muqeet, Kyuchul Lee, Sanghyun Seo.

Conference on Computer Vision and Pattern Recognition(CVPR) Workshop on AI for Content Creation (AI4CC), (top-tier CS/AI Conference), Seattle, US, 6/17, 2024.
[Paper], [Poster]

For face re-aging with portrait style transfer, we devise 1-stage manner based on StyleGAN architecture.

Harnessing Synthetic Data for Enhanced Detection of Pine Wilt Disease: A Imag Classification Approach
Yonghoon Jung, Sanghyun Byun, Bumsoo Kim, Sareer UI Amin, Sanghyun Seo.

Computers and Electronics in Agriculture 2024. (SCIE, IF8.3, Q1 : 0.9%)
[Paper]

To address data requirements problem and improve the classification performance, we leverage the synthetic dataset while designing specific virtual world in game engine, i.g., Unreal5.

Transfer Learning based Parameterized 3D Mesh Deformation with 2D Stylized Cartoon Character
Sanghyun Byun, Bumsoo Kim, Wonseop Shin, Younghoon Jung, Sanghyun Seo.

KSII TRANSACTIONS ON INTERNET AND INFORMATION SYSTEMS, VOL. 17, NO. 11, (SCIE), 2023, [Paper]

We proposed mesh deformation system with 2D stylized cartoon character.

Minecraft-ify: Minecraft Style Image Generation with Text-guided Image Editing for In-Game Application
Bumsoo Kim, Sanghyun Byun, Younghoon Jung, Wonseop Shin, Sareer UI Amin, Sanghyun Seo.

Conference on Neural Information Processing Systems(NeurIPS) Workshop on Machine Learning for Creativity and Design(ML4CD), [Spotlight, Top25% (Oral Presentation)], (top-tier CS/AI Conference), New Orleans, US, 12/16, 2023.

[NeurIPS 2023 ML4CD], [Paper], [Code(Soon)], [Slide, Presenter: Bumsoo Kim], [Poster], [Project Page]

For Minecraft-World texture generation, we proposed user-friendly AI-tool which can generate frontal face texture and edit it by text-guidance

EfficientNet-based Dynamic Gesture Recognition using transformed Scalogram from Tri-Axial Acceleration Signal
Bumsoo Kim, Sanghyun Seo.

Journal of Computational Design and Engineering, 2022 (SCIE, IF3.42, Q1 : 8.4%)

Tri-axial acceleration singal-based scalogram classification system for dynamic gesture recognition.

Deep Learning Based Active Learning Technique for Data Annotation and Improve the Overall Performance of Classification Models
Sareer UI Amin, Adnan Hussain, Bumsoo Kim, Sanghyun Seo.

Expert Systems with Application, 2023 (SCIE, IF8.665, Q1 : 14.4%)

While focusing the high-confidence unannotated set with significant number of instances, we proposed DLBAL method, in contrast to previous machine learning based Active Learning system.

Intelligent Digital Human Agent Service with Deep Learning Based-Face Recognition
Bumsoo Kim, Sanghyun Seo.

IEEE Access, 2022 (SCIE, IF3.46)

Personalized service using individual face detection via intelligent digital human in the application aspects.

Extracurricular activities

Stanford CS230: Deep Learning | Autumn 2018 | Andrew Ng
Stanford CS236: Deep Generative Models | 2023 | Stefano Ermon
MIT 6.S184: Flow Matching and Diffuison Models | 2025 | Peter Holderrieth, Playlist

Learned mathematics

College-level: Calculus, Linear Algebra, Random Variables, Numerical Analysis, Topology,

Studied book

'Pattern Recognition and Machine Learning', Christopher M. Bishop, Springer, 2008.
'Deep Learning', Ian Goodfellow, Yoshua Bengio and Aaron Courvile, MIT Press (MA), 2016.
'Mathematics for Machinc Learning', Deisenroth, Marc Peter, Faisal, A. Aldo, Ong, Cheng Soon, Cambridge University Press, 2020.
'Probabilistic Machine Learning - An Introduction', Kevin P. Murphy, MIT Press (MA), 2021.
'Computer Vision: Algorithms and Applications - 2nd Edition', Richard Szeliski, Springer, 2022.
'GPU Gems 1/2/3', NVIDIA, Hubert Nguyen,

Page updated

Google Sites

Report abuse