Yoojin Jang

I’m Yoojin Jang (장유진), an Integrated M.S./Ph.D. student at UNIST AIGS, advised by Prof. Jaejun Yoo at LAIT. My research focuses on multimodal generation and evaluation, particularly for audio and video. I am interested in developing multimodal generation and editing models, as well as benchmarks for evaluating their quality and cross-modal consistency.

Research Interests

Multimodal Generation & Editing: Audio–video–text consistent, modality-aware generation and editing
Data-Centric Learning & Robustness: Benchmark redesign, data imbalance, synthetic-based data, and robust model evaluation
Multimodal LLMs: Multi-event semantic understanding, temporal reasoning, and real-world multimodal comprehension

News

Apr 15, 2026	AVENUE (Audio-Video EditiNg Understanding and Evaluation) — co-first-authored with Hayeon Kim — accepted to the Learning to Listen workshop at ICML 2026.
Mar 20, 2026	“Rethinking Video-and-Text-to-Audio Generation through Multimodal Coverage” (2nd author) accepted to the Sight and Sound workshop at CVPR 2026.

International conference

ICMLW

🚀 AVENUE: Audio-Video EditiNg Understanding and Evaluation

Hayeon Kim^*, Yoojin Jang^*, and Jaejun Yoo

In Learning to Listen: ICML 2026 Workshop on Machine Learning for Audio, 2026

PDF Dataset arXiv (coming soon)
CVPRW

Rethinking Video-and-Text-to-Audio Generation through Multimodal Coverage

Sangyeop Yeo, Yoojin Jang, Saad Mourafik, Arda Senocak, and Jaejun Yoo

In Sight and Sound Workshop, CVPR, 2026
BMVC
B-RIGHT: Benchmark Re-evaluation for Integrity in Generalized Human-Object Interaction Testing

Yoojin Jang^*, Junsu Kim^*, Hayeon Kim, Eun-ki Lee, Eun-sol Kim, Seungryul Baek, and Jaejun Yoo

In British Machine Vision Conference (BMVC), 2025

Abs Bib HTML Code

HICO-DET mixes severe class imbalance with uneven train/test counts, so HOI scores quietly inflate or deflate. B-RIGHT balances every class through automated generation-and-filtering and adds a zero-shot split — re-evaluation cuts variance and shifts rankings.
@inproceedings{jang2025bright, title = {B-RIGHT: Benchmark Re-evaluation for Integrity in Generalized Human-Object Interaction Testing}, author = {Jang, Yoojin and Kim, Junsu and Kim, Hayeon and Lee, Eun-ki and Kim, Eun-sol and Baek, Seungryul and Yoo, Jaejun}, booktitle = {British Machine Vision Conference (BMVC)}, year = {2025}, }
ECCV
Nickel and Diming Your GAN: A Dual-Method Approach to Enhancing GAN Efficiency via Knowledge Distillation

Sangyeop Yeo, Yoojin Jang, and Jaejun Yoo

In European Conference on Computer Vision (ECCV), 2024

Abs arXiv Bib HTML Code

DiME distills GANs by matching distributions in foundation-model embeddings via MMD; NICKEL keeps the student generator and discriminator talking through interactive compression. Together they reach FID 10.45 at 95.73% compression on FFHQ and still hold up at 99.69%.
@inproceedings{yeo2024nickel, title = {Nickel and Diming Your GAN: A Dual-Method Approach to Enhancing GAN Efficiency via Knowledge Distillation}, author = {Yeo, Sangyeop and Jang, Yoojin and Yoo, Jaejun}, booktitle = {European Conference on Computer Vision (ECCV)}, year = {2024}, }
NeurIPS
TopP&R: Robust Support Estimation Approach for Evaluating Fidelity and Diversity in Generative Models

Pum Jun Kim, Yoojin Jang, Jisu Kim, and Jaejun Yoo

Advances in Neural Information Processing Systems (NeurIPS), 2023

Abs arXiv Bib HTML Code

FID, IS, and precision-recall all rest on supports estimated from sample features, but nobody checks whether those supports are reliable. TopP&R keeps only topologically and statistically significant features, giving fidelity and diversity scores stable under outliers and non-IID noise.
@article{kim2023toppr, title = {TopP\&R: Robust Support Estimation Approach for Evaluating Fidelity and Diversity in Generative Models}, author = {Kim, Pum Jun and Jang, Yoojin and Kim, Jisu and Yoo, Jaejun}, journal = {Advances in Neural Information Processing Systems (NeurIPS)}, year = {2023}, }
AAAI
Can We Find Strong Lottery Tickets in Generative Models?

Sangyeop Yeo, Yoojin Jang, Jy-yong Sohn, Dongyoon Han, and Jaejun Yoo

In Proceedings of the AAAI Conference on Artificial Intelligence, 2023

Oral Abs Bib HTML Code

Oral Presentation

Strong lottery tickets ask whether a random network already contains a working subnetwork, no weight updates allowed. We bring this to generators: only a binary mask is learned over an untrained model, guided by moment-matching scores (MMD). With 10% of weights left, it matches the dense baseline.
@inproceedings{yeo2023slt, title = {Can We Find Strong Lottery Tickets in Generative Models?}, author = {Yeo, Sangyeop and Jang, Yoojin and Sohn, Jy-yong and Han, Dongyoon and Yoo, Jaejun}, booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence}, year = {2023}, }