Building K-Contents Specialized Text2Image Pipeline

---

Research on Building K-Contents Specialized Text2Image Pipeline

Jul. 2024 ~ Aug. 2024

---

Team members

Name	Role	Email
Seungeun Kang	AI pipeline Development Leader	haun620@kyonggi.ac.kr
Dahyun Lee	Data collection and preprocessing, Lora training	idahyun22@kyonggi.ac.kr
Sangbum Han	Serving and Optimization pipeline	hsb422@kyonggi.ac.kr

Software Stacks

---

Project Motivation

1. Global interest in generative AI has increased

• Text2Image generative model is utilized in various fields and is used as an innovative tool in fields that require creativity

2. Recent "K-Contents Boom" has increased interest in Korean culture

• created a global demand for traditional Korean art, hanbok, Korean animation, etc...

We were interested in building a pipeline that understands the Korean language and generates Korean-specific images in this technological and cultural context

---

Implementation

1. Collect Dataset and Preprocessing

• for text dataset, we used AI Hub Dataset(Korean English Translation Corpus Data - Parallel, Technical Science, Social Science)

• for image dataset, we used AI Hub Dataset(Korean traditional ink painting production data by style , Character Face Landmark Data in Animation and Web Crawled Data

• Data Preprocessing

2. Korean CLIP Text Encoder

• With BaseModel(Bingsu/clip-vit-large-patch14-ko) ...
• We trained CLIP TextEncoder using (BaseModel, text datasets)

3. Pipeline Design

4. Experiments

---

Outputs

• Publication Conference Paper in The 27th International Conference on Advanced Communication Technology, (Feb. 2025)

• Demo Web Page