Projects

LUCiD: AI Animation Trailer

LUCiD is an animation where everything, except for the editing, has been created using AI: images, videos, music, sound effects, and even voices. I utilized Stable Diffusion XL for images, Luma AI for videos, Suno AI for music, and MetaMind Creative AI along with some Freesound CC0 licensed audio for sound effects. Read more

Byte: AI Animate Digital Comic

Byte is an animate digital comic artwork that embodies a poignant reflection on humanity's future amidst the transformative wave of AI, both in content and form. The title, a nod to data measurement units, cleverly echoes the familiar canine action of 'biting', weaving a rich tapestry of meanings. Read more

Pet AI Profile Image Generator

Co-developed and launched Hellobot Pet AI Profile generator as two-person team, releasing it both domestically and globally. Oversaw the entire development of Pet AI Profile 3: National Athlete Edition, Pet AI Profile 4: Career Edition, and Pet AI Profile 5: Happy Lunar New Year Edition which generates anthropomorphized images of pets. Significantly contributed to profits by achieving sales of $100,000 in just three months.

Parts Demand Forecasting and Cost Optimization

As the leader of Team OpsIEdian in the 3rd POSTECH-UNIST-KAIST Data Science Competition, achieved 3rd place, winning a gold prize. Working with real data from LG Electronics, forecasted product parts demand using decision tree-based ensemble models and addressed inventory management problems with Auto ARIMA and heuristic decision-making. In the final round, our presentation discussed the strengths and weaknesses of the techniques attempted, suggestions for improvements in terms of data, model, interface, and evaluation metrics, and submitted a 58-page report.

Movie-to-Webtoon Project

Created movie prologue webtoon based on Barbie trailer using ControlNet and LoRA. Demonstrated at Artificial Intelligence Graduate School Symposium thingsflow booth.

Web Novel Episode 1 Generator

Demonstrated web novel generator at Artificial Intelligence Graduate School Symposium thingsflow booth, which produces covers and illustrations that fixed character traits via Low-Rank Adaptation (LoRA) Fine-Tuning.

AI QR Image Generator

Generated QR codes that fuse celebrity or character images using ControlNet and LoRA. QR image is natural and can be scanned with QR reader app to reveal hidden QR codes.

Developed Ball Jointed Doll Tangible User Interface (TUI) based on Stable Diffusion

Authored a research proposal that was selected for the UNIST Artificial Intelligence Challenge Program (AICP), receiving a grant of 4 million KRW for the study. Developed TUI that ControlNet-based generates images with desired poses and compositions by manipulating ball jointed dolls in reality and capturing them using webcam

Suggested Image Generative AI Service Interface Design Improvement Guidelines

Conducted case studies and thematic analyses on posts with more than 10 recommendations in Arcalive AI Art channel, right after emergence of NovelAI Image Generation. Defined Stable DiffusionWeb UI user behavior patterns, provided UX/UI, external service usability evaluations, and suggested guidelines through qualitative analysis. Evaluated UX/UI, extensions, external services, and usability and improved guidelines as follow-up study based on quantitative analysis. This follow-up study was presented as a first author paper at the IASDR 2023 Congress in Milan, Italy.

Business Prediction and Advertising Decision-Making

As the leader of Team The Quick Brown Fox Jumps Over The Lazy Dog in the 2nd KAIST-POSTECH-UNIST Data Science Competition, achieved 2nd place out of 43 teams, winning a gold prize. Based on real data from Hana Bank, predicted the probability of being self-employed using a CNN-based ensemble model and maximized expected revenue using simulated annealing and bounce rate. In the finals, our presentation proposed advancements in data, model, and service aspects. My interview as a representative of the winners was published in a school article.

Developed Creativity Support Tool for Line Drawing in Manga Production

Implemented pipeline that turns analog sketches or rough digital sketches into images with pencil texture line art and shading using state-of-the-art (SOTA) models.

Print.ly: Applied Style-Transfer in LG Household & Health Care’s Mini Tattoo Printer

Industry-academia Joint Research with LG Household & Health Care. Bulit pipeline that generates artist style-transferred image with YOLACT segmentation. Set up conda virtual environments, attempted built Docker container and image, and wrote README. Composed and delivered manual for operating the entire pipeline.

bbb Korea: Multilingual Voice Data Text Conversion App

Industry-academia Joint Research with bbb Korea. Implemented pipeline that converts dialogue voice files in 20 languages into text using AWS S3 and transcribe.

Shale Gas Production Prediction and Well Purchase Decision-Making

As the leader of Team JORDY in the 1st UNIST-POSTECH-KAIST Data Science Competition, achieved 4th place out of 38 teams, winning a silver prize. Utilizing real data provided by the Korea National Oil Corporation (KNOC), predicted production volumes using PyCaret and decision-making to maximize expected well purchase profits through binary integer programming. In the final stage, our presentation suggested adding the Reynolds number as a shale gas-related feature to structured data and introduced environmental scores in the context of ESG management.

Solutions for Industrial Accident Prevention

As the leader of Team 721, participated in the 2021 Campus Patent Universiade, advancing to the first round of presentations. Our selected topic was in the Patent Commercialization Sector: A7: Method and System for Motion Based on Interactive Service, proposed by the Electronics and Telecommunications Research Institute (ETRI). Proposed a solution utilizing CCTV and patrol robots to automatically build a training dataset for a model recognizing hazardous behaviors and accidents, aiming to prevent industrial accidents and respond effectively when occur. During this process, received advice from rofessors, companies, institutions, factory officials, and patent attorneys. Reviewed industrial accident reports and analyzed the market impact of the Serious Accident Punishment Act and government policies. Created a technology tree and OS matrix, and designed a patent circumvention strategy based on registered patent information. Using IPC patent classification codes in WIPSON, conducted patent searches, and provided ETRI with a detailed analysis using methodologies such as personas, PESTEL, 3C/FAW, SWOT, 4P, and the Business Model Canvas. Additionally, proposed a marketing strategy based on technology transfer and provided information on expected total sales profits and societal contributions based on expected growth rates, as well as potential areas for business expansion.

Developed Cat Arthritis Predictive Model Based on Behavioral Data

Authored a research proposal that was selected for the UNIST Artificial Intelligence Graduate School (AIGS) Creative Independent Project Research, receiving a grant of 5 million KRW for the study. Proposed the research topic and designed a controlled experiment, creating a collar with a 9-axis IMU sensor, MetaMotionS+, to measure cat movements. Obtained time-series data on the location, acceleration, and angular velocity of cat behaviors, and conducted preprocessing to contribute to the development of the model.

Refrigerator for User Healthcare

As the leader of Team 77, participated in the 2020 Campus Patent Universiade, securing 2nd place in the Patent Commercialization sector A4: Refrigerator for User Healthcare, a topic proposed by Samsung Electronics. Our team awarded the Samsung Electronics Vice Chairman Prize for Excellence. Proposed a smart refrigerator concept that facilitated interactive communication between trainers and users via a screen, and provided a platform for users to compete in health scores. In this process, conducted market research in the healthcare industry and non-face-to-face medical platform sectors. Developed an OS matrix and prior art research document for patent circumvention. Using IPC patent classification codes in KIPRIS, searched patents. Employed different methods such as personas, PEST, 3C, SWOT, STP, 4P, and the Business Model Canvas to provide Samsung Electronics with a detailed analysis. Our award was announced in a school article.