Sooran Kim

I am an undergraduate student at Yonsei University, double majoring in Psychology and Computer Science.

My interest in AI, especially in computer vision, began during a cognitive psychology course. I found it fascinating that the way the human visual system perceives objects was similar to how Convolutional Neural Networks (CNNs) process visual information. This realization sparked my curiosity about AI and led me to explore the field of computer vision, eventually deciding to pursue a second major in Computer Science.

Email  /  CV  /  Github

Research Interests

I'm currently exploring various areas in computer vision, with a particular interest in video-related topics.

Education

Mar 2022 - Present

Experience

Undergraduate Intern (Advisor: Seon Joo Kim) | Jan 2025 - present

14th & 15th Member, 16th President | Jul 2024 - Present

YAI, the first and only AI club at Yonsei University, focuses on studying the principles and applications of artificial intelligence. As a member, I actively engage in reading, presenting, and discussing computer vision papers with fellow members, and participate in various club activities, including projects.

Projects

Apr 2025 – May 2025

A core challenge in robotics is the lack of high-quality expert demonstrations, especially for diverse and physically realistic tasks. As part of the 6th YAI Contest (YAICON), our team tackled this problem by reproducing and extending the NIL (No-data Imitation Learning by Leveraging Pre-trained Video Diffusion Models) framework. NIL proposes training robot policies from videos generated by video diffusion models—without any real-world data. Since the official code was not released, we implemented the entire NIL pipeline from scratch and applied it to a humanoid locomotion task using the HumanoidBench simulator.

While NIL originally uses only synthetic videos, we hypothesized that abundant raw human videos might offer more physically plausible guidance. We trained policies using four demonstration types: raw human videos, mesh-rendered videos (via SMPL), MuJoCo-rendered humanoid videos (via SMPL), and diffusion-generated videos. Our experiments showed that although final performance was limited due to insufficient hyperparameter tuning, raw videos may serve as a promising direction for future imitation learning research.

You can check out the project on GitHub. While the model is not yet able to walk properly, the following result shows one of the more promising outputs.

Oct 2024 - Nov 2024

As part of the 5th YAI Contest (YAICON) held from October to December 2024, I collaborated with a team of five to develop "Make Your Own Emoji," a tool for creating personalized video-like emojis from 3-5 photos. My role focused on fine-tuning the DreamBooth model on a GPU server to incorporate user-provided photos, enabling high-quality, customizable outputs. The fine-tuned images were converted into videos using the MOFA model, with features allowing users to define motion trajectories. This project was awarded 2nd place in the YAICON.

You can check out the project on GitHub. Below are qualitative inference results:

Aug 2024

As part of a toy project in the YAI, I worked with a team of four to develop a simple face mosaic tool using Detectron2. Two team members were responsible for downloading datasets and implementing the data loader. My role focused on fine-tuning the Detectron2 framework using a GPU server, particularly training a pre-trained Mask R-CNN model to ensure accurate face detection. Another teammate handled the implementation of the mosaic effect. This project gave us the opportunity to explore practical applications of computer vision techniques in a collaborative setting.

You can check out the project on GitHub. Below is a simple inference result:

Honors and Awards

Fall 2024

Fall 2024, Spring 2023, Fall 2022

Aug 2022

Services

Tutor | Mar 2024 - Jun 2024

Provided 15 hours of personalized tutoring across 8 sessions to a student from Kazakhstan, covering the course 'Understanding Psychology' and Korean conversation.

Mentor & Instructor | Mar 2022 - Dec 2023

Taught mathematics to middle and high school students. Based on my performance, I was contracted as an instructor in 2023, delivering structured lessons and accumulating 206.5 volunteer hours.

Things I Love

I don’t always enjoy working out, but I love how I feel afterward. These days, I mostly go to the gym or do Pilates.

I'm a huge fan of Bong Joon-ho, and Parasite is my all-time favorite movie. I watched it three times in theaters and countless more times on Netflix—I've seen it so many times that I can recite the lines by heart.

Website template from Jon Barron.