Changho Choi

Computer Science undergraduate at Korea University specializing in computer vision and generative models, with 6+ years of applied machine learning experience in vision and audio. Actively seeking PhD opportunities in computer vision and general machine learning.

From Mar 2025 to Dec 2025, I was an Undergraduate Researcher at the Vision & AI Lab under Prof. Jinkyu Kim, conducting research on causal sequential visual encoders. Prior to that, from Oct 2024 to Dec 2024, I worked as an AI Researcher at OptimizerAI where I contributed to research and development for the Text-to-SFX Model v2. I also worked as an AI Engineer at Pion Corporation developing deep learning models for optimal product images. I was an AI Scientist at Maum.AI from Mar 2019 to May 2023, where I experienced various computer vision tasks through extensive literature review, utilization of open-source resources, and paper implementation.

I am an Undergraduate at Korea University majoring in Computer Science and Engineering, with graduation expected in August 2026. I previously graduated from Gyeonggi Science High School in Feb 2017.

Email  /  CV  /  GitHub  /  Google Scholar  /  LinkedIn  /  X

Publication

I'm interested in computer vision, machine learning, optimization, graphics and robotics.

MambaEye: A Size-Agnostic Visual Encoder with Causal Sequential Processing

A causal sequential visual encoder that achieves input-size agnostic charateristic like human vision by using a pure Mamba2 backbone. We designed a strictly unidirectional approach and relative move embeddings, MambaEye ensures translation invariance and adaptability to arbitrary resolutions.

SAOInstruct: Free-form Audio Editing using Natural Language Instructions

A model based on Stable Audio Open for flexible, free-form natural language audio editing. Trained on a novel dataset of audio editing triplets, it generalizes to real-world audio and unseen instructions.

LatentSwap: An Efficient Latent Code Mapping Framework for Face Swapping

A lightweight face swapping framework that generates latent codes for pre-trained generators without external datasets. It features a fast training process with a simple three-term loss, producing high-resolution results comparable to state-of-the-art models.

Work & Research Experience

Besides my work on the research and publications above, a sampling of my past works

Vision & AI Lab, Korea University

Conducted research on causal sequential visual encoders under Prof. Jinkyu Kim. Proposed MambaEye, a size-agnostic visual encoder that treats image recognition as a causal, sequential process, effectively mimicking human saccadic vision.

OptimizerAI

Contributed to research and development for the Text-to-SFX Model v2, achieving state-of-the-art human preference win rates. Update with further details if desired.

Pion Corporation (vcat.ai)

Developed deep learning models for optimal product image recommendation and online image clustering using CLIP and Triton Inference Server.

Maum.AI

Led the computer vision research team and worked on various CV tasks (Super Resolution, Face Swapping) through open-source utilization and paper implementation.

Other Projects

These include coursework, side projects and unpublished research work.

HifiFace PyTorch Implementation

Developed an unofficial PyTorch implementation of HifiFace, the state-of-the-art face swapping model at the time (Stars: 300+).

FaceShifter Pytorch Implementation

Implemented FaceShifter, one of the first GAN-based face swapping models, from scratch in PyTorch (Stars: 600+).

Design and source code from Jon Barron's website