Kyu Song

Building efficient generative models for image and video

Kyu (Hyoung-Kyu) is an Applied Scientist at Captions, working on Mirage, a diffusion-based video generation model. His research focuses on generative AI and model optimization for LLMs and visual content.

Previously, he worked on Stable Diffusion, VLM optimization, and talking face generation at startups in South Korea. He has presented at CES and GTC, published at CVPR and ECCV, and served as a CVPR 2025 Outstanding Reviewer. In 2022, he led the Brain team at MAUM.AI during its KOSDAQ listing.

Kyu Song

Seeing Voices: Generating A-Roll Video from Audio with Mirage

BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion

LD-Pruner: Efficient Pruning of Latent Diffusion Models using Task-Agnostic Insights

EdgeFusion: On-Device Text-to-Image Generation,

Shortened LLaMA: A Simple Depth Pruning for Large Language Models

A Unified Compression Framework for Efficient Speech-Driven Talking-Face Generation

Talking Face Generation with Multilingual TTS

Deep User Identification Model with Multiple Biometric Data