Hyelin Nam

I'm a PhD Student in Computer Science and Engineering at the University of Michigan, where I'm advised by Prof. JJ Park. Before that, I completed my master's at KAIST, where I was advised by Prof. Jong Chul Ye.

Recently, I have worked on video diffusion models, exploiting their controllability to better capture physically plausible dynamics in generated videos. Moving forward, I am interested in extending these models toward world modeling and broader science-driven applications, including robotics.

Email  /  CV  /  LinkedIn  /  Scholar  /  Github   

Research

A camera-controlled video retaking method that improves geometry guidance with Sierpinski triangle pattern cues and reference-video conditioning.

A cascaded framework that bridges Text-to-Motion and video diffusion models for coherent, camera-aware human video generation.

An inference-time method that boosts VideoLLMs via parallel frame streams for richer temporal reasoning.

A text-to-3D method using a video generation model to jointly generate diverse camera poses and realistic 3DGS for unbounded scenes.

A zero-shot inference-time steering method that enhances geometric alignment in 3D/4D scene generation by integrating scene reconstruction using pose-free geometric reward functions.

Prompt optimization driven by an optical flow discriminator to enhance temporal consistency and natural motion dynamics in video diffusion models.

A simple fix to CFG that enables lower guidance scales, improves sample quality and invertibility.

Ensure structural correspondence by leveraging diffusion features during the score distillation process.