I work on reconstructing, generating, and understanding the visual world.
I am a Research Assistant at the Computational Photography Lab, National Yang Ming Chiao Tung University (NYCU), supervised by Prof. Yu-Lun Liu.
Recently, I led Splannequin (WACV 2026) and a diagnostic study of world-space human recovery under fast camera motion. I am now exploring visual understanding in vision-language models (VLMs). My goal is to build systems that perceive and reconstruct the visual world as effectively as humans do.
Previously, at UCLA, I worked on privacy-preserving AI and IoT security (MobiCom, IEEE IoT-J).
Camera motion turns into human drift when camera and human estimates are combined at mismatched scales. We introduce the K-Pop stress test to expose it, and show that one closed-form scale removes the drift with no retraining.
InfiniDream turns a single text prompt into a photorealistic, open-ended 3D world: a VLM plans the scene and precisely edits sophisticated procedural code, so every asset stays high-quality and a new seed grows a new world.
GaMO reformulates sparse-view 3D reconstruction as multi-view outpainting, expanding the field of view with geometry-aware diffusion to achieve consistent, high-quality reconstructions efficiently from very few input views.
Voxify3D is a differentiable two-stage method that converts 3D meshes into stylized voxel art with discrete palette control. It preserves semantic structure using multi-view pixel-art supervision and CLIP-guided optimization.
Splannequin freezes dynamic Gaussian splats into crisp 3D scenes from monocular videos by anchoring artifacts to more reliable temporal states.
