Yuxuan Xue

AvaImg, a multi-stage optimization pipeline, is presented: it enforces body-inside-clothing constraint via signed winding numbers, is made viable by a three-level efficiency cascade (~10x runtime reduced, ~95% storage saved), and recovers fine surface detail using coarse-to-fine displacement optimization.

GeoEdit achieves unprecedented geometric accuracy and identity fidelity in object editing, overcoming the limitations of existing diffusion-based methods.

Jointly modeling 3D geometry and relighting in a diffusion framework unlocks physically plausible single-image relighting that surpasses previous pipeline-based or geometry-agnostic approaches.

Achieve realistic human-scene interaction 50x faster by amortizing geometry-based fitting into a feed-forward transformer that iteratively refines human meshes.