End-to-end pipeline: single RGB photo → animatable 3D avatar deployable in AR (Adobe Aero). PIFuHD for mesh prediction, IPNet for parametric fitting, AIST++ for motion, Substance Painter for texturing. Published at VRST '21.
An end-to-end pipeline that converts a single RGB photo into a realistic, animatable 3D avatar — usable directly inside VR/AR tools like Adobe Aero. The pipeline stitches together state-of-the-art neural networks, each solving one sub-problem: image → mesh, mesh → rig, rig → pose. Published at VRST '21.
techniques
PIFuHD — multi-level pixel-aligned implicit function for high-resolution 3D human digitization (CVPR 2020)
IPNet — combining implicit function learning with parametric body models for clean 3D reconstruction (ECCV 2020)
AIST++ — music-conditioned 3D dance generation, used here as a re-pose source (CVPR 2021)
Substance Painter for texturing; Adobe Aero for AR deployment
gallery
Pipeline stages — A) input photo · B) PIFuHD mesh · C) IPNet parametric fit · D) AIST++ re-pose · E) AR + texture
