ICON
ICON stands for Implicit Clothed humans Obtained from Normals. It is a deep learning framework presented at CVPR 2022 for reconstructing high-fidelity 3D clothed human models from single-view images. The method leverages implicit neural representations to infer detailed surface geometry and clothing texture, overcoming limitations in traditional mesh-based approaches. ICON integrates normal estimation to guide the reconstruction process, enabling the generation of realistic 3D avatars with complex attire details. The software is built on PyTorch and PyTorch Lightning, offering modular components for flexibility. It supports various Human Pose and Shape (HPS) estimators including HybrIK, PARE, PIXIE, and BEV to initialize the underlying body model. Key features include a dedicated cloth-refinement module for enhancing garment quality and the ability to deploy via Google Colab or Hugging Face Spaces. The project provides both inference and training code, allowing users to fine-tune models on datasets like THuma