ICON

Why we included this project

ICON turns a single RGB photo into a full 3D mesh of a clothed person, combining predicted normal maps with the SMPL body model so it holds onto clothing detail and body shape even in poses that break simpler image-to-mesh methods. The repo ships more than the core reconstruction: it can output segmented humans, body and garment normal maps, and a rotating video of the result, and trained checkpoints plus a Colab notebook let you judge the output without building the whole stack yourself. There are also training and evaluation harnesses on standard datasets if you want to retrain on your own data. The catch is licensing: this is a non-commercial research license, so teams building avatars, metaverse assets, or animation inputs should treat it as an academic reference rather than a drop-in production dependency. It comes from a CVPR 2022 paper, and the authors point to ECON as the current successor.