Duplicated from fantaxy/FitGen2
Learning Flow Fields in Attention for Controllable Person Image Generation