Another reason I saw was that models were trained on 512x512 "portrait" images including very few hands. Added to the inherent complexity of hands, this throw off their generation.