This was a random experiment to see how loras and lycoris differed in learning concepts. I trained on the first 50 images I found when searching for bubble porn. It works best on women at the beach. The most successful is LoHa at network_dim-16, network_alpha-8, conv_dim-8, conv_alpha-1. The rest are in a zip file, along with the dataset. It needs weight of 1.5-2 for really showing the concept. Generally I don't post loras that are over or undercooked but someone asked about it and its just for fun.
It was trained on photos of course, presumably it will work on photograph model I just dont use them.