This lora should be the biggest challenge since I started playing SD. There are only two pictures, and they are still pictures generated by AI. Then I divided each picture into 3-4, and then mirrored it, and finally got a better picture, 14 in total. Zhang, the most important thing is that the model only has one posture, sitting with legs spread, which is particularly easy to fit. It can be fitted after running for 100 steps. It completely disrupted the way I trained the model before, and the learning rate also dropped again and again. Before and after, we have trained a total of more than 50 versions, from number 0266 to 0271. After repeated testing, or the final version, it is temporarily considered the optimal version, 2.5D, two-dimensional, realistic, all have been tested repeatedly, and sometimes it may be necessary When you draw cards repeatedly to obtain them, one key point is that you cannot violate gravity. When you stand up, the branches in front of your vagina will disappear, haha. The branch on the chest is still healthy.