I've retrained this TI from scratch after pruning some of the images in the dataset. I got rid of certain pics where her face was not clearly distinguishable, a few others that were too low quality and a couple which contained artifacts that could cause problems during training. In the end, I kept just 18 out of the original 30 images. After getting a number of decent preview pics, I stopped the process at 375 steps and tested the embedding. I think the result is pretty good. For starters, it's not overtrained (the previous version was, making it necessary to reduce its weight to about 0.95 in many cases in order to obtain good results). I also consider the result is more accurate in general, as her facial features are a bit less pronounced and probably look more realistic.