(Updated Nov 11 2025)

• I just realized that I may have set the UNet learning rate too high (0.0005) in version v1

• I updated the model (v1b) by keeping the same dataset and parameters as v1, and simply reduced the UNet learning rate to 0.0002

• Compared to v1, it seems to produce good results even without using ControlNet (I haven't tested it yet)

• Feel free to use it if you need

• For me, one of the problems using Stable Diffusion is that when I want to generate a character wearing tokusatsu or sentai bodysuit without helmet, the bodysuit often doesn't fully cover the neck skin even "turtleneck" is added in the prompt

• Personally, I CAN'T ACCEPT THAT

• This Lora tries to fix the problem but it's not perfect

• Trigger word: covered neck

• Also add "(color) neck", "(adjective) neck" if necessary

• Weight: 0.8-1

• Recommend to use with ControlNet Canny. For details, see HERE

• You can use it as usual, but it may change the output of the image completely (this is the biggest issue)

• It was not tested on male characters. It may work fine or not