Trained on about 2400 images of kawaii-style lingerie advertisements from Japan, Korea, China, Taiwan, etc. as well as related media like lingerie/bikini gravure idol photos and magazine covers. The women in the photos are usually either Asian or European/Caucasian though I don't see why it won't work for other races.

Use <lora:kawaii7.1.3_obj20:0.75:0.25> to call the LoRA with the optimal strengths for the UNet and Text Encoder though you can probably get away with simply <lora:kawaii7.1.3_obj20:0.75> depending on your model and other LoRAs and adjust the strength accordingly.

I recommend using Objective Reality as the base model though it works surprisingly well with the original SD1.5 as well as most reasonably good photorealistic models.

You'll also want text-related things in your negative prompt. Unfortunately, most of the training images exhibiting the "kawaii" style have a lot of text on them due to being advertisements or magazine scans, getting unwanted text is a side effect that I've not really found a way around without negative prompting.

Use the following negative prompt as a start: text, watermark, signature, logo, Japanese characters, Chinese characters.

If unwanted text is a severe problem, try: (text: 1.8), (watermark: 1.8), (signature: 1.8), (logo: 1.8), (Japanese characters: 1.8), (Chinese characters: 1.8)

If you get a lot of shadows and dark faces, put "poor contrast" in the prompt (or "good contrast" in the negative prompt.)