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 &lt;lora:kawaii7.1.3_obj20:0.75:0.25&gt; to call the LoRA with the optimal strengths for the UNet and Text Encoder though you can probably get away with simply &lt;lora:kawaii7.1.3_obj20:0.75&gt; depending on your model and other LoRAs and adjust the strength accordingly.I recommend using <a target="_blank" rel="ugc" href="https://huggingface.co/ostris/objective-reality/tree/main">Objective Reality</a> 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.)

Files with SHA256: e4d29826e875c0addbf48954ab0783573af343182fdb13b5389b50901774ce7e