<h3 id="heading-1564">WARNINGS to get the best results:</h3><ul><li>Do not use her last name anywhere in your prompt! Doing so almost guarantees to give her an incorrect nose and/or wrong face shape due to SD1.5 being convinced it already knows what this person is supposed to look like. ;P I had to choose an activation token of just her first name omitting the apostrophe "aulii" in order to dodge SD1.5's bias towards what it inexplicably thinks that she's meant to look like. </li><li>Lowering lora weight to 0.8 or so helps the results a lot (which might be due to overtraining on my part.. I'm still learning) </li><li>LORA is not great at getting her face right from a distance.. this can be addressed by touchup inpainting though. People who actually understand how the adetailer extension works might know other ways to address this in post, but me and that extension simply do not get along heh heh</li></ul>Trained on 223 images at 640x640. For me this was also an experiment to create a character lora with no captions aside from the activation token, and I think it came out well at least insofar as I don't see any problems that captions clearly would have improved and captioning can be a huge headache y'know? ^_^

Files with SHA256: 8494dbbef1a76a925516aeac51156fa376640292d98c79d48ae9e26922b9b454