<h1 id="heading-11"><span style="color:rgb(250, 82, 82)">This Lora learned the makeup and appearance of some Korean beauties. The purpose of this lora is not to accurately reproduce a certain beautiful girl.</span></h1><h1 id="heading-29"><span style="color:rgb(250, 82, 82)">I'm not very satisfied with the current two versions, because I need the pictures produced by Lora to have higher quality. The third version hopes to be the final version, which will be released after 1-3 days of testing. At that time, I will also publish a comparison chart to help understand the difference between version 3.0 and version 2.0.</span></h1><h1 id="heading-65"><strong><span style="color:rgb(250, 82, 82)">You should use version 2.0 instead of 1.0</span></strong></h1><h1 id="heading-66">I fixed some overfitting problems in Lora and greatly optimized the details that Lora can output. Lora's tendency to draw dirty skin is also gone.</h1><h1 id="heading-67">Now, with the right keywords, it is possible to paint very detailed skin without using Hires Fix. Moreover, when the frame occupied by the face is too small, this Lora can still try its best to draw the face well. Compared with the 1.0 version or most of the models, the 2.0 can well guarantee the facial details when the face only occupies a small frame.</h1><h1 id="heading-68">I also added more example images, which are basically using Chilloutmix and majicmix.</h1><h1 id="heading-12">If you want this Lora to work correctly, please make sure to use any checkpoints that have studied Asian beauties, and avoid using models without the concept of Asian beauties such as realistic vision.</h1><h1 id="heading-13">The generalization of the model is high, and it can accurately perform most of the text input under the weight of 1.</h1><h1 id="heading-28">If you encounter any issues while using it, please don't hesitate to point it out in the comments.</h1>

Files with SHA256: a0259ea74b9f9588e6b35ae2b798b4b96e040c4704899388d96377d939410299