<p><a target="_blank" rel="ugc" href="https://www.instagram.com/annie.shr">Who is Annie?</a></p><h1 id="heading-1487">Notes</h1><p>This is the first Lycoris Locon I've released. It's different from the standard lora in one important way: the file size is far smaller. The file size of a lora is a function of its <a target="_blank" rel="ugc" href="https://github.com/bmaltais/kohya_ss/wiki/LoRA-training-parameters#network-rank-dimension">"dim" or "network rank"</a>. With a standard lora, I use a dim of 96, resulting in file sizes around 110mb. In SD15 land, that's cool. But looking forward to SDXL, the size skyrockets due to SDXL's larger size. And while you can simply create standard loras using a smaller dim in SDXL to help alleviate this, another approach is to use a more advanced style of ancillary model: the Lycoris Locon.</p><p>With a Locon, you can get the same quality of training with a much smaller dim, and hence a much smaller file size.  This locon uses a dim of 24, and clocks in at around 45mb.</p><h1 id="heading-1488">Training data</h1><p>I have included the full training suite of data used to create this Locon in the attached Files. This includes the training images, the repeats used for them, their WD14 generated captions, and the training configuration used by Kohya in json format.</p><p>You can reference this information if you want to create your own locons.</p>

Files with SHA256: 75a0e441f122e80341f238ee370c730b1efa2c1e0bf078731a95796a6ca9e521