<h2 id="compound-style-multi-artist-style">Compound Style - Multi Artist Style 🧪</h2>--[<a target="_blank" rel="ugc" href="https://civarchive.com/articles/20103">Pinboard Updated🥝</a>]--<hr /><h2 id="compound-d">Compound D</h2>(Neocoill pre-2022🎨 + Bokuman pre-2022🎨 + Kionant 🎨)with this compound LoRA instead of having it finetune an existing to be up to date I wanted to revert the selected style to a previous state in their evolution.I also tried to stack 3 styles now<h3 id="usage-tips-(normal)-:">Usage Tips (normal) :</h3><ul><li>set the LoRA weight at 0.8</li><li>add the name of artist A <code>neocoill, </code>or artist B <code>bokuman, </code>or artist C <code>kionant,</code><ul><li>merging order counts</li></ul></li></ul><hr /><h2 id="compound-c">Compound C</h2>(Kakure Eria 2025🎨 + Kelvin Hiu 2025🎨)in this one I'll be training the evolved style of Kakure Eria 2025🎨 along with Kelvin Hiu 2025🎨, i love how both of these artists reshaped their styles this year, for this compound im focusing on exaggerated form and soft lines which both style are known for .<h3 id="usage-tips-(normal)-:">Usage Tips (normal) :</h3><ul><li>set the LoRA weight at 0.8</li><li>add the name of artist A <code>kelvin hiu, </code>or artist B <code>kakure eria, </code>or both<ul><li>merging order counts</li></ul></li></ul><hr />About experiment :This LoRA is part of a experiment like focusing on training efficiency, since there are tone of art styles that Illust XL knows already but have evolved over the natural growth of the artist, it would be less efficient to train each new form of known style again with a high dim, for both the trainer and the user who will be storing these LoRAs later on.so what i did is trained the artist the based model know and need update on together in a same LoRA hence "Compound LoRAs"<ul><li>used the same artist name as the activation</li><li>i used lower count subset (selected by year/new style form)</li><li>i used artist with a chosen matching trait so training them together will further amplify it</li><li>when all the trained names are used together they will merge with almost the same consistently then prompt merges or even stacking individual LoRAs together</li></ul>Training Note: --so far i think 3 styles within a single rank16 LoRA is the way to go about making these as long as the model knows a bit about said styles--<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/073c45f1-4b70-46ea-a6e2-c39b71e467a1/width=525/073c45f1-4b70-46ea-a6e2-c39b71e467a1.jpeg" />

Compound Style - Multi Artist Style

Compound_C inner styles:<ul><li>kelvin hiu 2025</li><li>kakure eria 2025</li></ul>

Files with SHA256: c17663bd2e4ac26e63d9967ecaccf2c7cce0ecf82301a280411f70641f454bae