<p><em>If you would like to support me so I can make more LORAs - </em><a target="_blank" rel="ugc" href="http://boosty.to/mv_qt/donate"><em>boosty.to/mv_qt/donate</em></a><a target="_blank" rel="ugc" href="https://boosty.to/mv_qt/donate"><em> </em></a><em>❤️</em></p><h1 id="tags:-cfl1s50nv">Tags:</h1><h3 id="anime-style:-u9ffdujjw"><strong>Anime Style</strong>:</h3><pre><code>anime coloring, muted color, film grain, </code></pre><p><em>I trained style into these tags, I decided to split off "muted color" as separate tag for better flexibility and "film grain" as potentially unwanted part . Remove or decrease weight* of these tags, depending on how much style flexibility/accuracy you need. You can also try adding them to negative.</em></p><p><em>*You can change tag weight like this -</em> <strong>(anime coloring:0.5)</strong></p><h3 id="appearance:-986th55zg"><strong>Appearance</strong>:</h3><pre><code>NaomiMisoraMV, 1girl, solo, black hair, long hair, grey eyes,</code></pre><h3 id="outfit:-3wq6syopv"><strong>Outfit:</strong></h3><pre><code>black jacket, turtleneck sweater, black gloves, shoulder bag, jeans, </code></pre>

Naomi Misora - Death Note

<p>I decided to try to make LORA which can both generate style accurate images and be somewhat style flexible when needed. To achieve this I tagged style with appropriate tags (If style isn't tagged it usually becomes part of character tag). I haven't done much testing yet, but when tagged in such way LORA seems to be less prone to overfitting, even though dataset wasn't not very good, it didn't seem to have affected results.</p>

Files with SHA256: 15601609f3724c1e2187e4dbb323d788ec99fb93cb82384bb3879842944f25cd