Um... Strictly speaking it's not stable (two-person image).The main training set images come from the anime screenshots of the theatrical version.<s>Maybe there will be a new version in the future?</s><s>I think using the following prompt words as negative may improve the stability of the two-person image.</s><pre><code>(blonde hair:0.5),(yellow eyes:0.5),</code></pre>style<pre><code>anime screenshot</code></pre>

Files with SHA256: 519c0bb6d15e8a8d541c58a7162d06dad650c8630197f606e2a1cd47c2e796b2