<h2 id="description">Description</h2>LoRA for generating images in a stylized aesthetic inspired by the works of Dizzyspells.Original author: <a target="_blank" rel="ugc" href="https://x.com/Dizzyspellss">https://x.com/Dizzyspellss</a>The model focuses on reproducing the following visual traits:<ul><li>characteristic jaggy lines — sharp, broken, and uneven outlines;</li><li>recognizable color palette;</li><li>stylized body and facial proportions.</li></ul>The main focus of the model is feminine and androgynous characters. It works best for generating femboy characters, as well as characters with exaggerated or pronounced body shapes, especially hips and breasts.Due to the nature of the style, male characters may often appear more feminine or androgynous. For more masculine results, additional prompt clarification or a lower LoRA weight may be required.<h2 id="changelog">Changelog</h2><h3 id="v2.0b">v2.0b</h3><ul><li>Reduced the total number of training steps by approximately half.</li><li>This version is focused on further reducing training time and cost.</li><li>Training time: 1h 16m.</li><li>Training cost: 350.</li></ul><h3 id="v2.0a">v2.0a</h3><ul><li>Reduced the number of epochs.</li><li>Increased the number of repeats.</li><li>Reduced training time and cost without significantly reducing the total number of steps.</li><li>Training time: 1h 19m.</li><li>Training cost: 400.</li></ul><h3 id="v2.0">v2.0</h3><ul><li>Increased the dataset.</li><li>Split the main style categories:<ul><li>default;</li><li>jaggy lines;</li><li>pixel art;</li><li>oekaki.</li></ul></li><li>Balanced the dataset by categories and aspect ratios.</li><li>Added atypical works by the same artist to the dataset to expand the style boundaries and reduce attachment to a single dominant pattern.</li><li>Training time: 2h 5m.</li><li>Training cost: 700.</li></ul><h3 id="v1.0">v1.0</h3><ul><li>Initial release.</li><li>This version is primarily oriented toward characters and recurring patterns of the artist rather than universal style transfer.</li><li>Due to mixed image types in the dataset, the linework is unstable and may shift between regular linework, jaggy lines, and pixel-style rendering.</li></ul>

Style (dizzyspellss)

dataset parameters:<pre><code>
| | brush | oekaki | jaggy | pixel | total |
| -------- | ----- | ------ | ----- | ----- | ----- |
| 1:1(sqr) | 4 | 4 | 4 | 4 | 16 |
| 2:3(vrt) | 4 | 4 | 4 | 4 | 16 |
| 3:2(hrz) | 4 | 4 | 4 | 4 | 16 |
| total | 12 | 12 | 12 | 12 | 48 |</code></pre>training parameters:<pre><code>
| Version | Parameter | Value |
| ------- | ---------- | ----- |
| 3.0 | Steps | 1 008 |
| | Epochs | 12 |
| | Repeats | 7 |
| | Images | 48 |
| | Batch Size | 4 |
| | Clip Scip | 2 |
| | Dim | 64 |
| | Alpha | 32 |
| | Ai Toolkit | true |</code></pre>

Files with SHA256: eff8b6b5ffbce82af129f5b4da39368efda8c40b8e5215c7ee1b1398d249f2ee