<h1 id="flat-color-style-illustrious">フラット塗り / Flat Color Style — Illustrious</h1>A light flat-color style LoRA: fewer shading steps, cleaner lines, calmer colors. It pulls Illustrious-based models away from their default heavy rendering toward the look of color manga pages and doujin covers — flat, but not posterized.No trigger word needed. Just load the LoRA. Recommended weight: 0.8 alone / 0.4 (max 0.5) when stacked with a character LoRA Trained base: Illustrious XL<h2 id="usage">Usage</h2><pre><code>1girl, (character / hair / outfit / scene),
masterpiece, best quality</code></pre>Negative: <code>grin, smug, tan, dark skin, bad quality, worst quality, worst detail, sketch</code><h3 id="tips-(tested)">Tips (tested)</h3><ul><li>Keep the quality tags short. <code>masterpiece, best quality</code> is enough. Tags that push rendering — <code>intricate details</code>, <code>cinematic lighting</code>, <code>highly detailed</code>, <code>depth of field</code> — cancel most of the effect.</li><li>Do not write a trigger word. It is not needed, and writing one can print stray letters into the image.</li><li>Add <code>grin, smug, tan, dark skin</code> to the negative. Without it the LoRA leans toward a smug grin and slightly tanned skin (a dataset bias).</li><li>With a character LoRA, stay at 0.4–0.5. At 0.6+ some seeds start to pull hair color or add stray effects.</li><li>Plain backgrounds (<code>white background</code>) come out very light grey rather than pure white.</li></ul><h2 id="what-was-verified">What was verified</h2><ul><li>Same seed, with and without the LoRA, 4 SFW scenes over several seeds: hair color tags are respected (black / blonde / silver), no extra props, no color shift at 0.8.</li><li>Complex multi-figure poses: 4 compositions × 3 seeds, alone (0.8) and stacked with a character LoRA (0.4). No melted or fused anatomy in 24 images.</li><li>Stacked with a character LoRA at 0.4: the character's horns, two-tone hair and accessories stay intact.</li></ul><h2 id="training">Training</h2>43 hand-picked Danbooru <code>flat_color</code> illustrations from 37 different artists (max 3 per artist), chosen for white/plain backgrounds, thin lines and one-step or no shading. Only the style block of the U-Net was trained (the B-LoRA style block, <code>up_blocks.0.attentions.1</code>), which keeps the LoRA from changing composition, props or the character.<h2 id="caveats">Caveats</h2><ul><li>Strong composition prompts still win; the effect is a nudge, not a full repaint.</li><li>The style is "flat-ish anime" in general, not any single artist's style.</li></ul>

Files with SHA256: 2a734c9f83dbf43fe37db11236c3ed04b1ade4740e4b27f86518d236f03e83f9