<h1 id="aerilune-2.0"><span style="color:rgb(34, 139, 230)">Aerilune 2.0</span></h1><p></p><h2 id="aerilune-v2-what's-improved"><span style="color:rgb(34, 139, 230)">Aerilune v2 — What's Improved?</span></h2><ul><li><p><strong>Improved captions:</strong> Used a better captioning model. Curious how much better? The generated baseline images have compositions almost identical to the original images.</p></li><li><p><strong>Improved baseline quality:</strong> Improved the quality of generations without the <code>aerilune</code> trigger token.</p></li><li><p><strong>Adjusted the dataset ratio:</strong> Changed the ratio of <code>aerilune</code>-tagged images to baseline images (without the <code>aerilune</code> token) from 1:1 to 1:0.25. I felt there was little benefit in having the model learn too much from images without the trigger token, so I reduced their proportion.</p></li><li><p>Highly recommend <a rel="ugc" href="https://civarchive.com/models/2723760/diving-anima">Diving-Anima!</a></p></li></ul><p></p><h2 id="style-lora-inspired-by-the-ai-artist-&quot;ao-nori&quot;."><strong>Style LoRA inspired by the AI artist "</strong><a target="_blank" rel="ugc" href="https://www.pixiv.net/en/users/2724969"><strong>ao-nori</strong></a><strong>".</strong></h2><p>I really like ao-nori's work. With an effort into post-processing each image, and the final result is something that AI can't easily create in a single run.</p><p>Even this LoRA can't fully reproduce the original artwork. It's just my attempt to bring some of the style into a LoRA.</p><p>This style focuses on beautiful girl illustrations with an <strong>airy atmosphere, soft blue colors, soft lighting, and gentle shading</strong>.</p><p>This artwork in particular made me fall in love with the style:<br /><a target="_blank" rel="ugc" href="https://www.pixiv.net/en/artworks/143657333">https://www.pixiv.net/en/artworks/143657333</a></p><p>I didn't use any character names in the dataset, but you can try it with <strong>Shiny Colors</strong> or <strong>Gakumas</strong> characters.</p><h2 id="don't-forget-to-use-trigger-word-:">Don't forget to use trigger word :</h2><pre><code>aerilune</code></pre><h2 id="weight-recommend-:-0.7-1.0">Weight recommend : 0.7-1.0</h2><h2 id="some-helpful-assist-prompts-by-gpt:">Some helpful assist prompts by GPT:</h2><pre><code>delicate colored linework, soft organic contours,
semi-painterly anime rendering, translucent color layering,
soft watercolor and gouache texture,</code></pre><pre><code>airy, airy atmosphere, light and weightless,
soft blue, powder blue, pale cyan, cool pastel tones,
soft lavender, dusty pink, pale peach, ivory, mint,</code></pre><pre><code>high-key natural lighting, luminous soft highlights,
soft diffused light, gentle sunlight, subtle bloom,
cool lavender-cyan shadows, warm peach skin tones,</code></pre><pre><code>pastel color palette, low contrast, lifted blacks,
soft tonal transitions, gentle color grading,
soft atmospheric depth, hazy background,
light-filled atmosphere, translucent atmosphere,</code></pre><pre><code>dreamy, serene, peaceful, gentle, delicate,
fresh, clean, refreshing, tranquil,
romantic, nostalgic, summery, quietly joyful</code></pre><h2 id="lora-combo-by-me:">Lora combo by me:</h2><p>Aerilune + <a target="_blank" rel="ugc" href="https://civarchive.com/models/2857532/color-and-contrast-enhancement-for-anima-vivid-illustration">Vividillu</a> + <a target="_blank" rel="ugc" href="https://civarchive.com/models/2646250/animastrong-contrast-between-light-and-shadow">Kinrolstyle</a></p>

Aerilune Style

<p>Aerilune v2 — What's Improved?</p><ul><li><p><strong>Improved captions:</strong> Used a better captioning model. The generated baseline images have compositions almost identical to the original images.</p></li><li><p><strong>Improved baseline quality:</strong> Improved the quality of generations without the <code>aerilune</code> trigger token.</p></li><li><p><strong>Adjusted the dataset ratio:</strong> Changed the ratio of <code>aerilune</code>-tagged images to baseline images (without the <code>aerilune</code> token) from 1:1 to 1:0.25. I felt there was little benefit in having the model learn too much from images without the trigger token, so I reduced their proportion.</p></li></ul>

Files with SHA256: 4cd0815bf71b7054c6c231ee853d66fa8fb8f5dd1519ec7c52c73137ac4e1d33