Trained on a handpicked collection of AI generated images (Midjourney, DALL-E) that screams Fantasy and beauty. Originally, I was planning to make something like <a target="_blank" rel="ugc" href="https://civarchive.com/models/5414?modelVersionId=6297">Pastel-Mix,</a> but I couldn't get good data that doesn't break anatomy like the character is a piece of floating cloth. :pensivePatting: Note: some images use Controlnet in samples<h1 id="-pgzr84xgq"></h1><h2 id="how-to-proompt-pfhj6lzqp">How to Proompt</h2><ul><li>This is <a target="_blank" rel="ugc" href="https://civarchive.com/models/257749?modelVersionId=290640">Pony XL</a> Model. (Important)</li><li>XL models don't require any "Omega LUL Schizo 100KM long random string that even I won't use for a password" Negative Prompt. You don't need any negative prompt at all unless you want more colorful gens for which you usually add <code>greyscale, monochrome </code>to neg. And <code>3d</code> just in case you don't want 3Dish look. So to surmise, you only need <code>3d, monochrome, greyscale, </code>That's it.</li><li>Since this Model is pony diffusion, It follows the same proompting for positive <code>score_9, score_8_up, score_7_up, source_anime,</code> Just use this to start and add your proompt.</li><li>No, you don't need quality embedding use them only if you know what you're doing not because your favorite Youthuber or some obscure rentry which the writer boasted on being a proompting God told you to.</li><li>Make use <a target="_blank" rel="ugc" href="https://github.com/Bing-su/adetailer">adetailer</a> for face with @Anzhc <a target="_blank" rel="ugc" href="https://civarchive.com/models/293448/anzhcs-face-segmentation-prototype-or-yolov8-or-adetailer-model">face segmentation model</a> (it's better than default).</li></ul><h2 id="recommended-config-uapihjsd9">Recommended config</h2>Sampler: Euler aSampling Steps: 20-30CFG Scale: 6-7Upscaler: 4xNomos8kDAT, R-ESRGAN 4x+ Anime6B, 4x-UltraSharpDenoising strength: 0.2-0.7VAE: Baked in, but you can use any sdxl vaeEnable Quantization in K samplersCons:<ul><li>The Model sucks at skin colors.</li></ul>

Files with SHA256: a442744686600051f4e1b9662035b3e3475f7c50ac538a46316a1922e81512e8