<h2 id="heading-54">Usage instructions:</h2>Please use DPM++ 2M Karras/DPM++ 2M SDE Karras, not Euler A.Avoid using extremely unusual resolutions; stick to standard ones like 768x1024, 1024x1024, etc. (add +256 if needed).Whether to use pre-made images or add details is entirely up to you (there's not much difference, but you can redraw 2-3 images if necessary).As this is my first hastily crafted SDXL model, it may have some flaws.There is a case of an unattractive woman being accidentally learned by the model, which might result in her face appearing in some generated images.Support for male images might not be optimal, as I didn't train the model on male photos, and it could lead to a slightly effeminate appearance (though it still works reasonably well).If color blocks, distortions, or overexposed colors occur, reduce the weight of your tags. Adjust the CFG value between 3.5 and 6.<h2 id="heading-55">Recommended tags:</h2><ul><li>(8k, RAW photo, best quality, masterpiece:1.2)</li><li>(realistic, photo-realistic:1.4)</li><li>ultra-detailed, extremely detailed CG 8k wallpaper</li><li>(crystalstexture skin:1.2)</li><li>(extremely delicate and beautiful)</li><li>detailed skin</li><li>Please try to provide comprehensive tags related to image quality. Half-body images perform well, but there might be slight difficulties with foot and facial details.</li></ul>————————————————————————————————————————<ol><li>不要使用Euler a，请使用DPM++ 2M Karras/DPM++ 2M SDE Karras</li><li>分辨率不要过于诡异，请按照标准768x1024,1024x1024以此类推（+256）</li><li>要不要图生图补充细节完全看你自己（其实没什么差别，如果需要请重绘2~3）由于这是我匆忙炼制的第一个SDXL模型，所以还有一些缺陷</li></ol><ul><li>（训练集中有一个丑女恰好被它学进去了导致大头照可能会出现那张脸）</li><li>对于男性的支持可能不太好因为我没有训练男性照片（我自己跑着其实也还可以，就是有点娘炮）</li><li>如果出现色块，畸形，过曝的色彩，降低你的tag权重，CFG值调整在3.5~6之间推荐tag：(8k, RAW photo, best quality, masterpiece:1.2),(realistic, photo-realistic:1.4), ultra-detailed, extremely detailed cg 8k wallpaper,(crystalstexture skin:1.2), (extremely delicate and beautiful),detailed skin</li></ul>由于本人tag写得非常之烂所以在测试生图之中借鉴了吐司各位玩家的咒语，如有冒犯，多有得罪，此模型存在一些过拟合问题，由于我没有充分的照片拿来做训练集所以远景的时候人物脸型会崩得很厉害（<s>老生常谈的问题了</s>），请尽量补全关于画面质量的tag，半身照表现良好，不漏脚脸部稍微吃力。此模型记住了一些训练集中的脸部特征（<s>你可能会看到纯熟知性的深田姐姐</s>）

Files with SHA256: 57aea594ce7695b34061c8b733a2ee4fb9810f7bf7c0ed29121ef8105c31bed8