<h3 id="before-use"># Before Use</h3><p>This LoRA is trained based on Illustrious-XL-v2.0.safetensors. Therefore, it is recommended to use it with models of the same <span style="color:rgb(76, 110, 245)">Illustrious</span> architecture to achieve optimal generalization performance.</p><p></p><h3 id="introduction"># Introduction</h3><p>The main purpose of this LoRA is to address the insufficient details in realistic-style content generated by the original Illustrious model. These details include skin texture, body shape, and more importantly, facial style. Most Illustrious models for realistic styles primarily feature Western facial characteristics, which do not align well with the general aesthetic preferences of Asian users. Additionally, the original model's adjustments to body shape tend to reflect personal preferences, such as hip width proportions.</p><p></p><h3 id="known-issues"># Known Issues</h3><p>1. Since no training was conducted specifically for hands, the generation of <span style="color:rgb(76, 110, 245)">fingers may be suboptimal</span>.</p><p>2. The skin tone of characters in the dataset is relatively dark. This is because normal photos, without post-processing, tend to exhibit a natural darkening effect under indoor lighting. However, this can be appropriately corrected using the prompt <span style="color:rgb(76, 110, 245)">"pale skin"</span>.</p><p></p><p></p><h3 id="about-1118"># About 1118</h3><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/f2f169df-135a-4a06-aa04-1953791e07f7/width=525/f2f169df-135a-4a06-aa04-1953791e07f7.jpeg" />I further optimized the details of LoRa training, and now this LoRa model presents more specific results and completely avoids overfitting. You can see its specific effects on different models. In general, it mainly focuses on advanced training for larger chests, wider hips, and fleshier thighs. In addition, skin texture has also been optimized, making it very milf overall.</p><p></p><h3 id="about-1125"># About 1125</h3><p>Strictly speaking, version 1125 does not belong to this sequence because about 50% of the training dataset was reused. However, the adjustment of training parameters and the addition of other higher-quality datasets have made this LoRa change significantly.</p><p></p>

Files with SHA256: 44b712544482e6f45b4870ecf2cff38b34ec4ea7c537403a6b422e0519f31541