Please note that this LoRA is still in an experimental stage, and there are still many issues that have not yet been resolved(especially in nude/NSFW mode).仍处在实验阶段，问题很多（尤其是裸体模式）182 images,4 batch size,12 epochs,5 repeat timesThis is an experimental Flux LoRA model focusing on the Asian female physique.Tired of the pervasive trend of overly slender representations of Asian women in AI art, this project aims to explore and cultivate a more voluptuous and full-figured body type.<hr />The current issues are:<ul><li>hand and foot details need more additional images as support, or existing hand and foot images in the dataset need to be removed/cropped to avoid negatively influencing the base model.</li><li>delete <code>fuller form</code> to remove pregnant effect</li><li>wrinkles in the waist area of the data images need to be processed.</li><li>more images with various body orientations and different clothing types need to be added. Otherwise, there's a risk of being influenced by NSFW data, potentially resulting in exposed breasts.</li><li>enable T5XXL training and review captions.</li></ul>这不是一个通用类型的的身材模型，是一种特定身材的 LoRA。对于前者（通用类型身材），使用定点优化的方法将 LoRA 的学习目标限定在特定的身体部位：腰部、胸部、臀部和髋部，可以尽可能地减少 LoRA 中身材特征对其他部位的影响（个人感受就是尽可能减少对手部、脚部以及腿部的干扰，可以让人物的随机姿势/朝向少出现畸变，同时手脚也细节最多，最难绘制），减少伪影现象，避免了 LoRA 之间的互相污染且兼容性强（这是本站某位优秀作者身材作品的思路分享）。对于后者（特定身材）来说，前者的训练倾向性导致缺乏细节和个性感，比如腿型、肩宽、身材的肌肉感、 脚部和手部的细节这些身材特点就没办法重新复现.但特定身材 LoRA 的缺点是需要更多的数据集图片和更精细的分类与标注，因此也就需要更长的训练周期。这是两种训练思路之间的取舍：追求细节，还是追求更稳定的控制。

Files with SHA256: 5789ab990b5bc9fd13db75a59ce0e8e88e765c84297770a735f4d188e1e83101