少量数据集训练的一次尝试。原本只有一张我用Flux生成的图片，经简单重绘改为两张，多次裁剪（对称变换）凑出11张图片进行训练。尝试了用正则化去掉白背景带来的影响（不清楚是否成功）。最终效果还算满意，比我想象的好很多！只是部分动作表情难出图，可以提升guidance.
模型预览图都是由文生图一次性直出，没有经过后续优化，可能有细节上的问题，请忽略。
触发词是“girl”，如果prompt只有1girl，效果可能不好，建议初始词使用“anime girl”。
This project represents an attempt at training a LoRa model using a minimal dataset, built upon the Flux AI art generation model. Initially, the dataset consisted of only one image created by me using Flux, which was then slightly re-drawn to produce two distinct images. Through multiple cropping processes (employing symmetrical transformations), the dataset was expanded to 11 images for training purposes. An effort was made to mitigate the effects of white backgrounds on the model by applying regularization techniques, though it's uncertain how successful this approach was. The final outcome is quite satisfactory and exceeded my expectations to a great extent! However, some poses and expressions proved challenging to generate, indicating that guidance scale could be improved for better results.
The preview images of the model are all generated in one go from text-to-image without any subsequent optimization, which might result in some detail issues; please overlook these.
Additionally, the trigger word for the model is 'girl'. If the prompt only contains '1girl', the results may not be optimal. It is suggested to use 'anime girl' as the initial keyword for better outcomes.
Description
少量数据集训练的一次尝试。原本只有一张我用Flux生成的图片，经简单重绘改为两张，多次裁剪（对称变换）凑出11张图片进行训练。尝试了用正则化去掉白背景带来的影响（不清楚是否成功）。最终效果还算满意，比我想象的好很多！只是部分动作表情难出图，可以提升guidance.
模型预览图都是由文生图一次性直出，没有经过后续优化，可能有细节上的问题，请忽略。
触发词是“girl”，如果prompt只有1girl，效果可能不好，建议初始词使用“anime girl”。
This project represents an attempt at training a LoRa model using a minimal dataset, built upon the Flux AI art generation model. Initially, the dataset consisted of only one image created by me using Flux, which was then slightly re-drawn to produce two distinct images. Through multiple cropping processes (employing symmetrical transformations), the dataset was expanded to 11 images for training purposes. An effort was made to mitigate the effects of white backgrounds on the model by applying regularization techniques, though it's uncertain how successful this approach was. The final outcome is quite satisfactory and exceeded my expectations to a great extent! However, some poses and expressions proved challenging to generate, indicating that guidance scale could be improved for better results.
The preview images of the model are all generated in one go from text-to-image without any subsequent optimization, which might result in some detail issues; please overlook these.
Additionally, the trigger word for the model is 'girl'. If the prompt only contains '1girl', the results may not be optimal. It is suggested to use 'anime girl' as the initial keyword for better outcomes.