Because the training dataset is very small, the generated images may easily lean toward a loli appearance and visible blush. You can add loli and blush to the negative prompt to reduce these effects.

The training dataset "https://www.pixiv.net/artworks/134431368" was trained using only the first 6 images, so the character outlines may tend to appear slightly reddish.
--------------------------------------
因为训练数据很少，生成的图片很容易偏向loli和脸红，可以在负面提示词加入loli, blush以降低效果
训练数据《https://www.pixiv.net/artworks/134431368》只使用了前面6张图片进行训练，人物轮廓线会偏红