Objective: This is a Lora model designed to improve the FLUX model, specifically aimed at enhancing the generation of genital regions in anime-style imagery.

Dataset: The model was trained using seven hundred carefully selected images, all of high quality and avoiding the use of mosaics or censorship bars.

Annotation Method: The images were annotated using the vila-u automatic tagging tool, with all annotations being long texts to more accurately describe the details of the image features.

Training Details: A customized kohya-ss/sd-scripts training script was used, and the model was fine-tuned for twenty epochs.

Test Results:

• Anime Style: The model remains stable when generating anime styles, with effects that are generally reliable and meet expectations.

• Genital Structure: The genital structures are mostly stable, effectively representing the desired details and precision.

• Nipple Generation: Currently, the model performs poorly in generating nipple details, requiring further improvement and optimization.

Each example includes a workflow that can be directly downloaded and used.

The English translation was provided by ChatGPT.

目的：這是一個針對 FLUX 模型進行改進的 Lora 模型，旨在改善其在生成動漫風格中對陰部的呈現效果。

數據集：模型訓練使用了七百張精選圖像，這些圖像都是高質量且避免馬賽克、海苔條。

標記方式：圖像的標記使用了 vila-u 自動標記工具，標記內容皆為長文本，以更準確地描述圖像中的細節特徵。

訓練細節：使用了客製化的 kohya-ss/sd-scripts 訓練腳本，並進行了二十個 epoch 的微調訓練。

測試結果：

• 動畫風格：模型在生成動漫風格時保持穩定，風格效果相對可靠且符合期望。

• 陰部結構：陰部結構在多數情況下較為穩定，能夠有效表現出所期望的細節和精度。

• 奶頭生成：目前在生成奶頭的細節時表現不佳，仍有提升空間，需進一步優化。

每張範例皆包含workflow可以直接下載使用