Pony｜LoRA｜明日方舟 黍｜Arknights Shu

V1 uses full TAG annotation.

手蠻容易崩的，不知道是因為我太菜還是 LoRA 的問題（我是菜雞，不太會處理手的問題）

基本人物（base character）：

shu \(arknights\), multicolored hair, hair intakes, hair between eyes, dragon horns, pointy ears, hair stick, beads, jewelry, earrings, tassel, tassel earrings, very long hair,

衣著（outfit）：

strapless, shirt, strapless shirt, white jacket, open jacket, ???? shoulders, off shoulder, white pants, ????????????, long sleeves, dragon tail,

黍的大蒜 玉琮（Shu's weapon / Jade cong tube）：

sword, holding weapon, holding sword

第一次嘗試訓練 LoRA。

模型使用 417 張圖片訓練，其中 40 張來自 MMD 影片截圖，其餘的來自 Danbooru，每張圖片訓練 100 次 。
模型格式為 bf16，訓練底模使用 WAI-ANI-????-PONYXL v11 waiANINSFWPONYXL_v11.safetensors，您可以嘗試使用其他的模型，但我不確定結果如何。

推薦參數：CFG scale 0.5~0.7。（0.7 以上輸出品質明顯下降）

目前模型對頭髮有 overfitting 的感覺，導致頭髮有機率會在背景亂飛，並且背面和武器的訓練資料太小。預計12月底考完期末會重新製作一次資料集並發布更新的 LoRA。

This is my first attempt at training a LoRA.

The model was trained on 417 pictures, 40 are screenshots from MMD videos, and the rest are from Danbooru. Each pictures is trained 100 times.

The model is in bf16 format, and wass trained on WAI-ANI-????-PONYXL v11 (waiANINSFWPONYXL_v11.safetensors) base model. You may try using it on other model, but I am unsure of the results.

Recommended parameters：CFG scale 0.5~0.7。（Image quality is bad when >0.7）

Currently, the model seems to be overfitting with hairs, causing it to appear as if it is floating against the background. Additionally, the dataset lacks image of Shu's back and Shu's weapon.

I plan to create a new dataset and release an updated LoRA model here once I finish this semester at the end of December.

大蒜比青蔥還長，也比較粗，明顯比較像黍拿的東西吧。（一般都是用鱗莖，蒜苗比較少用。）