Pure Innocent girl 純真女孩

Important reminder: It is recommended that all users who previously used this LoRA download the new V2.5 version. The new version has been adjusted by LoRA Block Weight, which not only preserves facial effects but also reduces contamination of clothing, background, and style.

重要提醒：建議之前所有使用這個LoRA地用戶下載新地V2.5版本，新地版本經過LoRA Block Weight調整，既保留臉部效果又减小了對服飾、背景、風格地侵染。

The name "Pure Innocent Girl" is given because it is not a specific character's face, but rather a blend of styles that can produce a young and beautiful innocent girl's face, and a more natural and realistic (less plastic surgery sensation) oval face; It is necessary to enable Face restoration, with a weight between 0.25 and 0.85. The higher the weight, the more likely it is to produce childish features that look too childish. It is recommended to use it alone with a value of 0.5, and the maximum value should not exceed 0.9. If abnormalities occur when using it with other models, Lora, or certain samplers, it is recommended to randomly adjust the weight.Otherwise, there will be deviations that affect the style. Suitable for semi realistic and realistic effects, but not tested in anime styles.

If you like this LoRA, please leave a comment or image!

取名叫“純真女孩”，是因為它不是特定人物的臉，而是更像一種風格混合，能產生年輕美麗的清純少女臉，並且是更自然真實（整形手術感覺更少）的鵝蛋臉；一定要開啟臉部修復，權重在0.25～0.85之間，越高越容易產生看起來太孩子氣的幼稚特徵，單獨使用建議0.5，最高不要超過0.9，否則會產生偏差影響風格。如果搭配其它模型、Lora或者某些採樣器使用發生了异常，建議隨機調整權重。適合半寫實和逼真效果，但並沒有在動漫風格中測試。

如果你喜歡這個LoRA，請留下評論或圖片！

On May 23rd,

Xindi V2.5 versionKnown problem: "Long Range Full Body" will collapse on the lower face.

5月23日，

新地V2.5版本已知問題：“遠景全身”下臉部會崩。

On April 22nd,

The V2 version is the final version of this LoRA, with the following adjustments made:

The model was trained with "stable_diffusion 15 pruned" and "vae ft mse-840000" to improve the universality of LoRA;

Trim tags, remove some tags and add new tags such as clothing and posture, while simplifying the abbreviation keyword to "PI_girl". Of course, as long as LoRA is loaded or the entry is written<lora: Pure_ Innocent_ Girl2:0.4>, can be used without triggering words;

Updated the dataset, removed some potentially harmful images, added some high-quality high-definition new images, trained with about 300 images, and also improved the training resolution;

Increase the dim value to 64, increase the number of training steps, and adjust the learning rate.

Known issue: Using excessive weights (exceeding 0.7) can still affect the character's clothing and background style.

Example of dataset labels:

1girl，best quality，masterpiece，realistic，solo，looking at viewer，face，eyes，nose，mouth，lips，smile，makeup，

Pose: suggestive_ spread_ legs_ pose，lying，on back，Standing，sitting，kneeling，Squatting，

Clothing: Ice princess dress,

4月22日，

V2版本是這個LoRA的最終版本，做了以下調整：

模型改用“stable_diffusion_1_5-pruned”與“vae-ft-mse-840000”訓練，以提高LoRA的泛用性；

修剪標籤，删除了一些標籤和新增了一些服飾、姿勢等新標籤，同時簡化縮寫關鍵字為“PI_girl”，當然，只要加載了LoRA或者詞條寫上<lora:Pure_ Innocent_ girl2:0.4>，無需觸發詞也可使用；

更新數据集，删除了一些可能導致不良的影像，添加了一些高品質的高清新影像，用了約300張圖片訓練，同時也提高了訓練分辯率；

增大dim值調整為64，新增訓練步數，以及調整了學習率。

已知問題：使用過高權重（超過0.7）依然會影響人物服裝和背景風格。

數据集標籤示例：

1girl，best quality，masterpiece，realistic，solo，looking at viewer，face，eyes，nose，mouth，lips，smile，makeup，

姿勢：suggestive_ spread_ legs_ pose，lying，on back，Standing，sitting，kneeling，Squatting，

服飾：Ice princess dress，

On April 10th,

The V1 version is an experimental LORA model that used approximately 200 images to train in "ChiloutMix" and generate "vae-ft-mse-840000" VAE. This is a very featureless LoRA, which did not meet my original expectations. It still looks like a mixed race with many Korean faces.

4月10日，

V1版本是一個實驗性的LORA模型，用了約200張圖片在“ChilloutMix”中訓練和“vae-ft-mse-840000”VAE生成。 這是一個非常沒有特點的LoRA，並不符合我原本的預期，看起來還是混合了很多韓國面孔的混血。