Miiyamoto hikari from Summer Lesson.

Trigger Words: miiyamotohikari

You could find the example prompts from the images above.

Here are two files: one with 130 epochs named as miiyamotoHikari, and another with 80 epochs, which was compressed as zip file with smaller size. You only need to download one based on your preference.

The 80-epoch version is more flexible, but the details are not as refined as the 130-epoch version and easy broken. The 130-epoch version leans more towards a game modeling style, with more stable character appearances, but the images tend to look more like 3D models, resulting in somewhat stiff character dynamics. You could refer to the post I shared below for more details.

Recommended weight: 1~0.6, adjust as needed until the character's appearance meets your requirements.

Upscale value recommendation is around 1.5, denoising strength is 0.2

It is recommended to add '3D' to the negative prompt.

Looking forward to your comments and pictures.

Reference prompt:

School uniform:

miiyamotohikari
1girl,solo,black hair,ponytail,bangs,pink scrunchie,brown eyes,
school uniform,white shirt,blue plaid skirt,blue plaid necktie,red wristband,black kneehighs,bandaid on knee,leather shoes,

White camisole:

miiyamotohikari,
1girl,solo,black hair,bangs,ponytail,blue scrunchie,brown eyes,
white camisole,jeans,tight clothing,choker,jewelry,

Sleeveless white dress:

miiyamotohikari,
1girl,solo,black hair,ponytail,bangs,pink scrunchie,brown eyes,
sleeveless white dress,white collar,white wristwatch,white flat shoes,

Black camisole:

miiyamotohikari,
1girl, solo, black hair, bangs, ponytail, black scrunchie, brown eyes, 
camisole, black camisole, jeans, tight clothing, belt, choker, jewelry, black wristband,

Sleeveless blue plaid dress:

miiyamotohikari,
1girl,solo,black hair,ponytail,bangs,blue scrunchie,brown eyes,
sleeveless blue plaid dress,white collar,white wristwatch,

Cyan tank top:

miiyamotohikari,
1girl,solo,black hair,bangs,ponytail,pink scrunchie,black wristband,black shorts,brown eyes,cyan tank top,

20240624

This version has better generalization performance. I have tried many training methods and finally had to admit that adding some images of different styles to the dataset improves the model's learning effect.

Initially, I only used in-game screenshots, and the results were indeed very similar to the game. However, being too similar to the in-game 3D models made the generated images very stiff. If you use '3D' in the negative prompt, in most cases, it will lead to the loss of the character's features, resulting in a completely flat appearance.

My goal is to train a model that can generate images retaining the original style and character features but not entirely 3D, with strong generalization ability.

In the end, I added some images other than game screenshots to the dataset, such as illustrations and 2D-style images. However, these images were carefully selected to ensure that the character features and style were close to the game style or came from images generated by previous versions in other styles.

Currently, if more high-quality images of other styles are added to the dataset, the training performance of this version may improve further.

---

20240519

The style of this version is more consistent and closer to the game.

It performs better in prompts like 'full body'. However, the hands still tend to turn out poorly. This time, all images not sourced from the game were excluded from the training dataset, in an attempt to make the generated images closer to the original game style. As a result, the initial training outcomes were severely overfitted. After several attempts, the overfitting issue was somewhat alleviated by adjusting the Scale weight norms to 1 and the Network dropout to 0.5.

Currently, this LoRA model still has some overfitting problems, which can be mitigated by adjusting the LoRA weights within the range of 0.6 to 0.8. or add a negative prompt "3d". If you want good performance in prompt like "full body", "full shot", you might set the upscale value to 1.6 or above.

To further improve this model, I believe adding more images of this character from various angles and ensuring more precise and accurate annotations would likely yield better results.

If you use the examples above, please remember to change the tag lora:miiyamotoHikari1512 as lora:miiyamotoHikari

---

20240401

You could get her example tag from the picture above.

First attempt to make a lora.

Collecting more than 80 images from internet and original game as datasets.

Setting its weight value as 0.8~0.9 would get a good performance. There are several problems in this model. If u want a good performance in tag like "full body", u could set upscale as 1.4 or above, otherwise the character's face would be really bad. Her hands is easy to become bad now.

The model have trained 130 epoch now, I guess training more than this may get a better result.