I have trained this model at least ten times. Now it has reached a state of balance. But her hands are still easy become bad, and her swimsuit outfit is not learned well enough.

Trigger Words: misorahibiki

The training of this model and the images it generates are solely for learning purposes.

You could find the example prompts from the images above.

My prompts are basically composed in the order of [character traits] + [style] + [expression] + [clothing] + [camera and action] + [background], and you can delete or modify them as needed.

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

Upscale value recommendation is around 1.2, denoising strength is 0.2

I suggest add "3D" in the negative prompt to get good performance.

20240830

The results of the model's training in the first few attempts were very poor, with continuous occurrences of distorted limbs, hands, and chaotic images.

Through continuous debugging and the method of controlling variables, I have roughly identified the possible reasons:

1. There are several images in the dataset that show hands holding a guitar, and these images are too complex for the model to recognize.

2. There are some single images in the dataset that appear only once and have no association with other images, and these images also have complex hand conditions that the model has not learned well.

3. The number of images prepared for the dataset is still not enough, and the model has not fully learned the concepts in the dataset.

4. The quality of the images in the dataset is not good enough, with too many chaotic pigments.

In the end, I removed all problematic images and added new ones. At the same time, I redrew the old images in high definition. This has achieved the current effect.

Other methods could also solve these problems too, I think: changing to another checkpoint or training with flux.

描述:

训练词语: misorahibiki