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

I did nothing, I'm just a porter.

You could find the example prompts from the images above.

Recommended weight: 1.0~0.6, adjust as needed until the character's appearance meets your requirements.

Upscale value recommendation is around 1.5~2.0 (especially in the case of full body.), denoising strength is 0.2

Continuous adjustment of the prompt words is needed to achieve a good representation.

If the model tends to exhibit blurriness, that can be alleviated by adjusting the model's weights. Adding 'score_6,score_5,score_4,' to negative prompt, 'highly detailed,highres' to positive prompt should work too.

20240725

The model should perform better now that the dataset has been replaced with clearer images.

There are still some overfiting problems.

20240721

I think better image datasets are needed to improve the model's performance.

Because the quality of the dataset is not good enough, the model tends to exhibit blurriness, which can be alleviated by adjusting the model's weights. Adding 'score_6,score_5,score_4,' to negative prompt, 'highly detailed,highres' to positive prompt should work too.

Tips: I read about this method in an article on CIVITAI (I forgot which one). The original dataset did not include images of full body or back views, to address this issue, I extracted full body and back view images of a similar style and added them to the dataset. Then, during caption, I only tag the characteristics of the images without adding trigger words. After training, I found that the model's learning effect was good, and the characters could be displayed in full body and back view images normally.