Dorothy from MÄR

The trigger word is: doroshii

You could find the example prompts from the images above.

Recommended weight is 0.8~1. Also recommend setting the upscale value to 1.4 or above for better quality image output.

Currently, there are some issues with the dataset annotations, which may lead to some inaccuracies in Dorothy's clothing when using the tags I provided. This is not a big problem, and I will gradually correct it over time. The tags I provided should largely reproduce the character's appearance.

I collected about 80 images for the dataset, mainly from anime screenshots, aiming to maintain the consistency of this model's style as much as possible. I manually cropped and adjusted the image quality, and manually annotated them. My training method uses Prodigy's default configuration. I set the scale weight norms to 0.5 and the network dropout to 0.8 to reduce overfitting. The learning rate scheduler is set to cosine_with_restarts with 3 restarts. The character was trained for approximately 945 steps. I suspect that training for more steps might yield better results.

For dataset organization, I annotated each image with an "anim style". This is to allow for adding the anim style to the negative prompt when using this LoRA, reducing the influence of the original dataset on the final output, and increasing the model's generalization ability. Of course, you can also add “anim style” to the normal prompt, which will make the output image style closer to the original anime style.

I'm happy to see everyone to upload the images generated by this model to this post, whether there are issues or not. I am always happy to see someone using this model. The most important thing is that I can deepen my understanding of my training methods through these your feedbacks.