Character of manga Gachiakuta called Riyou. This lora was trainend with 4 images and the test was made in anythingv5. Therefore the lora maybe func...
#Riyou | Gachiakuta
9 messages · Page 1 of 1 (latest)
Based
Impressive you trained it with just 4 images
How did you manage that?
Numeratic's method is best trained with 20 images or more
Last time I trained some characters with less than 20 images I got shit results
Yep The better is to train with 20 images but i like test the dataset with few images.
The Lora has a few problems, for example, the pattern background is repetitive, and the clothing and hair are not always exact.
I trained the Lora model using 100 steps and 10 epochs, but I noticed that the best results usually occur around the middle. I believe this is because the model reaches a balance between underfitting and overfitting during that stage. In order to achieve flexibility, I experimented with adjusting the weights in the prompt and incorporated OpenPose. This allowed me to enhance the model's ability to handle various poses and movements.