I updated this model becuase I found out there's some plently bad images that I use in dataset. Also, I added newer, high quality, and high definition images in my dataset. Image quantity is kinda samey (only increased in single digit images), but has much higher quality ones. Here's some patch notes and updates in this model:
Increased number of images in dataset with much more higher quality and varied images
Increased LoRA size
Increased flexibility in terms of clothings and model compatibility
Increased image resolution trained on from 512 to 640
All images trained in this model already upsized by 4x
Notes:
This model still has issue in racial bias a little bit. Crasher Wake supposed to be a Japanese middle-aged man, rather than Japanese-Cuacasian in his 40s. Idk why is this, probably the checkpoint I used is bit in bias or I didn't add "japanese, middle-aged, 40 years old" like that. Btw, you can try that.
I updated this model becuase I found out there's some plently bad images that I use in dataset. Also, I added newer, high quality, and high definition images in my dataset. Image quantity is kinda samey (only increased in single digit images), but has much higher quality ones. Here's some patch notes and updates in this model:
Increased number of images in dataset with much more higher quality and varied images
Increased LoRA size
Increased flexibility in terms of clothings and model compatibility
Increased image resolution trained on from 512 to 640
All images trained in this model already upsized by 4x
Notes:
This model still has issue in racial bias a little bit. Crasher Wake supposed to be a Japanese middle-aged man, rather than Japanese-Cuacasian in his 40s. Idk why is this, probably the checkpoint I used is bit in bias or I didn't add "japanese, middle-aged, 40 years old" like that. Btw, you can try that.