V2 Changes: because I found in trying the 1.5 lora I made that every image generated, almost always reverted to the image imperfections in the dataset, so:
1: Try to cut out and remove, or fill in and cover, all of the white edges, black edges, grey edges, and signs of breakage in the dataset that have appeared in the corners and borders of the images due to rough scanning work.
2: Remove dirt and suspected handwriting from the image.
3: Removed easy-to-handle fonts from the image, and things that I thought would taint the background, well, I'm a little regretful that I didn't try to keep Tsukasa Jun's signature as much as possible.
4: Found some new character images, cut the images to separate the characters from the images and put them in the dataset.
5: Added tag: 1990s (style) to the txt of all the old style images, so typing 1990s (style) into Prompt will help generate the pre-artist style, hopefully
Software Translations:
V2 Changes: because I found in trying the 1.5 lora I made that every image generated, almost always reverted to the image imperfections in the dataset, so:
1: Try to cut out and remove, or fill in and cover, all of the white edges, black edges, grey edges, and signs of breakage in the dataset that have appeared in the corners and borders of the images due to rough scanning work.
2: Remove dirt and suspected handwriting from the image.
3: Removed easy-to-handle fonts from the image, and things that I thought would taint the background, well, I'm a little regretful that I didn't try to keep Tsukasa Jun's signature as much as possible.
4: Found some new character images, cut the images to separate the characters from the images and put them in the dataset.
5: Added tag: 1990s (style) to the txt of all the old style images, so typing 1990s (style) into Prompt will help generate the pre-artist style, hopefully