After one week, 18 trainings, and 14,719 buzz spent, I have a Lora of my favorite manga art style (hopefully I picked the right training lol). It doesn't capture it perfectly, and hands are a bit unstable sometimes, but I feel it's as good as I'm gonna get for a simple Civitai-trained Lora with a pretty mediocre dataset. I tried to make it look as good as possible without any special settings or prompts, maybe some settings will make it look better.Trained on 58 stylistically-consistent and hand-tagged images of the main characters from the 90s to the 2000s. Not all images were high quality, but it's hard to find high quality scans of this stuff.<h3 id="generation-advice:">Generation advice:</h3>I put the "m0dernstyl3" tag on images of the high quality cover scans. Adding it to the prompt makes it look slightly more like them.For some reason it likes to add straps or sleeves to Belldandy's dress, try messing with the negative prompt to avoid that.Put "torn clothes, broken collar, broken bracelet" in the negative prompt. One image had Belldandy with broken jewelry and ripped clothes. I probably should've removed it, but it's fine if you put those tags in the negative.Also put "blue eyes" in the negative when generating any character that isn't Belldandy, Urd, or Lind. It should mostly handle eye colors correctly without any tags in the positive prompt.All the characters generate pretty well except Lind, she only had 2 images. She doesn't look terrible, but not great either. Also Megumi is hit or miss.Adding "traditional media, scan" might make it look slightly more authentic.

Files with SHA256: b544c3c0e19b763ab7c8c6e8e1960964a80efc08919a69360a19c055fefc91c1