This is a LoRA for Goru ゴル(@glty07)'s style, created upon request.
I will remove this if requested by the original artist.
Info:
Trigger word: @g0rug0ru,
Description
Trained on Anima Base v1.0.
FAQ
Comments (4)
How do you train styles with artists who exclusively use so much text and multiple views in their works? I wanted to train this artist's style but figured it would be too much work cleaning the images.
Speaking of Anima, you just need a few images out of several dozen training images where only one character is shown from a single angle, with no text or speech bubbles. For the other images that contain multiple views, speech bubbles, text, or sound effects, you just need to tag them properly. This will allow you to generate simple images like the samples on this page. I think those single-character, text-free images act as regularization images. Of course, if you do not input anything, it will end up generating images with various angles, speech bubbles, sound effects, and so on, but you can easily avoid this by adding those tags to the negative prompt.
@toghashie441 That makes sense, but as a follow up question, do you go through and manually change each images' tags? I'm assuming you use an autotagger, but how would you know if it happened to miss or add any tags that you want/don't want?
@Sfdwackys After auto-tagging, I manually add the necessary tags. I mostly just tolerate and leave any incorrectly added tags alone. However, I manually add tags that should be there if they were missed. In particular, you need to tag things like speech bubble, narration, and sound effects properly; otherwise, the trained LoRA will generate them even without explicit prompts, and you won't be able to prevent it with negative prompts either.