My first LoRA. And I just made it to use it myself. So no negativity, please
Description
FAQ
Comments (7)
I like helping new creators improve, so I hope you don't see this as negative.
The one thing I can absolutely recommend right away for any future LoRAs is choosing a different trigger word. "Bleach" is highly likely to be misunderstood by the generator for its innate understanding of the word and apply that on top of your word without the user intending it to. Things like everyone looking like the main characters (since the IP is big enough to be recognized), bottles of bleach, bleached hair etc. . Plus it may trigger Illustrious's base understanding of the style, which is imperfect and contaminated by such a mix of data. The trigger used doesn't (and generally shouldn't) actually have to be anything the base model actually understands. Some people use 1337 speak triggers (e.g., "6l3ch"), I use letters separated by underscore than make a relevant initialization (e.g., "T_K_B_S" for Tite Kubo Bleach Style), but anything is better than a common word used straight.
Can you answer the following questions so I can be more specific? They're addressing
1: What kind of images did you use? I notice there's nothing on "monochrome" which I'd expect from a manga style, so I'd guess the data is (mainly?) from covers and/or a colorized re-release (it's fine if it is, just helps to know helps when giving feedback). Some types need a few special tags or treatments (e.g., blank out speech bubbles and use "blank speech bubble" as a tag)
2: How many images did you use and did they have "real" (non-flat color/gradient) backgrounds? Styles require a lot of images (100+ for a "general purpose" style), and those that don't have backgrounds in the training data tend to struggle with them.
3: How were your training images tagged (e.g., just auto-label only, auto-label with some basic correction, hand tagged)? Styles are more forgiving to automation,
4: How big a variety of subjects did you use? It's important to not "over represent" characters in a style dataset, and you actually can (and should) include images without people in them (e.g., scenery pictures)
Thanks for the good comment! I didn't know about trigger words, and I'll change it now.
1. I found a person on the website who used the style of the original manga (colored) for art, and I really liked it, so I used his work for training. So, there were no monochrome images.
2. There were approximately 260 images. About half of them had backgrounds.
3. I was a bit lazy, so I used automation.
4. Apart from the images with the characters on blank background and whith painted background, nothing else was used.
Unfortunately, I haven't found a way to change the trigger words for LoRA. As I understand it, they can only be changed before publication. I noticed that i can change the trigger words in LoRA description, but it doesn't seem to be the right.
@bagivan19 For changing triggerword you're supposed to retrain the model with new triggerword as the firs tag with keep tag setting set to 1.
Not sure if it affects style too much but: setting keep tag to 2 and setting all character tags as the second tag should make the model learn the characters even better, but not sure if it affects the ability to change characters to something else.
something like this: bleachstyle, kuchiki rukia
The style trigger should just be anything that the model doesn't know of, literally anything, but keeping it simple is better
You could change the triggerwords on the site like this: edit version details right under the download and thumbs up button. The place where all the details are written about the model. Doing that only changes what's shown there but the model didn't learn it. For the model to learn it you should retrain it like I mentioned above.
@bagivan19 By the answer to 1, do you mean you used generated data? I would not recommend for a style. It means the LoRA learns AI artifacts are supposed to be there, which certainly explains a lot of what I see in the example images.
What Lopodelas said above is correct. With an IP this well known tagging characters by their booru tags can help it avoid learning "male characters look exactly like Ichigo", though I do think it requires enough manual tagging it can be exhausting at the scale of a style.
Thanks for the advice, I'll keep it in mind in the future!
Love this one. First lora? Bravo! It's much better for me than others with Bleach style. Thank you.