Trained on Anima BaseV1.
Lora trained on the style of Mr. Fuga's art.
And as always let me know if you like the Lora or not. For any questions or requests, dm me.
If you gen something really nice, please share it to further showcase the Lora ✌️
Buzz tips are also appreciated!
Description
FAQ
Comments (6)
Another fuga? Cool. though I can see you didn't bother to fix the watermarks even on 70 images, AND YOU DIDN'T EVEN ADD A TRIGGER - for which I specifically criticize you and advise you to retrain the dataset. But I'll run some tests using my own style, which used about 307 images for training. I have some theories
Triggers are unnecessary and bothersome (mostly only useful for concepts, or character loras), we're past the SD 1.5 days where you have to hold the AI's hand to get a result you want, if you want to train your loras with them go ahead. But I personally have absolutely 0 need for them.
As for watermarks yeah I didn't remove them, I don't bother with something you can remove later on from the final image (Photoshop, lama cleaner, or hell even inpaint it out). As for images, anything after 90 is diminishing results with the settings I run, and do believe me I tested that thoroughly.
Not quite sure why you felt the need to leave such a comment but you do you.
@ToasterKing The developers’ page itself clearly states that the model processes styles most effectively in the ‘@xyz.’ format. All other loras that do not follow this principle suffer from appalling stability and reduced performance, and also blend less effectively with other loras. I don’t get how anyone could miss this note on the developers’ website, or think that the ‘neural network era lora training bla bla bla’ is over, but I don’t know, mate – it’s not that much effort to write a script that’ll inject a trigger into all the text files in the folder. Anyway, do as you like, but it’s not really that difficult.
@portableshot The developers can say one thing, people will always do a different thing. I always blend several loras together with 0 issues when I do my own things. And not once did I have an issue with the final output because I didn't put in a trigger tag for the style lora. If anything think of it this way, majority of your dataset will share one tag anyway, which will in the end act as the trigger when used. Be it 1girl, 1boy, some bodypart, act, etc. Hence I see genuinely no reason to stuff it in even if it is as easy as putting it into the tagger as a mandatory included tag for every image. And yes, it really isn't difficult, takes few seconds tops. I just choose to not do it cause I see no reason to.