Excel character saga from a 1999 - 2000 anime Excel Saga
Description
This is the base version of this Lora
FAQ
Comments (16)
Very great! I love Excel Excel! Hail Ilpallazo!
Thanks a lot, glad you like it :)
Hail Ilpallazo!(I didn't know how to spell it, so thanks [Across] Constituent.)
Definitely and others have also made a Hyatt LoRA, hope that there is a LoRA containing both
I believe I'll make other models to, using only the original materials as always, no gelbooru random scraps
@New_Fossil Look forward to it. :)
@New_Fossil probably for capturing the leotards with shoulder pad design since those do look good
I do warn in advance that I plan to make V2s of this and other Excel loras as available materials are poor quality and scarce at best so loras are not in that quality I would like to make them (Had the same problem with V1 Rockna), so I plan to find the series and make my own timed screencaptures (my personal code in making loras, I use only original source materials)
@New_Fossil in that case with you can publish the auto-capture tool in the future, but at the same time will have to see how pose and outfit classification are handled (assuming background is trivial).
@TomLucidor The thing is that I like to pre-edit materials in photoshop as well before using them, i notices that lora quality is noticeable better when all surrounding and unwished artefacts are removed, because of that auto captures were as of yet not used, but I need to find good quality series version first before I even start
@New_Fossil wait then how do you "naturalize" the image in content of its environment?
@TomLucidor sorry for my lack of in deep details, can you explain what do you mean by naturalizing the image content? Not really sure what you mean.
But to put it simply what i do is I photoshop out all unnecessary details around character I don't want in training like, logos, text, background people or details... I leave only character itself and simple backgrounds as to have training focus on character details and not surroundings.
No more in deep then that, I keep it simple as of yet
@New_Fossil if every training image has a white background, won't that cause the contents to not be able to fit into dynamic backgrounds from prompts or other background-oriented LoRA?
@TomLucidor ah now I follow, well to surprise you with the answer, no. Not that I train many of my loras with white backgrounds, that's mostly in case of manga based loras.
My Tomo lora from Iinazuke Kyoutei is trained with white backgrounds only and I must admit that it never caused any problems so far. As for non manga (meaning anime and art based materials) I leave the backgrounds as they are, only artefacts like objects and people (stuff that can compromise training) is removed and capture edited with simple tools like clone tool of similar
@New_Fossil Indeed it was surprising that it is okay to have partial images to function.
It probably is worth documenting the distinction between poster data vs manga data vs anime data (or illustration vs webcomic vs video), and how to handle each and blending them into a single LoRA of the same topic.
@TomLucidor though never crossed my mind but you're totaly right, it is worth documenting it.
I think I'll put an article here on civit about how to work with manga/ comics in monochrome and facts I gathered while working with them, as there is a catch in it.