Another Dagger Ear this time NOT an artists OC I like and is clearly from a more mainstream video game and my first "1boy" Character LoRA!
This is Axel from Disgaea 2 who was once a famous demon action hero and rock star known by many as the Dark Hero at one time and as we see in Disgaea 2 he is trying to get his fame back.
Anyways this model because Axel does not really have ANY proper character sheets (at least ones I could find scouring the Internet and fan Wika sites for a few days) was largely trained on for at least a good 75%-90% of the guidance data was trained on fan art of him and with Nano Banana Pro doing a majority of the heavy lifting (understatement of the year) for the actual training I think it came out not too bad IMO overall for what I generated from the best Epoc for a character that as far as I know NEVER had a model made of him before for Stable Diffusion since there is really so little of him to work with overall.
The final product LoRA MAY (keyword) be a little scuffed but with the moment I tested the Epoc in my local environment with the Checkpoint (Nova Anime v11) I used because once again not a subscriber here but overall I think it came out decently well all things considered.
To Summon Axel use this:
1boy, solo, Akutare, male_focus, purple_eyebrows, blonde_hair, short_hair, purple_eyes, pointy_ears,
His default outfit is as follows (after a bit of help from ChatGPT 5.3 for Danbooru style prompts):
open_clothes, open_jacket, jacket, white_jacket, long_sleeves, bare_chest, midriff, navel, pants, purple_pants, belt, white_belt, shoes, white_footwear, necklace, pendant, earrings, stud_earrings, buttons,
I'm going to be playing with him more later on today/tonight since I got work in the next little bit and I still have a few more characters I got to test before I shove off for the next few long hours.
Enjoy and have fun!
Description
Comments (3)
"Dis-gay 2"? Lol.. I had a feeling I recognized the art style. I never personally got into the games, but there was another similar game called Arknights, which I also didn't get into, cuz I'm not really into that type of game(Mostly just know about it cuz of a character named Aak; cute, cool design). However, I have been playing Arknights: Endfield*, and that one's pretty decent so far.
If you need more material to train on, use the best images that you generate from his lora. Just be sure to keep those images consistent, so the lora trains better. It can confuse the lora during training, if he has different body types (slender, vs. big buff/manly), especially if not tagged consistently to reflect those traits.
Endfield; It's gacha, so that part sucks, and leans heavy into female characters/representation. So, yea.. Could be better with more balanced representation overall. That said, I've already got all the male characters, so at least it's not too hard to get those. And I haven't even spent much on it(just the usual bit for the monthly pass(optional), and any cheap one-time packs(the one that's like, a $1.50 and gives you an endgame/BiS weapon for the protagonist character (iirc..))).
And the crafting system is pretty unique/interesting.
As well there's a custom zip-line system in the game, that makes area traversal pretty fun.
Thanks for telling me I can use the actual generations Stable comes up with if they are of decent quality for training models farther. Never thought you could do that or you CAN but you run into getting into a whole ChatGPT Yellow Image problem since it trained itself on it's own stuff when it scoured the internet over the years so it's not really recommended overall.
I am already starting to think later this year of revisiting and updating some of my models already down the road that I know they can be better.
@LaharlTheCreator "Yellow Image problem"
Haven't heard of that one specifically, but I've heard the general concept (models effectively canibalizing themselves, by being retrained from the same data). However, this applies far less when talking about unique character loras and the like.
I think to some extent, it will start to happen with nude women, cuz those have been WAY overtrained into the base models, and people keep insisting on adding more in the loras, many of which, then get merged into base models and re-released. But, you'll see evidence of that, long before this boy starts lookin messed up from over training.
Just, like I said, make sure of two things, quality, and consistency. And the quality and composition are something that you should be especially picky about. Especially with visual errors that are consistent across a dataset, cuz that's when you start getting reproduction of those errors in generations. Anatomical errors, is a big one, cuz they can often slip by unnoticed, cuz they look alright at first glance, until you look closer and realize that 'Hey.. The human body ain't supposed to do that..'.
Things like, stretched torsos, stretched limbs, extra eyelashes, uneven eyes, off-center necks, belly buttons 2 inches away from nipples, tails coming out from halfway up the back (on anthro characters), etc.. many thing that are even often missed by professional artists. This is also a part of why AI messes up a lot of humanoid features, cuz it was trained on the art of many artists that didn't pay attention to fundamental anatomical details.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.