3 version:v1 v2 v3 lora for sd 1.5NoobAI-XL-v1: sdxl lora for <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294?modelVersionId=1046043">this</a>All versions were trained on approximately 350 illustrations and fanart of MapleStory. additional tags: small body, big head to achive the atonomy/style body of maplestory character

Maplestory style

version 3: BIG CHANGES, this is not Lora but Locon (idk what is the true power of locon but i got the better result now so i guess it helped a lot) or <a target="_blank" rel="ugc" href="https://github.com/KohakuBlueleaf/LyCORIS">LyCORIS</a> (they must be a big fan of lycoris anime lol) trained with dim 32 alpha 16 and 768 RESOLUTION of course with the help of adafactor again and I also reduced the dataset by removing unnecessary pictures that are either of low quality or don't have a big head and small body , althought i still trained it based on maplestory dataset but the style is like an upgrade from the original style and color (, but this model is quite overfit so you have to use hires fix), what i mean is that this locon now have more vibrant colors, more sharper, eyes quality is much more better now , because the dataset is now more focus on the maplestory illustration so now it can draw closer to maplestory character style than the version 2 and 1 maybe some people will like or dislike it, but they can still use versions 1 and 2, which have more original colors I figured out that the problems of version 1 and 2 are the shorter the prompt the more less accuracy lora can draw the style that's why if you put only the chibi 1girl in the prompt then it gives you the perfect style but if you copy the long ass prompt then the style will be immediately distorted but version 3 is better now example : with 8k, 4k, ,1girl, chibi,solo,city,full body, you get thisbut if too specific about the character you may have this (long big body now) Negative prompt and (highres:1.1), best quality, (masterpiece:1.3) (my old prompt) have a significant impact on the style. I don't know which is the best for you all, but I am currently using 4k, 8k, and easynegative embedding, which give me the best results Different models give different resultshands still fucked up that's it i'm very sorry about four fingers if you want more "big head, small body" just do the trick (chibi:1.x) this maybe the last locon i made unless the LoHa from LyCORIS has anything even more significant impact on the quality of the styleThe only reason I made this lora/locon was just to draw my character, but it seems like I failed. Perhaps other people can make better use of this lora/locon than I can version 2 I trained new lora with the same old dataset with new learning rate,text lr , with the help of adafactor, the result is surprise, big changes are 1. more colorful 2. can regconize stuff better 3. line is kinda thin now i guess 4.many variants for 1 character (time to gacha) In my opinion, this one will have more quality results. but sometimes version 1 style will be closer to the game i just did a experiment with this (chibi:2) and the result is still good, i feel like it style even more closer to the usual character in illustration in some cases lmao version 1 lora trained on hundred of maplestory illustration,npc artwork,fanart,some other fanart that could give a better nsfw pic in case you need, Recommend options:put chibi in prompt to ensure the style lora good at 1. and 0.7-0.8 when combined with other loras like 2 vtuber pics below i mainly use DPM++ 2M KARRAS to generate pics . Don't put nsfw in prompt if you want sfw because i want it good at nsfw toosometimes it maybe not look like the style much just click the generate button again lol, ok now i will be back to grinding to 275,btw my main is shade, but it's oke if no one knows this dude

Files with SHA256: f4bf047d56cbe9df4b9b8b13cbcdf322626adb477270ca9b015819e9d0f0cfe8