Time to share my own private(not commision ones) LoRAs of SD1.5.
Donation link: https://boosty.to/blessedworld/donate
Description
FAQ
Comments (15)
Hello randomizer!
As you can see from my gallery, i'm a big fan of Urusei Yatsura world (Lum, Sakura, Ran, Benten... etc) and usually tip creators to support their LORA... the only pity is there are already dozen of LUM on CivitAI and few secondary characters...
Your is the second LUM came out just today!
(see also : https://civitai.com/models/394485/lum-or-urusei-yatsura-lora-ponyxl )
So, I will surely try your Lum, but I would also hope you will create other characters from this anime... : D
hello i hope you try all 3 versions. And sorry, i don't think i made others on sd1.5, and don't think if i do it on PonyXL(other charas) i share it to community in time it actual for me(models be private/paywalled in time i use it for patreon/pixiv posts myself).
@randomizer89 Well, you can see I tipped 8000 buzz here:
https://civitai.com/images/4343321
and 16000 buzz here:
https://civitai.com/models/352594/lum-urusei-yatsura?modelVersionId=427884
Since, in this last case, the creator was preferring Kofi donations, he send me back the buzz and I donate 16 Euros on his Kofi. No problem.
The 2008 version is from some OAV?
yes, style is much different -> i do it.
I have a problem with the 1981 LORA.
It seems to have the tiger bikini not separated during the train.
Infact I dont see such trigger word.
The effect is the bikini is often mixed if you try to use another cloth.
I don't think i fix this LoRAs, i just do workable ones for PDXL later (~in 7 days i hope, but don't sure).
P.S. you can beat this problem on 80's(by negative of tiger print bikini/skirt), but on 2008 it hard to beat.
P.P.S. if just tagging of it on train not a true solution for this char, need frames with other dress(not a problem, she have it on later eps/in movies) for it works correct.
@randomizer89 if you wanna work on characters, better choose another one from Urusei Yatsura. For a unfortunate coincidence, as told, this one is already well implemented by other user., on both SD1.5 and SDXL Pony. For example, male characters for SD 1.5 are very rare and poor trained and for XL there are any at all. If needed, I can help with the dataset.
@settima_ai i want to do all main charas in one model PDXL(1.5 works bad by design, it can produce stable only simple images). And "help with dataset" is much harder than you think, as i see when fast look at it title have 3-4 different styles inside(i really think it be realised by different group of artists) and only one of it is good looking for me. it need to find best eps first, and later find 30-40 images for each char needed. too Hard work. (yes i don't interesting to train on arts/googlable content).
And sad, movie styles is different too, and it not better. Great movie 2 is not a best styled, i can't use screens from it.
@randomizer89 Yeh, I know. I have the DVD with episodes and I know the serial very well, different styles included.
All I can help is to grab image stills from DVD, choosing best quality one and various poses. If necessary, I can edit to crop/exclude unnecessary details. Recently a creatot told me DSXL can train also on few images because is smarter. This would be precious for rare charactes, seen on serial just on few episodes.
Another surprising fact about the "LUM dataset" shared by NocheKaiser is the images resolution of any size.
I was bellieving it request all same resolution imaged, but may be SD XL is more elastic on this (?). If confirmed this would be another big help: when there is a group we can just crop out the wanted character, instead of photoshop the image...
Soon I will post the links for some dataset I'm preparing, just to give the idea...
@settima_ai
Ataru's Father
Ataru's Mother
Benten
CAO-2
Ginger
Fujinami, Ryuunosuke
Kotatsu-neko
Kurama
lum
Lum's Father
Lum's Mother
Mendou, Ryouko
Mendou, Shuutarou
Misuzu
Miyake, Shinobu
Moroboshi, Ataru
Onsen-Mark
Oyuki
Pepper
Perm
Principal
Ran
Ran's Mother
Rei
Ryuunosuke's Father
Sakura
Sakuranbou
Satoshi
Sugar
Ten
Ten's Mother
List of charas i want to be in model(maybe i lost few mentioned lvl ones, i just look at list from MAL and copypast charas i don't forgot) :) and for main ones need to select good style not random one.
Ofc i can made it in different LoRAs too if you want and if you pay some small money for this spreading(i lost 500 buzz/LoRA for small ones -> 50 additional cents for it).
P.S. if you made screenshots you made more than 50% of work. It be very very helpfull.
P.P.S. poses are not really important for PDXL, but a couple or three fullbody and back frames per 30 screens is desirable. ~
@randomizer89 Would be a dream! At the moment I have completed my KURAMA frame harvest.
here it is:
https://mega.nz/file/O41iQb4R#e3sumGuZyWpGbYnrirOspM467CNpov_x3J3HxYcE4H8
read the instructios in the english TXT file (don't care for the italian version, is just a sort of internal diary)
The principle I followed is to harvest all usable images avaiable, because this charater is rare (just 7 episodes)
So this wants to be a complete image collection to buid any future dataset for SD1.5, Pony or the incoming SD 3.0 -> user must check images careful, deleting want not usefull for his training dataset.
( And, yeah, I took a lot of hours to harvest all 4 episodes... )
Now I'm curious to understand what do you really need to "clean and prepare" images for your dataset.
If you can please share, for example, the dataset you used for this LUM, it would help me a lot to understand...
@settima_ai join to my discord group and write me PM if you can, it much easier to use discord for it, than chat on civitai where i can easilly not read answer. I lost where i write anything about clean and prepare - but anyway i just known what better to crop/clean and what can no touch/just tag in text.
@settima_ai additional: 1 think i try now do LoRA of her, just for test,with drop most frames from ep71(i didn't like quality(it drawned in different style most frames))
@randomizer89 Sorry, no Discord. But, since would be a long and big project, please take note of the mail I use for collaborations:
settima.ai@gmail(dot)com [where (dot) is just a . ]
(Please, contact me, possibly with the LUM dataset, if you agree to show me. Tomorrow I will remove the mail adress from here, for safety reasons)