100danganpack
Important The lack of prompt information here makes this release a bit premature. But I figured since I was going to be testing this for a while I might as well put it up.
What is this in theory? An attempt to train a lora on every ultimate (and most significant non or post graduation ultimates) from the Danganronpa franchise with some students from The Hundred Line: Last Defense Academy as well as Shinigami and Pucci from Rain Code added in as well. Basically a big pile of Kodaka and Komatsuzaki's combined works. I also have a few characters from the Danganronpa Another fan games.
I also wanted to try boosting how well single-shot gens of of multiple characters would do. So I specifically trained on some images of two or three characters with thematic or canon links. Haven't had a chance to test how well that worked yet. I'd assume that the more represented in the larger danbooru data the characters are together the better this would work out. So Hajime and Chiaki are probably going to go well. Tsumugi and Gonta not as much.
What is this in practice? No idea. I really just decided to try it out because I was curious what would happen and I had 'just' enough spare buzz to do it. So far testing has it working better than I'd feared but not as well as I'd hoped. But given the scope it's way too early to say. If I try again I think just doubling the dataset size would be enough to make this viable. Which I'll probably try once I've had time to collect and caption all the 100 line character sprites from the game.
Issues found so far.
There should be a couple images of Darumui from 100 line in the main gallery to highlight the problem. It 'mostly' gets her in the hemoanima form. But it still misses on some details that differentiate her from any other anime girl with gradient hair wearing a dark uniform. And even more so with the standard look. I'm assuming that's going to be how it'll go for a lot or all of the 100 line characters. The Danganronpa cast 'usually' had a strong presence in the Illustrious training data anyway. But here things pretty much begin and end with this lora training on a very small amount of imaes per character. It's the difference between just strengthening what's already there and adding something new. Basically the 100 line characters I trained on are in there and they 'kinda' work but you're better off just looking for a dedicated lora for them.
Usage tips The ideal would be to mix it with a specific character lora. But if you're going kind of scattershot with prompts, an alternative would be to go for loras that 'also' have a wide scope of danganronpa-related training. My Tsumugi beta loras have a fair amount of non-tsumugi stuff in there too. And I tried to get a very wide range of characters in my danganronpa anime lora. So right now I've been playing around with a mix of this lora in upper mid range strength, the danganronpa anime lora at full strength, and the Tsumugi experimental lora (v2) at very low strength. This is a bit of a mess to be sure, but here's my current base prompt with all the loras:
lora:Danganronpa_Locations:0.1, lora:Tsumugi_Shirogane_Danganronpa_Illustrious_V2:0.2, lora:Danganronpa_Anime_Style_Mix_V2_Illustrious:1, lora:100danganpack-illustrious:0.6, 100danganpack, danganronpa_3_(anime), danganronpa_3_(anime)(style), danganronpa(series),anime_screenshot , anime screencap, anime coloring, official_art, komatsuzaki_rui_(style) , YOUR_ACTUAL_PROMPT_HERE
Or to push it even further with a bit more booster from ccdanganart:
lora:Danganronpa_Locations:0.1, lora:Tsumugi_Shirogane_Danganronpa_Illustrious_V2:0.2, lora:ccdanganart:0.1, ccdanganart, lora:Danganronpa_Anime_Style_Mix_V2_Illustrious:1, lora:100danganpack-illustrious:0.6, 100danganpack, danganronpa_3_(anime), danganronpa_3_(anime)(style), danganronpa(series),anime_screenshot , anime screencap, anime coloring, official_art, komatsuzaki_rui_(style) , YOUR_ACTUAL_PROMPT_HERE
The downside of course is that with every extra lora things are all getting further dilluted into each other. Which is a pain if you want a specific art style like the danganronpa anime. Though some with cross-lora training, like using komatsuzaki_rui_(style) , won't suffer from that nearly as much. However, even then there's different examples of komatsuzaki_rui_(style) in each lora, going with his CG over sprite work, etc.
What was the dataset curation methodology? I started out with a base of official artwork from danganronpa and then added in all the higher quality pieces from danbooru that had a crossover between 100 line and danganronpa. Then I scripted out a process to check how many images a specific character had and to then fill in the gaps with quality sorted, tag blacklisted, items from danbooru. For characters who still had gaps I filled them with the best synthetic data that I could. I don't like using synthetic data, but after the amount of gens I've done Ihave a fair amount I think are viable to beef up characters like Aika. In total that didn't make for a huge amount of images per character given that I was using civit's trainer and its limit of 1,000 images. My hope was that it wouldn't be a huge issue for modern Illustrious models given that they generally have a pretty good grasp of the major danganronpa characters anyway. Having a nudge in the right direction for well represented characters and a blow to the head to drive in the minor characters better would be enough. I'm more curious how Flux will do. Given that Flux has no knowledge of the franchise but better adaptability to training on smaller datasets.
Please just forgive me already! Might as well call me Leon, because when I was putting together the list of major characters I forgot Kanon Nakajima.
Prompting style? Underscored, comma seperated, danbooru style tags. Yes, for the Flux one as well. I was thinking more along the lines of using it for Chroma so didn't want to go through the process of re-captioning everything in a narrative style too.
Irony
So far the strongest recognizable training example for me in the dataset is Tsumugi. The character I already have many, many, many loras for.
What are the tags? The specific tags are going to take a while to get up. I'm in the process of documenting all the danganronpa characters, tags and descriptive, in all the official outfits. But until the just finding an example on danbooru should be the best bet for trying this. Or using any of the images I've posted as a guide.
Anything special about the tags or clothing? Most of the Danganronpa 10th aniversery "prom" outfits were in the training data. Not to a huge extent, they're sparse, but they're in there. Again, specific tag lists to come.
What Danganronpa characters were trained on for this lora? iruma_miu
kamukura_izuru
alter_ego_(danganronpa)
ikusaba_mukuro
naegi_makoto
kirigiri_kyoko
togami_byakuya
fukawa_toko
monokuma
ishimaru_kiyotaka
asahina_aoi
yamada_hifumi
enoshima_junko
fujisaki_chihiro
owada_mondo
celestia_ludenberg
kuwata_leon
maizono_sayaka
ogami_sakura
ikusaba_mukuro
hagakure_yasuhiro
hinata_hajime
komaeda_nagito
nanami_chiaki
koizumi_mahiru
pekoyama_peko
kuzuryu_fuyuhiko
tanaka_gundham
sonia_nevermind
soda_kazuichi
owari_akane
nidai_nekomaru
saionji_hiyoko
tsumiki_mikan
hanamura_teruteru
togami_byakuya_(danganronpa_2)
naegi_komaru
daimon_masaru_(danganronpa)
shingetsu_nagisa
utsugi_kotoko
kemuri_jataro
towa_monaca
mitarai_ryota
munakata_kyousuke
sakakura_juuzou
kizakura_kouichi
great_gozu
yukizome_chisa
kimura_seiko
ando_ruruka
izayoi_sounosuke
gekkogahara_miaya
akamatsu_kaede
saihara_shuichi
amami_rantaro
shirogane_tsumugi
yumeno_himiko
harukawa_maki
shinguji_korekiyo
k1-b0
gokuhara_gonta
momota_kaito
oma_kokichi
tojo_kirumi
yonaga_angie
chabashira_tenko
shirogane_tsumugi
amami_rantaro
akamatsu_kaede
umesawa_aiko
matsuda_yasuke
otonashi_ryouko
iranami_satsuki
kurokawa_mikako
What pairings of Danganronpa characters were trained on? shirogane_tsumugi, gokuhara_gonta
momota_kaito, harukawa_maki
naegi_makoto, kirigiri_kyoko
naegi_makoto, maizono_sayaka
maizono_sayaka, shirogane_tsumugi
kuzuryu_fuyuhiko, pekoyama_peko
sonia_nevermind, tanaka_gundham
akamatsu_kaede, saihara_shuichi
saihara_shuichi, momota_kaito
saihara_shuichi, momota_kaito, harukawa_maki
k1-b0, iruma_miu
chabashira_tenko, yumeno_himiko
amami_rantaro, akamatsu_kaede
amami_rantaro, shirogane_tsumugi
k1-b0, shirogane_tsumugi
shirogane_tsumugi, enoshima_junko
towa_monaca, enoshima_junko
kamukura_izuru, nanami_chiaki
What 100 Line characters were trained on? sumino_takumi
kashimiya_karua
yakushiji_takemaru
shizuhara_hiruko
amemiya_darumi
aotsuki_eito
kawana_tsubasa
maruko_gaku
tsukumo_ima
tsukumo_kako
ginzaki_shouma
kirifuji_nozomi
magadori_kyoshika
omokage_yugamu
oosuzuki_kurara
mojiro_moko
What Rain Code characters were trained on? shinigami_(rain_code)
pucci_lavmin
General in progress tag list: I'll try to ust gradually get tags for everyone in here.
Darumi Hemoanima
1girl,amemiya_darumi,armband,black_boots,black_bow,black_hair,black_sailor_collar,black_serafuku,black_shirt,black_skirt,blood,blood_from_eyes,blue_hair,blunt_bangs,bodysuit,boots,bow,collarbone,full_body,gradient_hair,hair_bow,highres,holding,holding_knife,holding_weapon,hole_in_chest,hole_on_body,knee_boots,knife,komatsuzaki_rui,last_defense_academy_school_uniform,long_hair,long_sleeves,looking_at_viewer,miniskirt,mole,mole_under_mouth,multicolored_hair,neckerchief,official_art,open_mouth,pleated_skirt,purple_eyes,red_eyes,sailor_collar,school_uniform,score:4,second-party_source,serafuku,shirt,simple_background,skirt,smile,solo,standing,tachi-e,the_hundred_line_-last_defense_academy-,tongue,tongue_out,transparent_background,two_side_up,v-shaped_eyebrows,very_long_hair,weapon,white_armband,white_bodysuit,white_neckerchief
