I wouldn't fully recommend actually using this model as it's mainly for testing out some training ideas to see what sticks. There's some great stable Tsumugi loras on this site, including mine. Still, this one should hopefully offer something a little different, for better or worse.Models with multiple versions are instances where I'm really just throwing things at the wall to see what sticks. At the time of writing, that's why there's a bucket of flux versions. After having done more testing the e25 is looking like it's going to be the winner among the Flux versions. It's just 'very' finicky about prompts and the strength level with different models and other loras. Sadly flux crawls on my system so testing these is a slow process since it monopolizes my GPU. So testing takes some time. Right now the main intent for this "branch" is to try out methods of balancing narrative and tag based captioning to hopefully get the best of both worlds. Along with seeing differences with how various models take to it. This is also part of a larger effort to get my stuff ready to train loras for chroma and get some of my illustrious loras working with flux as well. At the moment the Flux loras 'might' work with Chroma. I've had bad results with previous attempts, but as of Chroma v46 the model seems to be picking up the gist of the lora. Messing up a lot of smaller details like the buttons on her jacket,but still better than the last time I tried using one of my Flux loras with Chroma.Chroma aside, if I've uploaded a model here it's 'probably' usable for image generation with Tsumugi Shirogane or Chiaki Nanami from the Danganronpa franchise. Hopefully with both tags and descriptive prompts. How well theory holds out in practice remains to be seen. But in general you should be able to just toss danbooru style tags at it reliably. And narrative descriptions probably won't work as well, but you should hopefully get something. And a combination of the two should be fine as well. Or at least slightly better than you might expect from a base model without this lora. The biggest caveat there is Stable Diffusion 3.5. My results with 3.5 are improving, but it's pretty unpredictable. At least with my workflow.Training data consists of official, higher quality unofficial , and a smattering of synthetic art of Tsumugi Shirogane with all her official outfits (V3, Danganronpa S, 10th anniversary dress). I've balanced it out with a variety of images from the Danganronpa animes and games to push that style a bit while also hopefully allowing for direction of image generation to go with each style. I've also included a fair amount of data for Chiaki Nanami. Not nearly as much, but enough that it should be able lto reliably generate images of her as well even if a base model doesn't have much or anything in the way of danganronpa knowledge. Likewise it should do a good job of having both of them together in the same image. The base tags for Tsumugi are: 1girl, solo, danganronpa_(series),danganronpa_v3:_killing_harmony, shirogane_tsumugi,tsumugi shirogane, long_hair, round_eyewear, sidelocks, very_long_hair, wavy_hair, blue_hair,dark_blue_hair, blue_eyes, breasts,medium_breasts And for her standard Danganronpa V3 uniform: bow, bowtie, buttons, collared_jacket, collared_shirt, corset, black_corset, dress_shirt, glasses, jacket, black_jacket, lapels, layered_sleeves, long_hair, long_sleeves, notched_lapels, open_clothes, open_jacket, orange_bow, orange_bowtie, pleated_skirt, black_skirt, round_eyewear, shirt, sidelocks, skirt, very_long_hair, white_shirt, shoes, brown_footwear, black_socks For changing outfits you should be able to just swap out the uniform tags with something else.To get both Tsumugi and Chiaki with tags, try a prompt like this:(((2girls))), multiple_girls,side-by-side, danganronpa_(series), BREAK, danganronpa_v3:_killing_harmony, shirogane_tsumugi,tsumugi shirogane, long_hair, round_eyewear, sidelocks, very_long_hair, wavy_hair, blue_hair, blue_eyes, medium_breasts,BREAK,danganronpa_2:_goodbye_despair, nanami_chiaki, chiaki nanami, breasts, pink_hair,medium_hair, flipped_hair, pink_eyes, spaceship_hair_ornament, hair_ornament,large_breasts , danganronpa_(series), danganronpa_v3:_killing_harmony, shirogane_tsumugi,black_socks, blazer, blue_corset, blue_hair, blue_jacket, blue_skirt, bow, bowtie, breasts, brown_footwear, buttons, collared_shirt, corset, glasses, jacket, kneehighs, layered_sleeves, long_hair, long_sleeves, medium_breasts, medium_skirt, open_clothes, open_jacket, orange_bow, orange_bowtie, pleated_skirt, pocket, round_eyewear, shirt, shoes, sidelocks, skirt,socks, very_long_hair, wavy_hair, white_shirt, BREAK, danganronpa_2:_goodbye_despair, galaga, nanami_chiaki, animal_bag, black_jacket, bow, brown_hair, brown_skirt, cat_bag, hood, hood_down, hooded_jacket, jacket, long_sleeves, medium_hair, miniskirt, neck_ribbon, pink_bag, pink_bow, pink_eyes, pink_ribbon, pleated_skirt, ribbon, shirt, shoes, skirt, white_shirt, zettai_ryouiki,BREAK, PUT THE REST OF YOUR PROMPT HEREOr in their Danganronpa S bathing suits: And here's the prompt for them both in their bathing suits: 2girls,BREAK,Danganronpa S,bow,breasts,chiaki nanami,nanami_chiaki, pink_eyes,pink_hair,cleavage,collarbone,danganronpa_(series),danganronpa_2:_goodbye_despair,danganronpa_s:_ultimate_summer_camp,flipped_hair,frills,hair_ornament,spaceship_hair_ornament,large_breasts,mole, mole_on_breast, white_bikini,white_bikini_chiaki,white_lace, medium_hair,navel,BREAK, shirogane_tsumugi, tsumugi shirogane, danganronpa_v3:_killing_harmony, blue_eyes,blue_hair, orange_bikini,orange_bikini_tsumugi,light_orange_bikini,long_hair,medium_breasts, round_eyewear,sidelocks,stomach,very_long_hair,wavy_hair My default descriptive prompt is: Tsumugi Shirogane is wearing her uniform from Danganronpa V3. Tsumugi Shirogane has knee-length dark teal-blue hair parted in the center, which is straight and smooth at the base and gradually splits into waves near the tips. Her eyes are large, round, and are a soft teal-green color. Tsumugi wears round, rimless glasses with white temples. Tsumugi's Danganronpa V3 school uniform consists of a white blouse with an orange bow tied at her collar underneath a two-button, pocketed, teal-black school blazer adorned with two silver buttons on each lapel and a single button on each gorge. She wears a knee-length teal-black pleated skirt with two dark teal stripes lining the hem. Her legwear consists of dark navy knee-length socks and teal-gray loafers.However, so far I seem to be getting good results with this more concise version:Tsumugi Shirogane and Chiaki Nanami. Tsumugi Shirogane wears her Danganronpa V3 school uniform - white blouse with orange bow, teal-black blazer with silver buttons, knee-length pleated dark navy blue skirt, navy knee-high socks, and teal-gray loafers. Tsumugi's long knee-length dark teal-blue hair is parted in the center, straight at the base and wavy at the tips. Tsumugi wears round rimless glasses. Chiaki Nanami wears her Danganronpa 2 uniform - white button-up shirt with dark teal-gray placket, light pink ribbon bow, pale beige pleated skirt, dark teal-gray hooded cardigan with gray lines, black thigh-high socks, and light pink Mary-Jane shoes. Chiaki's neck length dusty light pink hair curls outwards with inward-curling face locks, straight bangs split right, and her hair held in front with a white Galaga spaceship hairclip.<h1 id="experimental-stuff">Experimental Stuff</h1>You may or may not get extra impact from using some of these prompts with this lora. But like I said, experimental. The positional and "one person cosplaying but not the other" stuff probably isn't represented well enough in the dataset to really have much impact but I was curious. If nothing else the positional tags do seem to help the models understand that there should be "two" people in the image rather than three or more.komatsuzaki_rui_(style)official_art, komatsuzaki_rui_(style)danganronpa sprite style, danganronpa_sprite_styleThis is an anime screencap from Danganronpa 3. , anime_screencap_danganronpa3, danganronpa_3_(anime), anime_screencapThis anime screencap is from Danganronpa the Animation. , anime_screencap_danganronpa_the_animation, anime_screencapglossy surfaces,high contrast, intricate details,masterful shading,shiny skin,smooth shading, modern anime, anime1cosplayer2cosplayerschiaki_left, tsumugi_righttsumugi_left, chiaki_rightmodern, modern anime, 2023, 2024, 2025

Tsumugi Shirogane [Experimental Branch] ( Danganronpa )

Chroma: This version is going to take me a while to really test out and get used to but I'm liking it so far. Compatibility with some of my Flux style loras seems...okay'ish but I'm going to be doing full retrains on those datasets over time. It should work equally well with both tags and narrative prompts. The biggest mistake I've been making with chroma is not using enough steps. If you're not getting the quality you expect, push up the step count. I'll hopefully have more examples up soon. I'm using ai-toolkit for lora training and apparently it doesn't save the generation metadata.I currently have the nononoway style training, and assuming that goes well the danganronpa anime style is next. But I'm seeing mostly ok results with this combo: hyper flux 8 step lora at a strength of 0.124, unlocked cfg lora at strength of 1, this Tsumugi lora at a strength of 0.8, and the flux danganronpa anime style lora at a strength of 0.5. Finally,, with a minmium of 15 steps at 1024x712, using the euler sampler and beta scheduler. However, that's VERY preliminary. My system is pretty slow so I'm really pushing for as low a spec on this as possible.

Files with SHA256: 74191b24a02cb2bd685e338e8a7265752cbc436d1af9fa5d0362d0575f0e2f02