Aiko Umesawa from the Danganronpa 3 anime. I consider this a 2.x version of my initial Aiko lora for Flux/Illustrious/SD 1.5/etc as I made significant changes to the dataset.With the Z-Image Turbo version you might want to keep the strength lower, around 0.7 rather than 1.0.<h1 id="basic-prompt">Basic prompt</h1>In theory tags, narrative prompts, or a combination should all work. With Chroma I've found that I seem to get the best results with a mix of tags and narrative prompting. Though with either tags or narrative you should try to start the prompt with "umesawa_aiko" and have "Aiko Umesawa" somewhere in the prompt. These are the basic prompts.Aiko in her standard Danganronpa uniform (tags): 1girl, hood_up, black_socks,blue_eyes,blush,bow,brown_hair,brown_shoes,brown_skirt,danganronpa_3_(anime),green_eyes,yellow_jacket,heterochromia,jacket,kneehighs,loafers,medium_hair,long_sleeves,open_clothes,red_bow,pleated_skirt,shirt,shirt_tucked_in,shoes,skirt,socks,white_shirtAiko in her standard Danganronpa uniform (narrative): Aiko Umesawa is wearing her uniform from Danganronpa 3. She's wearing a long-sleeved white button-up shirt with a red ribbon tied in a bow at the collar and a brown pleated skirt and black socks and brown shoes. Aiko Umesawa is wearing a yellow bunny hoodie on top of the Hope's Peak uniform. Aiko has heterochromia and her left eye is green and her right eye is blue. Aiko has light brown hair. Aiko's hood is up and it has yellow bunny ears.<h1 id="prompting-for-her-hood:">Prompting for her hood:</h1>In theory you should be able to prompt for her hood being up or down with the following. Though disparity between the amount of training material with her hood up and down means that the hood down generations have a good chance of messing up and still giving her the rabbit ears. This seems to be more an issue with the Chroma lora compared to the Z-Image lora. In theory at least the state of the ears should change depending on emotion. For example, if your prompt says she looks nervous the ears should droop. But in practice it's somewhat unpredictable so you might want to manually state the orientation of the hood's bunny ears. In the previous example prompts I set her hood to be up.Aiko's hood up: hood_up, Aiko's hood is up and it has yellow bunny ears.Aiko's hood down: hood_down , Aiko's hood is down.<h1 id="styles">Styles</h1>I trained the lora on two specific styles - a Danganronpa 3 anime style and a style that tries to mimic what's found in the Dangaronpa games. The Z-Image lora seems to have picked up on those style options a bit better than the Chroma lora. Personally I think that the anime style works best when compared to either the sprite style or prompting without style options at all and generally tend to go with that option.You can try using these styles with the following prompts:Danganronpa 3 anime style: danganronpa_3_(anime)_(style), This is an anime screencap from Danganronpa 3.Danganronpa sprite style: The image is in the Danganronpa sprite style.<h1 id="her-eyes">Her eyes</h1>Both the Chroma and Z-Image lora seem to do a pretty good job with her heterochromia. Every now and then it might get which eye is which color wrong. But for the most part as long as you're prompting for it the lora should be able to handle it better than my previous Aiko Illustrious/SDXL/Flux/etc loras.<h1 id="design-methodology">Design methodology</h1>The biggest challenge with Aiko is the lack of official material to train on. Officially she basically has that one scene in the Danganronpa 3 anime and a brief cameo in the Danganronpa 2.5 anime special. There's obviously fan art, but a lot of care needs to go into carefully curating it to make sure that fan interpretation doesn't drown out the original, that it works with the intended style options, etc. With 2.0 and my previous Aiko loras the best I could do is carefully curating the dataset while carefully captioning it all by hand since I wanted to keep the original dataset fully human made. With 2.1 I added a VERY small amount of synthetic data generated with the 2.0 lora. Just the absolute best of the best added to the 2.0 training data to make the 2.1 dataset. So far that experiment seems to be a success.

Files with SHA256: 4b6c412ce625d3553f6a5e10da549d34f66495130d0fda1c8e1324978981768f