Sorry to start with some long-winded talk, but since the “1.0” version will have quite a lot of changes, my friend suggested I upload this version now, so everyone can take what they need.
Previously on Pixiv, some senpai had posted AI-generated images of Kousuke and Midori, and on Tensor, a senpai uploaded a LoRA of Kousuke as well—but the results weren’t very satisfying. After getting familiar with the character LoRA creation process, I decided to bring to life my three favorite characters from Kyushu Sentai Danjija. Based on just a few images from the original work, I initially judged the difficulty as: Kouji > Midori > Kousuke. But things turned out very differently in practice.
Midori only took three training rounds to achieve a stable body type and a redesign of the outfit—like removing text and such. Kouji even managed to reproduce the X-shaped harness on his back. But Kousuke… in the first round, even the face shape and hairstyle couldn’t be decently recreated. I tried many times afterward, but they all just didn’t look right. Even when manually tagging the training set, most of the generated images looked androgynous or outright feminine—eyebrows, eyes, and let’s not even mention the clothes, which came out completely chaotic.
After reexamining the original images, I suddenly realized that Kousuke’s art style varied significantly between them, something I had subconsciously overlooked. With that in mind, this LoRA really ended up as a “doujin”. So in the future, Kousuke’s appearance in the “1.0” version will likely be closer to how he looked in some of the older doujinshi.
As for recreating Kousuke’s outfit—especially the one many consider the sexiest, the “black and red skintight suit”—I really exhausted all efforts and still couldn’t get it right. In order to avoid having the outfit trigger words pollute the clothing tags, I tried using some uncommon terms to shift how the AI learned the outfit, but this completely failed with Kousuke. Only when using “leotard” did the color and style come out somewhat close—but unfortunately, that word almost always leans toward a “female leotard” look, and even then, the red parts are still hard to get. The uploaded version of this LoRA is the one using “leotard” as a trigger. It hasn’t been heavily optimized and still includes the original images, so in many cases the eyes aren’t well rendered, and the eyebrows are inconsistently thick or thin.
Well, I’ll stop rambling now. Below are the trigger words, as well as recommended prompt and negative prompt settings that might help achieve better results.
Trigger words:
fukuokakousuke-leader, kousuke-red hair, kousuke-black eyes,
kousukei_battle_outfit, kousuke-red scarf, kousuke leotard,
kousuke-fingerless gloves, kousuke-red overskirt, kousuke-black pants, kousuke-red tabi,
Recommended prompt weights:
toned, young:0.8, manly:0.1, skinny:0.4
Adjust these values to make Kousuke look more like a boy depending on your checkpoint. In most cases, manly:0.1 doesn’t need to be changed.
Negative prompts:
turtleneck, zipper, zipper pull tab, wristwatch, clothes writing, flame print
These help keep the outfit clean.
arm belt will be generated by default when using the outfit trigger words. If you don’t want it, add it to the negative prompts.
Use short eyebrows, thick eyebrows as negative prompts to help Kousuke get back his original thin eyebrows.
As of the upload date, progress on the three LoRAs is around 5%. On a normal day, I can finish about 3%—but that’s if nothing unexpected happens. If the first version that includes all three characters ends up with too many bugs, then I’ll have to start over again.
See you next time~