One looks hot-blooded and reckless, but is actually unexpectedly calm and intelligent. One looks like a total baka... and well, he kinda is. One seems cold and emotionless, but is always quietly caring for his juniors, silently vowing in his heart to protect everything about them.
The unreliable adults have finally taken up the task of protecting the peace of Kyushu. The boys no longer need to fight, they can finally live out the vibrant youth that should’ve always belonged to them. And now, with your creativity, you can bring their everyday lives to life.
Here comes the usual nonsense at the beginning, but it’s really necessary to explain some potential misunderstandings in usage, as well as the meaning behind the name “doujin LoRA.”
Given that the resolution of the original character pics is quite poor and AI is inherently random, this LoRA does not aim to perfectly reproduce the original designs. The costumes have been modified: the "black zipper inner lining" was removed, the text on the clothing was deleted, and the weapons were removed. The "red long scarf" and "green long headband" lacked side and back view references, so the AI couldn't learn or generate them well. Same goes for color coordination—the LoRA can’t fully mimic the original color schemes.
Also, a few character traits got unintentionally exaggerated thanks to doujin fanworks—like Kouji’s thick eyebrows or Midori’s little fang—which weren’t that prominent in the OG designs. So this LoRA includes both “regular” and “enhanced” versions of those features, generated randomly.
To keep their body shapes consistent, I decided to base their physiques partly on canon, partly on personal taste. Here's the general idea:
Kousuke and Midori are about the same height, while Kouji is a bit taller.
Kouji is the leanest, lowest body fat, and has those sexy visible veins.
Midori has the broadest frame, similar muscle mass to Kouji, but slightly higher fat percentage.
Kousuke falls in between for bone structure, but has less muscle and more body fat than the other two.
Thanks to that, this LoRA can already generate solid body types just using the word “nude”—no need for complicated prompts. Add extra prompts only if you want to fine-tune.
This is also one of the few LoRAs that doesn’t need a fancy workflow, region control, inpainting, or img2img tricks. Just using prompts can generate 2boys or 3boys images—The catch is: prompt formatting, prompt order, and resolution are critical to getting the right composition. When those align, the output is surprisingly accurate.
That said… to get here, I had to retrain this LoRA multiple times, cherry-pick better results, and train again. So yes, the boys did drift a little from their original designs—guided by AI and a lot of my personal preferences. But, isn’t that exactly what doujin means?
Please don’t be annoyed by my “nonsense,” and don’t be too eager to jump to the trigger words. Since we can’t fight against AI’s randomness, listing the already-tested failures helps reduce wasted attempts during actual use.
Here are the recommended negative prompts (reasons will be explained later): “nose blush,” “angry.” I’ll write up a small reflection on this creation process later. If you’re interested, you can read it to avoid making the same mistakes. Due to the differences in the details of the three boys’ costumes, it’s hard for the AI to learn them to the same degree. Although it’s possible to generate them, the result is quite poor in quality, so I suggest not using them.
Resolution Tips: 2boys images are sensitive to resolution.
A 1.5 aspect ratio between 1024×1536 and 768×1152 is generally reasonable.
A 1.333 aspect ratio such as 1152×1536 and lower is also mostly safe, but might produce extra characters.
When the aspect ratio drops below 1.33, the closer it is to 1, the more likely it is to generate additional characters.
3boys is even more sensitive.
According to test results, when starting with 1152×1152, the aspect ratio should ideally not exceed 1.2.
Here are some settings with fewer errors:
Width × Height: 1152×1152, 1216×1216, 1152×1344, 1360×1280, 1360×1152, 1024×1216, 1216×1024, 1024×1360, 1360×1024, 1536x1360, etc.
Other resolutions aren’t unusable, but require a lot of trial and error to find suitable and properly ordered prompts for angles, actions, etc., and are also heavily dependent on the checkpoint used. It is recommended to use models based on illustrious 2.0 or the latest, popular ones with larger training volumes, or gay-themed optimized models. You can also pair with LoRAs optimized for multi-person generation.
Important note: Due to the structure of my training data, this LoRA is highly sensitive to high CFG values. Going above 5 often causes tons of artifacts. So stick to CFG 3–5 for stable results.
As for samplers and schedulers: The error rates don’t vary much if you follow the above settings. I usually use “dpmpp_2s_a”, which tends to give the most reliable output. Also, the “beta” scheduler almost always gives the best results.
At model weight 1, there is a higher probability of generating the body type differences introduced earlier. When the weight drops below 0.7, the error rate in multi-character images increases significantly.
Also, a experiment I tried: to prevent the AI from overly associating character traits with class tokens like “3boys,” I replaced them with trigger words with higher weights, It seems to be effective. So, without further rambling, here are the trigger words:
福岡紅輔： fukuokakousukeleader, kousuke-red hair, kousuke-black eyes,
kousuke sleeveless, kousuke-red scarf, kousuke-red overskirt, kousuke-black pants, kousuke-red tabi, kousuke-fingerless gloves,
熊本翠： kumamotomidoribaka, midori-green hair, midori-orange eyes,
midori-green hachimaki, midori-green haori, midori-green samue, midori-green rope belt, midori-black shorts, midori-green bracer, midori-green tabi,
鹿児島孝紫： kagoshimakoujisenpai, kouji-purple hair, kouji-red eyes,
kouji-purple kimono, kouji-purple tasuki, kouji-purple waist sash, kouji-black pants, kouji-purple tabi, kouji-purple fingerless gloves,
Negative prompts: nose blush, angry.
When generating solo images, you can use the following to reinforce traits—
But for multi-boys images, these may cause confusion, do not use:
Kouji: thick eyebrows, short eyebrows, bandaged arms, purple bandages, purple puttee
Midori: fang:0.1, fang out
Kousuke: Negative: thick eyebrows, short eyebrows
Positive: grey eyes, black eyes, spiked hair, sidelocks
For three-boys compositions, try replace “3boys” with “3danji.”
CFG: 3 – 5
Model Weight: 0.7 – 1
Sampler + Scheduler: dpmpp_2s_a + beta (or adjust according to checkpoint)
30 steps seems works best. If your composition is correct but there are minor bugs—like issues with eye color, pose, or hands—then, with the same seed, rerunning the process within 25–40 steps often fixes those errors.
Resolution (W × H):
Solo: no limit
2boys: 1024×1536, 1152x1536, etc.
3boys: 1152*1152, 1216*1216, 1152*1344, 1360*1280, 1360*1152, 1024*1216, 1216*1024, 1024*1360, 1360*1024, 1536*1360, etc.
When using "3danji", you can increase the resolution a bit. A very important point: because AI always tries to fill every inch of the frame and doesn’t understand “blank space,” the larger your resolution, the more “empty” background it sees, and ai always tries to fill it.
So you need to carefully adjust both camera distance and resolution to avoid extra people, stray heads, or missing characters. If you get extra characters, tighten the shot (like, upper body) or reduce resolution. If someone’s missing, zoom the camera out (like, cowboy shot) or increase resolution.
This isn’t just for this LoRA—nearly all SDXL‑based anime-style checkpoint follow this principle too.
Prompt Formatting:
1boy: No special requirements.
2boys: Generally okay, but performs better if following the subsequent format.
3boys: Use this structure to improve accuracy:
3boys, male focus,
A-character trigger words, expression, outfit, basic action (like, sitting, lying),
B-character trigger words, expression, outfit, additional action (like, leaning on person),
C-character trigger words, expression, outfit, action add-on (like, side-by-side),
other actions (like, grabbing, holding), other scene elements, lighting, etc.,
quality keywords as per checkpoint instructions.
To avoid errors, try swapping the actions that follow each character trigger words. Also avoid repeating similar actions for hands or legs across characters.
Be cautious with negative prompts: this lora is highly sensitive to them. Follow checkpoint guidance closely—using too many negative prompts can break the image. Only add extra negatives if you want to exclude specific elements.
And here comes a funny failure…a failed experiment led to a "prompt" you should be cautious with. I tried using certain prompt tokens to designate top/bottom roles. In theory: if the tag appears, it’s uke. If not, it must be seme. In practice: nope. It didn’t work at all. Worse, it polluted the tag “penis.” So now I recommend using descriptive words like “erection” replace "penis", or use “kouji penis,” “midori penis,” “kousuke penis” instead. Using “penis” has a high chance of producing a small and deformed result.
From the day I fell into the fandoms, but because of laziness and cowardice, I never thought of drawing or creating things myself. All the cp I loved, all the cold fandoms I followed... it always felt like I was just waiting on someone else to give them life. Never did I expect that one day I could actually create my own doujin works, over 20 years, I can’t quite put the feeling into words.
The next lora is already in the works. It’s about a chance encounter, a journey, a parting, a reunion — a bond between two boys nearly six years apart in age, a bond that’s not just about “brothers,” not just “partners.”
See you next time~