💗 Lover in Law 💗 Kozue, Mai and Tsutomu 💗
(Anima)

This one was a commission, and it turned into the biggest thing I have trained so far. Three characters in a single LoRA, about 1,591 images, roughly 30,000 steps. I went in expecting to fight it the whole way and it came out way better behaved than I deserved. 😅

What you get 💜

All three leads from the 2008 OVA in one file. You can call any of them on their own, put two in a frame, or run all three together. That was the whole point of training it this way instead of shipping three separate models.

The cast 🎭

Kozue Takama: call her on her own and she comes through clean.

Mai Katsuragi: same, and she is the one the model seems most confident on.

Tsutomu Katsuragi: the male lead, and honestly the reason a combined model was worth doing. Training him separately would have been a waste of a file.

Tips 💡

• 🔑 Name them explicitly. With three characters sharing one LoRA, the trigger word is doing real work. Be specific about who you want and the model listens.

• 👥 For multi character scenes, separate them in the prompt. 2girls or 1boy, 2girls plus each character tagged distinctly. Vague prompts are where you get face blending.

• 🎨 Lock their features down individually. Give each one their own hair and eye tags in the prompt. This is the single biggest fix for character bleed, where one face starts leaking into another.

• 🗺️ Regional prompting helps a lot. Regional Prompter in A1111 or masking in ComfyUI keeps each character in their own lane. Worth the setup on anything with three subjects.

• ⚖️ Do not crank the weight for group shots. Around 0.7 to 0.8 is the sweet spot. Higher and the characters start fighting each other for the same face.

• 📺 It is a 2008 OVA, so the source has that late 2000s look. Retro anime tags get you closer to accurate than modern rendering does.

• 🎲 Group compositions need more rerolls than solos. That is just how multi subject prompting goes. Queue extras and be patient.

Thanks 💜

Big thanks to the client who commissioned this one, it was a genuinely fun problem to solve and I learned a lot about multi character training doing it. 🙏 Post your gens, especially group shots, I want to see how far people can push it.

Fuel more of these here: ko-fi.com/deskgrenade ☕