Ganesha (ガネーシャ) — DanMachi LoRA

Please like the model, generate and post images, like my images, collect them, etc before just downloading pwetty please!
I'm almost out of buzz uwu..

ganesha_danmachi,

elephant_mask,

muscular, dark_skin, long_hair

A character LoRA for Ganesha from Is It Wrong to Try to Pick Up Girls in a Dungeon? (ダンジョンに出会いを求めるのは間違っているだろうか / DanMachi / ダンまち).

Ganesha is a male god — 190cm/6'3", extremely muscular build, always wearing his signature red elephant mask with tusks and a trunk. Dark skin, long black hair, gold bracelets, and traditional wrapped clothing.

Usage

Trigger word: ganesha_danmachi

Recommended weights:

• 0.85 — Standard scenes (action, canonical poses, bold compositions)

• 0.90 — Tonally mismatched scenes (quiet, casual, comedic). Add explicit identity tags: elephant_mask, muscular, dark_skin, long_hair

• 1.05 — Maximum before generation breaks down. Use for dramatic emphasis.

• Above 1.05 — Not recommended. Mask becomes exaggerated, muscle definition goes hyper-pronounced.

Best with: WAI Illustrious SDXL v17 and Illustrious-family checkpoints (KageIllustrious, etc.)

Prompt style: Danbooru tags. The LoRA responds well to standard anime/booru tag formatting.

Tips for difficult scenes:

• For scenes that tonally conflict with the character (reading, cooking, paperwork), add explicit identity reinforcement tags alongside the trigger word: ganesha_danmachi, elephant_mask, muscular, dark_skin, long_hair

• SDXL can't parse intent — if you get wrong compositions, rephrase the action rather than using negative prompts. Example: instead of "slam dunk" (always rear view), use "dribbling, facing viewer" or "jump shot, side view".

Training Details

Base model: waiIllustriousSDXL_v170.safetensors

Dataset: ~86 images (anime screencaps + synthetic generations), 5 repeats, with regularization images. Two-pass caption pipeline for tag variety.

Selected epoch: Epoch 9 of 18 — later epochs showed overfitting with loss dropping below 0.08.