Utsu / Utsuro - Gokurakugai LoRA (Illustrious)
============================================================
She looks sweet. She looks elegant. Don't let that fool you — she'll cut you the moment you look away.
Utsuro from Gokurakugai, Yuto Sano's underrated gem. Yandere, obsessive, dangerously in love. Short messy bob, teal inner gradient, sanpaku stare — the kind of character that makes you nervous in the best way.
Trained on 44 images — mix of close-up expressions and full-body poses. Painterly style with soft rim lighting. LoRA Manager compatible.
🏷️ TRIGGER WORD/ IDENTITY BLOCK
============================================================
utsurogokurakugai, black hair, bob cut, long side locks, teal inner hair, bangs, sanpaku eyes, dark red eyes, black pupils, thick eyelashes, black cheongsam, high mandarin collar, long sleeves, black dress, black pantyhose, black footwear
😈 EXPRESSION VARIATIONS
============================================================
The dataset covers 8 expressions. Add these to your anchor:
😬 YANDERE GRIN
(trigger word)
wide grin, yandere, dark room, crazy_eyes, crazy_smile, heart_eyes, blush
🥰 DEVOTED LOVE
(trigger word)
gentle smile, soft warm eyes, hand on cheek, golden hour lighting, blush, red cheek
😢 VULNERABLE CRYING
(trigger word)
crying, tears, sad, open mouth, blush, teary_eyes
😤 POUTING SULKING
(trigger word)
pout, pouting, pouting_lips, annoyed, looking away, sideways glance, bright room, cute
😊 MELANCHOLIC SMILE
(trigger word)
bittersweet smile, heavy-lidded eyes, big smile, cheerful expression, heart_eyes
👁️ MENACING STARE
(trigger word)
serious, expressionless, sanpaku stare, cold lighting, dark moody room, mad face, angry, looking_at_viewer
🌻 CUTE JOY
(trigger word)
joyful smile, hands together, sunlight, garden background, bokeh
😏 PLAYFUL TEASE
(trigger word)
playful smirk, tongue out, mischievous, looking at viewer
⚙️ RECOMMENDED SETTINGS
============================================================
LoRA weight: 0.7 to 1.0
CFG Scale: 5 to 5.5
Sampler: DPM++ 2M Karras or Euler A
Steps: 25 to 35
This LoRA can get dominated when stacked with other LoRAs. If the identity starts drifting — hair turning wrong color, eyes losing the sanpaku shape — bump the weight above 1.0 (try 1.1 or 1.2). The 16-token identity block helps, but sometimes brute force is the answer.
For the best results, run at native 1024x1024 resolution then upscale with your preferred upscaler. The painterly texture and fine hair details really shine after upscaling.
❌ NEGATIVE PROMPT
============================================================
lowres, bad anatomy, bad hands, text, error, missing fingers, cropped,
worst quality, low quality, jpeg artifacts, signature, watermark,
⚠️ KNOWN PROBLEMS
============================================================
Heart-shaped pupils sometimes leak from the training data. Add heart pupils to negative if unwanted.
At high LoRA weights (above 1.0), the painterly brush texture becomes too strong. Drop to 0.8 for cleaner results.
The teal inner hair can extend too far down at times. Add short hair with long locks to help contain it.
Background bleeding — the dark room and golden sunset backgrounds are strong in the dataset. Use plain white background in positive and add scenery, room, gradient, blurry background to negative if you want clean outputs.
📊 TRAINING DETAILS
============================================================
Base Model: Illustrious XL v20
Network: LoRA — Dim 16 / Alpha 8 (scale 0.5)
Epochs: 8
Batch Size: 2
Optimizer: AdamW with Cosine LR
Resolution: 1024x1024 (bucket training, min 512 max 2048)
Total Steps: 1,760
Final Loss: 0.098
Images: 44
Captions: WD14 Tagger + manual enhancement
keep_tokens: 11
Tag dropout: 0%
✅ COMPATIBILITY
============================================================
Tested and consistent on:
- WAI-illustrious-SDXL-v17
- bismuthIllustrious_v80
- ntrMIXIllustriousXL_xiii
- Illustrious XL v20 (trained on)
Results are consistent across all four, but slight adjustments may be needed per checkpoint — CFG, weight, or sampler. Works on other Illustrious SDXL checkpoints. Not tested on Pony or SD 1.5.
💡 TIPS
============================================================
This LoRA is LoRA Manager compatible. Drop the file into your ComfyUI models/loras folder and use the Trigger Word Toggle node.
The 16-token identity block is all you need. Copy, paste, add expression.
For painterly style images, keep weight at 0.8-1.0. For cleaner anime cel-shading, drop to 0.6 and add rough brushstrokes, painterly texture to negative.
When stacking with other LoRAs, bump Utsuro weight to 1.0-1.2. The full 16-token block anchors stronger than a single trigger word.
Say hi in the comments if you generate something cool!
🎨 COMMISSIONS & SUGGESTIONS
============================================================
Got a manga character you want turned into a LoRA? Want a specific outfit, pose, or style variation? I take suggestions — and yes, I do commissions.
If you want your own custom LoRA trained (character, style, outfit — anything), hit me up. I work on dataset curation and training with you to get the result you want.
Drop a comment or DM. Always happy to chat about new ideas.