# Mariabell Rosso (マリアベル・ロッゾ)
A SFW character LoRA for Mariabell Rosso of Greed (〝強欲〟のマリアベル・ロッゾ / "gōyoku" no mariaberu rozzo) — the Saintess of Greed and Princess of the Kingdom of Siltrosso, from "That Time I Got Reincarnated as a Slime" (転生したらスライムだった件 / Tensei Shitara Slime Datta Ken / Tensura / TenSura / 転スラ).
The granddaughter of Granbell Rosso and one of the Five Elders, Mariabell is a Reincarnated Otherworlder serving as a major antagonist in the Western Holy Church arc. She appears as a petite young girl with smooth blonde hair, choppy bangs, and signature pink/purple eyes — elegant on the surface, ruthlessly ambitious underneath, wielding the Unique Skill "Greed."
This LoRA was trained on Illustrious base, optimized for flexibility across dark fantasy, modern, and atmospheric scenes.
---
## ⚠️ Content Notice — SFW Only
This LoRA is intended for SFW use only. The character is depicted as a child in canon source material. use is not permitted and is blocked at the model level.
---
## 📺 Source
- Series: That Time I Got Reincarnated as a Slime (Tensura)
- Japanese: 転生したらスライムだった件 / 転スラ
- Character: Mariabell Rosso (マリアベル・ロッゾ)
- Epithet: Mariabell of Greed (〝強欲〟のマリアベル)
- Role: Princess of Kingdom of Siltrosso, Saintess, Five Elders
- First appears: Light Novel Volume 7 — Conflict Between Saints and Monsters / Anime Season 3
- Family: Granddaughter of Granbell Rosso, reincarnation of Maria Rosso (her late grandmother)
---
## ⚡️ Quick Start
Trigger word: mariabelle rosso
Minimal working prompt:
mariabelle rosso, blonde hair, purple eyes, long hair, medallion,
white capelet, red dress, [your scene]
---
## 🎯 Tested Settings
- LoRA strength: 0.9 – 1.1 (sweet spot: 1.0)
- Sampler: DPM++ 2M Karras
- Hires Sampler: DPM++ 2M SDE Karras
- Steps: 28 base + 16 hires
- CFG: 7 base / 5.5 hires
- Hires Upscaler: 4x-AnimeSharp or R-ESRGAN 4x+ Anime6B
- Hires Upscale: 1.5× – 1.7×
- Hires Denoise: 0.40
- Resolution: 1024×1024 or 832×1216
- Clip Skip: 2
Note on strength: This LoRA was trained with dim=8, alpha=4 (low-alpha design), meaning strength 1.0 is already "soft" by construction. No need to lower below 0.8. You can go up to 1.3+ for stronger expression.
---
## 👤 Default Canonical Appearance
For the signature Mariabell Rosso look from Tensura:
mariabelle rosso, blonde hair, purple eyes, long hair,
straight hair, choppy bangs, long sidelocks,
white capelet, red dress, long dress, long sleeves,
frilled dress, black bowtie, medallion, green bracelet,
white pantyhose, bow legwear, mary janes, black footwear
---
## 🎨 Outfit Variations
This LoRA uses detailed captioning — outfits are decoupled from character identity. You can freely swap clothing while keeping Mariabell recognizable.
Identity anchors (always include):
mariabelle rosso, blonde hair, purple eyes, long hair, medallion---
## 🔧 Training Info
- Base model: Illustrious XL v0.1
- Tested compatibility: WAI Illustrious SDXL v1.6.0
- Source: Tensei Shitara Slime Datta Ken — Light Novel / Manga / Anime
- Dataset: 12 hand-curated SFW images, 5 repeats each
- Resolution: 1024×1024 (with bucketing 256-4096)
- Network: dim=8, alpha=4 (LoRA, low-alpha design)
- Optimizer: CAME (betas 0.9, 0.999, 0.9999)
- Learning rate: 1e-4 (UNet) / 2e-6 (Text Encoder, effectively frozen)
- Scheduler: Cosine with 5% warmup
- Loss type: L2
- Min SNR gamma: 5
- Weight decay: 0.02
- Batch size: 2
- Max epochs: 15
- Noise offset: 0.0357
- Multires noise: 5 iterations, 0.25 discount
- Precision: bf16 (full_bf16 + fp8_base) with gradient checkpointing
- Captioning: Detailed (decoupled appearance from trigger), shuffle enabled, keep_tokens=1
---
## 📝 Known Limitations
Bow count and placement inconsistency. Mariabell's canonical outfit features multiple decorative bows (on stockings, accents, etc.). Diffusion models struggle with exact counts and consistent placement — bows may appear, disappear, or shift between generations. Use positional tags like bow legwear, bow on stockings and reinforce with (bows:1.2). For critical shots, ADetailer or manual inpainting gives the best results.
Medallion rendering issues. The signature medallion (green/silver floral cross design) is intricate and often comes out distorted, warped, or with simplified detail. The exact pattern is rarely reproduced 1:1.
Illustrious priors. Base model has strong tendencies. Always use the recommended SFW-focused negative prompt above.
Description
# v2.0 — Reworked dataset, detailed medallion
Full retrain on a rebuilt dataset. Same trigger word and base — drop-in upgrade.
What's improved:
- 🏅 Medallion now actually renders — the signature green/silver floral cross comes out as a recognizable detailed emblem instead of a vague blob. Trained as a dedicated concept with macro close-ups.
- 👁️ Cleaner, more consistent eyes — repaired eye detail, stable purple eyes, less color drift.
- 🎀 More consistent bows — the decorative bows render reliably now instead of shifting between generations.
⚖️ Trade-off (be aware):
- 🏅 The stronger medallion has a downside — because it's now baked more firmly into her identity, it bleeds into unrelated outfits more often (e.g. showing up as an emblem on a school uniform). Negative prompts won't fully it. If you specifically want medallion-free alternate outfits, v1 is the looser, more flexible choice — keep it around for that. v2 is the pick when you want her canonical look with the detailed medallion.
Recommended workflow additions (see description):
- Use ADetailer / hires on the face for best eyes.
- Prompt green bracelet over sleeve sometimes helps to keep the bracelet on the cuff (not the wrist).
Same trigger word mariabelle rosso), same Illustrious base, same recommended settings.
FAQ
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.






