ONOKO - AI Girl (SDXL char LoRA)

# ONOKO — AI Girl SDXL LoRA (v3 patch_a)

> **Original character LoRA** for stably reproducing **ONOKO**, an original Y2K / near-future street-aesthetic anime girl, on SDXL / Illustrious-family checkpoints.

---

## 🎯 Character

ONOKO is an original character with the following signature traits (baked into the trigger):

- Short, slightly messy black bob hair with vivid cobalt-blue **inner color** (front + bottom strands)

- Bright sapphire-blue eyes

- White-framed sunglasses with blue mirror lenses, usually worn on top of the head

- Star-themed accessories: blue star hair clip, star necklaces

- Slim, late-teen build

- Y2K / near-future / street style

---

## ⚡ Quick start

```

Positive:

<lora:onoko_0n0k0_sdxl_v3_patch_a:0.85> 0n0k0, black hair, blue eyes,

portrait, looking at viewer, simple background, studio lighting,

masterpiece, best quality

Negative:

lowres, bad anatomy, bad hands, watermark, signature, text, logo,

multiple girls, 2girls, malformed face, extra fingers, jpeg artifacts,

white hair, silver hair, gray hair, blonde hair, brown hair,

covered face, hair over face

```

- **Trigger**: 0n0k0

- **Recommended weight**: 0.85 (range: 0.75 – 0.95)

- **Include black hair, blue eyes explicitly** for best stability under strong-lighting prompts.

---

## 📷 Showcase prompts (the previews)

| Image | Prompt sketch |

|---|---|

| 01 — Main / Signature | <lora:...:0.85> 0n0k0, black hair, blue eyes, full body, transparent jacket, plaid skirt, sunset city background |

| 02 — Portrait studio | <lora:...:0.85> 0n0k0, black hair, blue eyes, portrait, looking at viewer, neutral expression, white background, studio lighting |

| 03 — Expression: angry | <lora:...:0.9> 0n0k0, black hair, blue eyes, portrait, angry expression, frown, white background, dramatic lighting (with hair-color-ban negative) |

| 04 — Expression: surprised | <lora:...:0.85> 0n0k0, black hair, blue eyes, portrait, surprised, open mouth, wide eyes, simple background |

| 05 — Outfit: school uniform | <lora:...:0.9> 0n0k0, black hair, blue eyes, full body, school uniform, sailor uniform, standing, classroom background |

| 06 — Composition: back view | <lora:...:0.85> 0n0k0, black hair, blue eyes, full body, from behind, looking over shoulder, sunset background |

---

## 🛠️ Training details

| Parameter | Value |

|---|---|

| Framework | kohya **sd-scripts** (commit 502cc3f) |

| Base model | n4mik4ILSFWNSFW_v20.safetensors (SDXL / Illustrious community checkpoint) |

| Network | LoRA-LierLa, **dim 16 / alpha 8** |

| Train target | UNet only network_train_unet_only=true) |

| Resolution | 1024 with bucketing (512–2048, no upscale) |

| Optimizer | **Prodigy**, lr=1, d_coef=1.0, decouple=True, safeguard_warmup=True |

| Scheduler | cosine, no warmup (Prodigy handles it) |

| Loss tricks | min_snr_gamma=5, multires_noise_iterations=6, multires_noise_discount=0.3 |

| Precision | bf16 + fp8_base + full_bf16 |

| Caching | latents (RAM) + TE outputs (disk) |

| Schedule | 314 images (269 base + 45 GPT Image 2 addon), 5 epochs × repeats (5/5), **3925 optimizer steps total** |

| Final loss | 0.118 |

| Trained on | RTX 4060 Ti 8GB |

**v2 → v3 (patch_a) delta**: 45 simple-background addon images generated via ChatGPT GPT Image 2 to fix v2's residual portrait + simple-bg weakness. Same optimizer / network / loss settings as v2 for clean attribution.

---

## ⚠️ Known limitations

1. **Hair color drift under dramatic lighting**: if you do **not** include black hair in the prompt, strong-lighting prompts dramatic lighting, hard rim light, etc.) can cause the model to render lighter / silver hair. The Quick-start template above prevents this.

2. **Face occlusion in full-body shots**: ONOKO's training data includes many "looking away" / hair-cascading poses. Add looking at viewer, face visible if you need a front-facing face in full-body compositions.

3. **Best weight is 0.85**: increasing weight to 0.95+ does **not** lock the character better; it only over-saturates style elements. Use prompt tags, not weight, to nail consistency.

---

## 🏷️ Suggested Civitai tags

```

character, original character, anime, sdxl, illustrious,

y2k, near future, street fashion, blue hair, blue eyes,

black hair, cyberpunk, character lora, oc, female

```

---

## ✅ AI / data disclosure

- This LoRA was trained on a mix of:

- Images generated by the prior onoko_ai_girl_n4mik4_sdxl_v1 LoRA (author's own prior model) on the **n4mik4ILSFWNSFW_v20** base checkpoint

- Images generated by **OpenAI ChatGPT GPT Image 2** with author-defined prompts (45-image addon for v3)

- All training images are AI-generated; **no real-person likenesses or copyrighted character references** were used.

- ONOKO is an **original character (OC)** created by the author; commercial use of this LoRA's outputs is permitted per the license below.

---

## 📜 License (suggested settings on Civitai)

- **Use the LoRA for image generation**: ✅ allowed

- **Sell images generated with the LoRA**: ✅ allowed (commercial use OK)

- **Use the LoRA as a building block in a merge**: ✅ allowed with attribution

- **Re-host / re-distribute the .safetensors**: ❌ no (point users to the original Civitai page)

- **Train other LoRAs on top of this LoRA**: ✅ allowed with attribution

(Adjust to taste on the Civitai upload form.)

---

## 🔁 Versions

| Version | Date | Note |

|---|---|---|

| v1 (SDXL) | 2026-05-18 | First SDXL release, 104 images, AdamW8bit, 768 res |

| v2 | 2026-05-19 | Major rewrite: 269 images, Prodigy, 1024 res, dim 16/8. Trigger changed to 0n0k0. |

| **v3 patch_a** | **2026-05-20** | **+45 GPT Image 2 simple-bg addon, portrait weakness fixed, expression range expanded** |

---

## 🙏 Credits

- Base checkpoint: n4mik4ILSFWNSFW_v20 (community SDXL / Illustrious model)

- Training framework: [kohya-ss / sd-scripts]( Tagger: [SmilingWolf / wd-eva02-large-tagger-v3]( Auxiliary asset generation: ChatGPT GPT Image 2 (OpenAI)

---

If you enjoy ONOKO, please ❤ the model and share your prompts in the comments. Feedback on edge-case prompts is especially welcome and feeds into v4 planning.