[Newbie Image] furry body structure correction -Test plan -interlude

# NewBie Anatomy Fixer: Bara Furry (LoKr)

Anatomy Improvement & Style Stabilization for NewBie Model

## ℹ️ Model Overview

This is a LoKr (Lycoris) model trained specifically to improve the anatomical correctness and "Bara Furry" style consistency of the NewBie base model.

NewBie is a promising but experimental architecture (Lumina-T2I). Its base state often struggles with furry anatomy, perspective, and limb logic. This model aims to mitigate these issues, though it is not a magic cure-all.

## 📊 Training Data (Quality > Quantity)

* Source: Filtered from a raw dataset of over 20,000+ images.

* Selection: Rigorously cleaned down to ~1,400 high-quality images.

* Criteria: The "Level 1" dataset focuses strictly on correct body structure, perfect perspective, and clear muscle definition (Bara style).

---

## ⚠️ Recommended Settings (Read Carefully)

| Setting | Recommendation | Notes |

| :--- | :--- | :--- |

| Network Type | LoKr | Ensure your UI supports Lycoris/LoKr. |

| Weight | 0.6 ~ 0.8 | Best stability. |

| Trigger Words | (Optional) | See example tags. |

> IMPORTANT:

> At weight 1.0, the model may become too rigid and introduce artifacts (e.g., 6 fingers or stiff poses).

> 0.8 is the sweet spot found during testing.

---

## 📝 Prompting Strategy

This model was trained with a complex tagging scheme. For the best results, I recommend a Hybrid Approach. Of course, you can also choose only one of them:

1. XML Structure (Please refer to the official model usage documentation for the officially recommended usage format: https://my.feishu.cn/wiki/NZl9wm7V1iuNzmkRKCUcb1USnsh).

2. Natural Language Caption (This section is similar to the prompt used in FLUX.).

3. Danbooru Tags (This section is similar to the prompt used in SDXL.).
(Please refer to the prompt metadata in the showcase images for the exact style.)

---

🛑 Scope & Limitations

• Target Subject: Strictly for Male Bara Furry characters.

• Nature of the Fix: The NewBie base model has inherent issues with spatial perspective and furry generation ("it can be a mess"). This LoKr improves the success rate and anatomical logic but cannot completely fix the base model's fundamental flaws.

• Architecture: NewBie is NOT an SDXL model. It uses the Lumina architecture. It is still in a very young, experimental stage.

🔗 About the Base Model

NewBie is a challenging model for both trainers and users, but it holds great potential for the future. If you want to learn more about the architecture or download the base checkpoint, please visit: NewBie Model