---
license: apache-2.0
base_model: astranime_V6
model_type: LoRA
tags:
  - stable-diffusion
  - sd-1.5
  - lora
  - lycoris
  - lokr
  - character
  - anime
trigger_word: gb_yuuma
---

# Homewreker — Yuuma Character LoRA

- **Repository:** `toonsquare/Breaking_the_Perfect_Couple`
- **Trigger Word:** `gb_yuuma`
- **Base Model:** `astranime_V6 (Stable Diffusion 1.5)`
- **License:** Apache-2.0

---

## Overview

![이미지 2222222222222](https://cdn-uploads.huggingface.co/production/uploads/63589a3199234d3790317371/MAM3bylNTmBhPHXK8tPyO.png)

This repository contains a **character-specific LoRA model** trained for **Yuuma**, a character appearing in the webtoon **_“Breaking the Perfect Couple”_**.

The goal of this model is to **preserve Yuuma’s facial features, hairstyle, and overall character identity** while enabling stable generation across a wide range of compositions, from close-up portraits to full-body shots.

---

## Trigger Word

- **Required trigger:** `gb_yuuma`

Including `gb_yuuma` in your prompt helps strongly anchor the generation to Yuuma’s character traits.

---

## Dataset Construction & Training Pipeline

![스크린샷 2026-01-29 16-43-19](https://cdn-uploads.huggingface.co/production/uploads/63589a3199234d3790317371/DUJth5OJWxOMdEAlGo2JZ.png)

This model was not trained on raw collected images alone.
Instead, a **character-consistency-first data pipeline** was designed and executed.

### 1. Original Image Acquisition (from IP Team)

High-quality reference images were provided directly by the IP team, covering the following compositions:

- **Close-up**
- **Bust shot**
- **Knee shot**
- **Full shot**

Each composition includes:
- **Front view**
- **Left-side view**
- **Right-side view**

This structure ensures that Yuuma’s facial features, hair, and silhouette remain consistent across different viewing angles.

### 2. Data Augmentation Using Flux

Using the original images as input, a **Flux-based generation pipeline** was used to create a large augmented dataset.

- **Total generated images:** ~1,000
- **Generation time:** ~12 hours

### 3. Manual Curation (Critical Step)

All generated images were **manually reviewed and curated**. The following were removed:

- Duplicate or near-duplicate images
- Incorrect generations (identity drift, malformed anatomy, etc.)
- Low-quality outputs (noise, broken composition, poor details)

After curation, approximately **500–600 images** were selected for training.

### 4. LoRA Training

The curated dataset was used to train a LoRA model on top of:

- **Base model:** `astranime_V6`
- **Architecture:** Stable Diffusion 1.5

The training focused on maintaining strong character identity while allowing reasonable variation in pose, framing, and context.

---

## LoRA Training Configuration (Reference)

> ⚠️ The following configuration reflects the settings used during training as closely as possible.
> Some values may not be exact and are provided **for reference only**.

<details>
<summary>Click to expand training configuration</summary>

```json
{
  "LoRA_type": "LyCORIS/LoKr",
  "LyCORIS_preset": "full",
  "adaptive_noise_scale": 0.005,
  "bucket_no_upscale": true,
  "cache_latents": true,
  "cache_latents_to_disk": true,
  "caption_dropout_rate": 0.05,
  "clip_skip": 1,
  "epoch": 20,
  "learning_rate": 0.0001,
  "unet_lr": 0.0001,
  "text_encoder_lr": 0.0001,
  "optimizer": "AdamW",
  "mixed_precision": "bf16",
  "network_dim": 100000,
  "network_alpha": 1,
  "max_resolution": "768,768",
  "min_snr_gamma": 5,
  "seed": 1234,
  "save_precision": "fp16"
}

```

</details>

---

## Usage Guide

### Basic Prompt

```text
gb_yuuma, [description]

```

### Tips

* **Positioning:** Place `gb_yuuma` early in the prompt for stronger character locking.
* **Composition:** Explicitly specifying the composition (e.g., close-up, bust shot, full body) improves control.
* **Style:** Extreme stylistic deviations may weaken character consistency.

---

## Intended Use & Limitations

* This model is designed for character-consistent image generation of **Yuuma**.
* It performs best when used with **anime-style SD 1.5 checkpoints** similar to the base model.
* Forcing radically different art styles or photorealistic prompts may reduce identity stability.

---

## License

This model is released under the **Apache License 2.0**.

You are free to:

* Use the model commercially
* Modify and redistribute it

Please ensure compliance with the Apache-2.0 license terms and any applicable IP usage policies.

---

## Credits

* **Character:** Yuuma (Breaking the Perfect Couple)
* **Training & Curation:** Toonsquare
* **Base Model:** astranime_V6

```

```