---
license: apache-2.0
base_model: beanimaXL_v10
model_type: lora
tags:
  - stable-diffusion
  - sdxl
  - lora
  - character
  - anime
---
# Our Azath — Sakuta & Yui Character LoRA

**Repository:** `toonsquare/Our_Azath`  
**Characters:** Sakuta, Yui  
**Base Model:** `beanimaXL_v10 (SDXL)`  
**Model Type:** Character LoRA (SDXL)  
**License:** Apache-2.0

---

## Overview

![이미지 333333333 (1)](https://cdn-uploads.huggingface.co/production/uploads/63589a3199234d3790317371/haMLHQsCY90kqpUg_JygK.png)
![이미지 4444444444](https://cdn-uploads.huggingface.co/production/uploads/63589a3199234d3790317371/mpj4Hs8V4YuYe9TOAJ865.png)

This repository contains a **character-focused LoRA model** trained for the characters  
**Sakuta** and **Yui** from the IP **_“Our Azath”_**.

The primary objective of this model is to preserve **clear character identity** while supporting stable generation across a wide range of compositions and visual variations.  
Special attention was given to **Yui’s hairstyle**, which is a key visual identifier of the character.

---

## Characters & Trigger Words

### Sakuta
- **Trigger word:** `az_bunta`

---

### Yui

![이미지 111](https://cdn-uploads.huggingface.co/production/uploads/63589a3199234d3790317371/6FJ47HN-WhMzEo81OO8L3.png)

Yui’s hairstyle is a defining feature of her character design.  
To improve generation accuracy and control, **Yui was trained with separate hairstyle-specific trigger words**.

- **Base trigger:** `az_yui`
- **Ponytail hairstyle:** `az_yui, ponytail hair`
- **Hair bun hairstyle:** `az_yui, bun hair`

Using the appropriate trigger word helps maintain consistency between hairstyle and character identity.

---

## Dataset Construction & Training Pipeline

![스크린샷 2026-01-29 16-44-40](https://cdn-uploads.huggingface.co/production/uploads/63589a3199234d3790317371/CPDj37miccYW46fVo7fdt.png)
![스크린샷 2026-01-29 16-44-56](https://cdn-uploads.huggingface.co/production/uploads/63589a3199234d3790317371/n18TqnhsvK-coyP0s0cH-.png)

This model follows a structured, repeatable pipeline designed to maximize character consistency and visual quality.

---

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

High-quality reference images were provided directly by the IP team for both characters.

For each character, images were collected across multiple compositions and viewpoints to ensure consistent identity representation.

---

### 2. Data Augmentation Using Flux

All reference images were processed using a **Flux-based generation pipeline** to produce large-scale augmented datasets.

- Generated images per character: ~**1,000**
- Purpose: increase pose, framing, and contextual diversity

---

### 3. Manual Curation

All generated images were manually reviewed.  
The following were removed during the curation process:

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

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

---

### 4. Yui Hairstyle-Specific Dataset Separation

Unlike Sakuta, **Yui required additional dataset separation due to hairstyle variation**.

To achieve this:

- **Ponytail Yui**
  - Separate reference images were provided by the IP team
  - Images were augmented via Flux
  - Manual curation applied

- **Hair bun Yui**
  - Separate reference images were provided by the IP team
  - Images were augmented via Flux
  - Manual curation applied

Each hairstyle-specific dataset was associated with its own trigger word (`yui_ponytail`, `yui_bun`) to allow precise control during generation.

---

### 5. LoRA Training

The curated datasets were used to train a **Standard LoRA model on top of an SDXL base model**.

- **Base model:** `beanimaXL_v10`
- **Architecture:** Stable Diffusion XL (SDXL)
  
The training process focused on:
- Preserving character identity
- Maintaining hairstyle consistency for Yui
- 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.  
> Values are provided **for reference only**.

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

```json
{
  "LoRA_type": "Standard",
  "epoch": 50,
  "train_batch_size": 3,
  "learning_rate": 3e-05,
  "unet_lr": 3e-05,
  "text_encoder_lr": 3e-05,
  "optimizer": "AdamW",
  "mixed_precision": "fp16",
  "network_dim": 32,
  "network_alpha": 32,
  "clip_skip": 1,
  "max_resolution": "1024,1024",
  "min_snr_gamma": 5,
  "gradient_checkpointing": true,
  "cache_latents": true,
  "enable_bucket": true,
  "save_every_n_epochs": 1,
  "sdxl": true,
  "save_precision": "fp16"
}
```

</details>

---

## Usage Guide

### Basic Prompt

```text
az_bunta, [description]
```

```text
az_yui, [description]
```

### Hairstyle-Specific Prompts (Yui)

```text
az_yui, ponytail hair, [description]
```

```text
az_yui, bun hair, [description]
```

### Tips

* Place the character trigger early in the prompt for stronger identity locking.
* Always use the hairstyle-specific trigger for Yui to avoid hairstyle mixing.
* Explicitly specifying composition (e.g., close-up, bust shot, full body) improves control.
* Extreme stylistic deviations may weaken character consistency.

---

## Intended Use & Limitations

* This model is designed for **character-consistent image generation** of Sakuta and Yui.
* Hairstyle accuracy for Yui depends on using the correct trigger word.
* Mixing conflicting hairstyle prompts may reduce consistency.
* Radical style changes or photorealistic prompts may degrade character identity.

---

## 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

* Characters: **Sakuta, Yui** (*Our Azath*)
* Training & curation: **Toonsquare**
* Data source: IP Team