File size: 6,123 Bytes
9aa860d 906fbd5 1af8afb 906fbd5 da0da36 906fbd5 67cfdf6 906fbd5 1af8afb 906fbd5 1af8afb 906fbd5 1af8afb 906fbd5 1af8afb 906fbd5 1af8afb 906fbd5 1af8afb 906fbd5 1af8afb 906fbd5 1af8afb |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 |
---
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
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
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
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 |