---
language: en
thumbnail: ./assets/example.png
tags:
- text-to-image
- diffusion models
- LoRA fine-tuning
- animagine-xl-3.0
- stable_diffusion_xl
- kohya_ss
- waifu2x
license: apache-2.0
model:
  name: GirlsFrontline2-SDXL-LoRA
  description: A model based on SDXL with LoRA fine-tuning for 《Girl's Frontline 2》 Text-To-Image generation.
  pipeline_tag: image-generation
  repo: https://huggingface.co/TfiyuenLau/GirlsFrontline2_SDXL_LoRA
  library: huggingface
  framework: pytorch
  version: 1.0.0
  pretrained_model: stable_diffusion_xl
  base_model: animagine-xl-3.0
  fine_tuner: kohya_ss
  data_augmentation: waifu2x
  task: text-to-image
---

# 基于SDXL模型LoRA微调实现《少前2：追放》文生图

![example](./assets/example.png)

## 一、Model Library
1. 微调数据集：[基于SDXL模型的《少女前线2：追放》LoRA微调数据集](https://www.kaggle.com/datasets/yukikonata/sdxl2lora)
2. 预训练模型：[stable_diffusion_xl](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl)
3. 底模：[animagine-xl-3.0](https://huggingface.co/cagliostrolab/animagine-xl-3.0)
4. SDXL LoRA微调训练器：[kohya_ss](https://github.com/bmaltais/kohya_ss)
5. 数据集画质增强：[waifu2x](https://github.com/nagadomi/waifu2x)

## 二、Prompt Dict
1. 少前2追放角色
* 佩里缇亚: PKPSP
* 塞布丽娜: SPAS12
* 托洛洛: AKAlfa
* 桑朵莱希: G36
* 琼玖: QBZ191
* 维普雷: Vepr12
* 莫辛纳甘: MosinNagant
* 黛烟: QBZ95
* 克罗丽科: Kroliko
* 夏克里: XCRL
* 奇塔: MP7
* 寇尔芙: TaurusCurve
* 科谢尼娅: APS
* 纳甘: Nagant1895
* 纳美西丝: OM50
* 莉塔拉: GalilARM
* 闪电: OTs14

2. Pixiv画师风格
* おにねこ（鬼猫）: Onineko26
* 麻生: AsouAsabu
* mignon: Mignon
* migolu: Migolu

## 三、使用方式
1. 安装部分环境（默认已安装pytorch等必要环境）
~~~sh
pip install diffusers --upgrade
pip install transformers accelerate safetensors
~~~

2. 使用Hugging Face下载并使用底模（animagine-xl-3.0）和LoRA模型
~~~python
import torch
import datetime
from PIL import Image
import matplotlib.pyplot as plt
from diffusers import (
  StableDiffusionXLPipeline, 
  EulerAncestralDiscreteScheduler,
  AutoencoderKL
)

# LoRA Hugging Face ID
lora_id = "TfiyuenLau/GirlsFrontline2_SDXL_LoRA"

# Load VAE component
vae = AutoencoderKL.from_pretrained(
  "madebyollin/sdxl-vae-fp16-fix", 
  torch_dtype=torch.float16
)

# Configure the pipeline
pipe = StableDiffusionXLPipeline.from_pretrained(
  "cagliostrolab/animagine-xl-3.0", 
  vae=vae,
  torch_dtype=torch.float16, 
  use_safetensors=True, 
)
pipe.load_lora_weights(lora_id)
pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
pipe.to('cuda')
~~~

3. 生成图像
~~~python
# Define Prompt
output = "./output.png"
prompt = "1girl, OTs14, gloves, looking at viewer, smile, food, holding, solo, closed mouth, sitting, yellow eyes, black gloves, masterpiece, best quality"
negative_prompt = "nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name"

# Generate Image
image = pipe(
  prompt, 
  negative_prompt=negative_prompt, 
  width=1024,
  height=1024,
  guidance_scale=7,
  num_inference_steps=28
).images[0]

# Save & Show
image.save(output)
image = Image.open(output)
plt.axis('off')
plt.imshow(image)
image.close()
~~~