TOUHOU VISION AI

A Streamlit-based Touhou Project character classifier built on PyTorch. The app includes Grad-CAM visualizations to explain model predictions and an in-app Pixiv crawler to help you collect and curate training data.

Pre-trained Release

Want to use the classifier without setting up a development environment?

Download the latest release

Quick Start (Release Version)

1. Download and extract the zip file
2. Double-click start_classifier.bat
3. If uv is not installed, it will be installed automatically
4. Wait for dependencies to install (first run only)
5. A browser window will open with the classifier

🩵Supported Characters (21 total)

• Alice アリス・マーガトロイド
• Cirno 琪露诺
• Flandre 芙兰朵露
• Inubashiri 犬走椛
• Kaguya 蓬莱山輝夜
• Kochiya 東風谷早苗
• Koishi 古明地恋
• Kokoro 秦こころ
• Konpaku 魂魄妖夢
• Marisa 霧雨魔理沙
• Meiling 紅美鈴
• Miku 初音ミク
• Mokou 藤原妹紅
• Reimu 博麗霊夢
• Reisen 鈴仙・優曇華院・イナバ
• Remilia 蕾米莉亚
• Sakuya 十六夜咲夜
• Satori 古明地さとり
• Shameimaru 射命丸文
• Suika 伊吹萃香
• Yuyuko 西行寺幽々子

Features

• Grad-CAM visual explanations
  • Generates heatmaps and overlays them on the input image
  • Clarifies why a character was predicted
• Pixiv crawler inside Streamlit
  • Search by keyword or tag
  • Filter by illustration type and options
  • Sort results to keep dataset collection organized
• Browser-based Pixiv refresh token flow
  • Obtain your Pixiv refresh token headlessly by providing your credentials in the .env file

Dataset

• Sample dataset: https://drive.google.com/drive/folders/1g0fvx9OwYUaohgFOwiKB1FCoqBD72DZA?usp=sharing

Installation (For who want to train the model)

1) Install uv

Windows:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

macOS:

curl -LsSf https://astral.sh/uv/install.sh | sh

2) Clone the project

git clone https://github.com/YYDongRo/TouhouCharClassifier.git

3) Enter your Pixiv credentials in the .env file

Feel free to use other images source

PIXIV_USERNAME=""
PIXIV_PASSWORD=""

4) Run the program for inference

uv run streamlit run app.py

5) Train your own model

Organize your images as one folder per class:

data/
  reimu/
  marisa/
  cirno/

Train the model using:

uv run python -m src.train