AI Image Auto Tagger

The successor of WD14 tagger and alternative to DeepDanbooru - now with metadata saving feature for nsfw-oriented gallery tagging (optimized for digiKam). Currently using wd-vit-tagger-v3 model (dataset as of 2024-02-28) by SmilingWolf which is newer than WD14 and DeepDanbooru. Using CUDA and ONNX library over Gradio WEBUI. Tested on Windows.

Features

• Supported models: Latest WaifuDiffusion v3 tagger architecture featuring SmilingWolf/wd-vit-tagger-v3 model
• Supported files: JPG/JPEG (recommended), PNG, WEBP, GIF, BMP(no metadata)
• Easy user interface: By utilizing gradio for GUI, the usage of this script should be smooth
• Process subdirectories: recursively goes through all directories within a given one. In case of writing tags to .txt it mirrors the folder structure
• User preferred threshold: Using the gradio slider, the user can adjust the threshold of the tagger model
• Hide rating tags: You can optionally choose whether to output the rating tags (Like "General", "Explicit", "Questionable", etc) or not by checklist the "Hide Rating Tags" box
• Character tags first: This feature makes the character name tag appear in front before other tags like general, copyright or rating. This feature is useful when training the text encoder with "keep n tokens"
• Remove separator: This function will remove the standard separator "_" of the tags in the output caption
• Overwrite existing metadata tags: wipes clean any existing tags in metadata before writing new ones (XMP:Subject and IPTC:Keywords)
• Output: There are 2 output modes: 1. embedding tags directly into image metadata 2 .txt files

How to run

Python >3.10 and CUDA GPU is required to run this script. Download from https://www.python.org/downloads
ExifTool >12.15 is required. Download from https://exiftool.org

Steps to run:

1. Clone this repository git clone https://github.com/Deiwulf/AI-image-auto-tagger.git OR download as a zip and extract
2. Navigate to directory cd AI-image-auto-tagger
3. Set up a virtual environment python -m venv venv *
4. Activate the new venv: *
   • Windows: venv\scripts\activate
5. Install the requirements pip install -r requirements.txt
   • Optionally visit https://pytorch.org/ and install one fitting your system for performance boost (~69% here)
6. Run the script python wdv3tagger.py OR use start.bat on Windows (using venv)

* Virtual environment is optional, but recommended to keep this isolated, you can skip to step 5 if you want to install and run on global environment

Disclaimer

This has been thoroughly tested, but it's still under development so do be savvy and backup before running, and report issues if any.