About ImgTagger

ImgTagger is a small, independent project that makes free tools for tagging and cleaning image collections. Everything runs in your web browser. There are no accounts and no uploads, and nothing is hidden behind a paywall.

Why it exists

The tools started as a personal toolkit for preparing image datasets. Anyone who has trained a LoRA or organized a large illustration archive knows the routine: weed out duplicates, throw away images that don't fit, fix the odd washed-out file, then write a caption for every single image. The existing options were either command-line scripts that need a Python environment and often a GPU, or web services that ask you to upload your files to someone else's server.

We wanted something in between: a page you can open, drop a folder onto, and work through visually, without your images ever leaving your machine. Modern browsers can now run neural networks with WebAssembly, and that made this possible.

What we make

• Anime & Danbooru Tagger and Anime Character Tagger: drop in one picture to get its tags or find out which character it shows, ready to copy into a prompt.
• Batch WD14 Tagger & Editor: runs the open WD14 tagger models by SmilingWolf to produce Danbooru-style tags, with tools to review, edit and export them as caption files.
• Duplicate Image Finder: groups look-alike images using perceptual hashing.
• Face & Expression Sorter: sorts images by whether they contain faces, how many, and the detected expression, using the open-source face-api.js models.
• Batch Color Fixer: white balance, histogram normalization and basic tone controls for many images at once.

Alongside the tools we publish guides on the things we had to learn the hard way: how booru tags work, how to choose a tagger model and threshold, how to caption a training set, and how to clean a messy folder.

How we write and check our guides

Every guide on this site follows a few rules:

• Claims about our tools come from the code. When a guide says the default threshold is 0.35, or that tag files are matched by filename, that's because the code does exactly that. We keep an internal facts sheet tied to the source and re-check it when the tools change.
• Numbers are measured, not guessed. Figures such as model download sizes and tag counts come from the published model files themselves, and the guides say where each one came from.
• Third-party details are checked against primary sources. These include the Danbooru wiki, model cards on Hugging Face, trainer documentation on GitHub and Google Search Central. Where practice varies, we say so instead of presenting one opinion as fact.
• We say what the tools can't do. If a model is weak on photographs, or a detector misses stylized faces, the guide tells you.
• Pages are dated. Each guide shows when it was last updated. If you spot something wrong or out of date, please tell us. Corrections are always welcome.

The technology

The tagger uses ONNX Runtime Web to run the models in WebAssembly, spread across a pool of Web Workers so that several images are processed at once. Model files are downloaded from Hugging Face and, when your browser allows it, cached in its IndexedDB storage so later visits skip the download. The face sorter uses face-api.js. The duplicate finder and color fixer use plain JavaScript and the canvas API. How in-browser AI works goes into detail, including how to confirm for yourself that nothing is uploaded.

Credits and licenses

The WD14 tagger models are published by SmilingWolf on Hugging Face. The model cards for the two v2 models (ConvNeXt V2 and Swin V2) state the Apache 2.0 license; the ViT repository doesn't state a license. The models were trained on images and tags from Danbooru, whose volunteer taggers built the vocabulary that most anime-style image tools now share. face-api.js is MIT-licensed open-source software by Vincent Mühler. ONNX Runtime Web (Microsoft, MIT License) runs the tagger models, and JSZip (MIT or GPLv3) builds the ZIP downloads. We're grateful to all of them.

How the site is funded

ImgTagger is free to use. Hosting and development are paid for by advertising through Google AdSense and by optional donations on Ko-fi. Our code never passes the images or tags you process to the ad scripts, and nothing in the tools sends them anywhere. Our privacy policy explains what advertising cookies do.

Get in touch

Bug reports, feature ideas and corrections are all welcome. See the contact page.