AI Dataset Tagger
Auto-label your image dataset for LoRA and Stable Diffusion training. Upload images, review and edit tags, download a ready-to-train zip — all in your browser.
The fastest way to prepare your Stable Diffusion training dataset
Training a LoRA, DreamBooth, or fine-tuned Stable Diffusion model requires every image to have a paired .txt file listing its descriptive tags — things like watercolor, long_hair, detailed_background, cel_shading. Writing these by hand for hundreds of images is tedious and inconsistent. This tool automates the entire process in your browser, with no Python, no GPU drivers, and no uploads required.
How the tagger works
Upload your images
Drop a folder of images or a zip file directly onto the page. JPG, PNG, and WebP are supported. No file size limits.
Configure your tags
Set your trigger word, adjust the confidence threshold, and add any tags you want excluded — like pose or clothing tags for style LoRAs.
Auto-tag the dataset
WD ViT Tagger v3 runs directly in your browser and tags each image with high-accuracy Danbooru-style tags sorted by confidence.
Review, edit & export
Flip through the image grid, remove bad tags or add missing ones by clicking, then download a zip ready to drop straight into kohya_ss.
Use cases
Style LoRA training
Training a model to replicate an artistic style — watercolor, anime lineart, oil painting, ukiyo-e — requires tagging only the technique, not the subjects. Use the blacklist to strip pose, clothing, and character tags so your LoRA trains purely on style signals. Set a custom trigger word like my_style_v1 and your model will reliably activate on it.
Character LoRA training
Training a specific character — from an anime, game, or original art — needs rich subject tagging: hair colour, eye colour, outfit, and distinguishing features. Leave the blacklist empty and set a low threshold (0.25–0.35) to capture fine details. The tagger identifies character-specific tags at a higher confidence bar automatically.
Fine-tuning & full training
Large-scale fine-tuning runs on hundreds or thousands of images. Manually tagging at that scale is impractical. Batch-tag your entire dataset in one session, then download the labeled zip and point your trainer directly at the extracted folder — no reformatting needed.
Why this tool beats the alternatives
The Stable Diffusion community has several tagging tools. Here's how they compare.
| Tool | Setup required | Batch tagging | Tag editing UI | Privacy | Cost |
|---|---|---|---|---|---|
| This toolYou are here | None — open browser | ✓ Unlimited | ✓ Per-image chip editor | ✓ 100% local | Free |
| A1111 WD14 extension | Python + SD WebUI install | ✓ | ✗ | ✓ Local | Free |
| kohya_ss CLI | Python env + dependencies | ✓ | ✗ CLI only | ✓ Local | Free |
| BooruDatasetTagManager | Windows app install | ✓ | ✓ | ✓ Local | Free |
| HuggingFace Spaces | None | ✗ One image at a time | ✗ | ✗ Uploads to server | Free |
The only tool that combines zero setup, batch tagging, per-image editing, full privacy, and a browser-based workflow in one place.
Powered by WD ViT Tagger v3
WD ViT Tagger v3 is the latest model from SmilingWolf, trained on millions of Danbooru images. It uses a Vision Transformer (ViT) architecture to predict thousands of Danbooru-style tags — from hair colour and clothing to art medium, lighting style, and composition. It's the same model used inside AUTOMATIC1111, ComfyUI, and kohya_ss, now running directly in your browser via WebAssembly.
The model outputs a confidence score per tag. The default threshold of 0.35 gives a good balance of coverage and precision for most datasets. Lower it to 0.2 for more tags, raise it above 0.5 for stricter quality. Character-identity tags are held to a higher bar automatically.
What the output looks like
Each image produces a .txt file with the same base name. The trigger tag is always first, followed by descriptive tags sorted by confidence score, comma-separated. For example:
000.txt
my_artist_style, watercolor, ink_wash, soft_shading, bokeh,upper_body, long_hair, looking_at_viewer, white_background
Drop the exported zip into kohya_ss, SD-scripts, or EveryDream as-is — no renaming or reformatting needed.
Frequently asked questions
Do I need a GPU or special hardware?▾
No. The tagger runs on WebAssembly (WASM) which works on any modern CPU — no GPU, no CUDA, no drivers needed. It's slower than a local GPU setup but fast enough for typical dataset sizes of 50–500 images.
What does the trigger tag do?▾
The trigger tag is a unique word or phrase prepended to every .txt file. Your LoRA model learns to associate this word with the common visual element across your dataset — whether that's an art style, a character, or a concept. Choose something specific and uncommon like 'tide_kubo_style_v2' rather than a generic word.
Why should I use a tag blacklist for style LoRAs?▾
When training a style LoRA, you want the model to learn the artistic technique — not the content. If you leave pose and clothing tags in, your model will partially learn those features too, making it less responsive to your trigger. Blacklisting subject tags (sitting, dress, 1girl, etc.) focuses the training signal entirely on style.
What image formats are supported?▾
JPG, PNG, WebP, and BMP are all supported. You can upload images individually, select a folder, or upload a zip file containing images. Mixed formats in the same dataset are fine.
Is this compatible with Flux, SDXL, and Pony Diffusion?▾
Yes. The .txt sidecar format is the same regardless of the base model. The tags generated are standard Danbooru tags which are supported by Stable Diffusion 1.5, SDXL, Flux, Pony Diffusion, Illustrious, and all community derivatives.
Can I edit tags after auto-tagging?▾
Yes. After tagging, every image shows its tags as clickable chips. Click the × on any chip to remove it, or type in the add field to insert a custom tag. Review and edit before downloading so your dataset is exactly right.
Does the tool work offline?▾
After the first use, yes. The tool is a Progressive Web App (PWA). Once you've used it once, both the app and the tagging model are cached in your browser and work without an internet connection.
Save this tool to your device for offline use
Use Dataset Tagger Offline
Install as a standalone app. AI models are downloaded after install so it works fully offline.