Which tool do I need?
| You want… | Use | What you get |
|---|---|---|
| A prompt from one anime-style picture | Anime Tagger | Tags to copy, with a slider for more or fewer, spaces or underscores. |
| A Danbooru-format string for an upload | Danbooru Tagger | Tags with post counts, current tag names and a rating tag. |
| Tags separated with underscores for an image board or caption | Booru Tagger | The same one-picture tagger, set up for booru-style tags. |
| To know which character is in a picture | Anime Character Tagger | One of 2,132 popular anime and game characters. |
| Labels for a whole training folder | Batch WD14 Tagger & Editor | Tags for every picture, searchable and editable in bulk, saved as .txt files. |
| To edit captions you already have | Batch WD14 Tagger & Editor | Drop the pictures and their .txt files together. |
| A sentence or a paragraph for each picture | Image Caption Generator | Short, detailed or very detailed captions, as .txt files or a CSV. |
| Alt text for a web page | Alt Text Generator | A short description for each picture, ready to check and paste. |
Tags or captions?
The same picture can be described two ways, and the right one depends on what will read it. Tags are a list of short words from a fixed vocabulary, such as 1girl, long_hair, smile, outdoors. Captions are ordinary sentences, such as "A woman with long hair smiling in a park". Many anime image models were trained on booru tags, so tags are their native language; newer models usually expect sentences.
| Tags (WD14 taggers) | Captions (Image Caption Generator) | |
|---|---|---|
| Looks like | A list of words from one vocabulary of 9,083 tags | Free English sentences |
| Trained on | Millions of tagged anime pictures | General pictures and their descriptions (Florence-2 base, by Microsoft, MIT license) |
| You control | A confidence slider: each tag comes with a score | The length, and a start phrase such as a trigger word |
| Where it goes wrong | A wrong or missing tag, which the score and the slider help with | Invented details: in our test of eight pictures, four of the detailed captions described something that was not there |
| First download | 388 MB (ConvNeXt V2, the default) | About 343 MB on a graphics chip, about 265 MB on a processor |
Our guide on tags versus natural-language captions goes into which one to use for which kind of model, and the guide on captioning a LoRA dataset shows both in a real training set.
Which WD14 model?
The Batch WD14 Tagger offers three models by SmilingWolf, and the one-picture taggers use the first. ConvNeXt V2 (388 MB, 9,083 tags, with character tags) is a good default. Swin V2 (455 MB) has the same tags and gives a second opinion. ViT (365 MB) has fewer tags (6,549) and cannot name characters. Our model comparison guide has the details.
A workflow for a training set
- Clean the folder first. Remove copies with the image organizer tools, so you do not tag the same picture twice. Resize and crop to your training size with the converter and resizer tools.
- Tag the folder with the Batch WD14 Tagger. Set the confidence slider (see our guide to thresholds) and look at a sample of results.
- Prune and edit. Remove tags you do not want the model to learn, add a trigger word to every picture, and fix mistakes in bulk. The tag pruning guide explains what to cut.
- Save the captions as one
.txtfile per picture, next to the pictures. - For sentence captions, run the folder through the Image Caption Generator and read every caption against its picture. A wrong detail in a training caption teaches the model the wrong thing.
The full order, with what to check at each step, is in our dataset workflow guide.
What to expect, and what not to
- These are suggestions, not facts. The WD14 taggers were trained on anime-style pictures, so photographs and unusual styles are outside what they know best. Every tag has a confidence score, which is why there is a slider.
- Captions can invent details. The caption generator and the alt text generator do not count well, cannot read text reliably and do not know people or places. Treat the result as a first draft.
- The batch tagger uses a lot of memory. Each background worker loads its own copy of the model: about 1.6 GB with 4 workers, and more on computers with many cores. Close other heavy tabs before a big run.
- Models download once. The first picture waits for the download; your browser keeps the files, so later visits are quicker.
Where the work happens
Tagging and captioning run in your browser tab, with the model files fetched once from Hugging Face, a public model host. Your picture files on disk are never changed: the results are text you copy or files you download. For what this site collects and keeps, see our privacy policy, and for the details of in-browser processing see our guide.
Frequently asked questions
What is the difference between the Anime Tagger and the Danbooru Tagger?
They use the same model. The Anime Tagger formats the tags for a prompt or caption. The Danbooru Tagger formats them for an upload to Danbooru, with post counts, current tag names and a rating tag.
Which tagger should I use for a LoRA dataset?
The Batch WD14 Tagger: it tags the whole folder, lets you prune and edit the tags in bulk, and saves one .txt caption file per picture. Use the Image Caption Generator as well if your model expects sentences.
How many pictures can I tag at once?
As many as your computer's memory allows: the batch tagger runs several workers, each with its own copy of the model. Split a very large folder if the tab struggles.
Are tags or captions better?
It depends on what will read them. Models trained on booru tags expect tags; models trained on sentences expect captions. See our guide on tags versus natural-language captions.
Can I edit the tags and captions?
Yes. In the batch tagger you can search, filter, add and remove tags across the whole set, and edit a single picture's tags. In the caption generator every caption is a text box you can change before saving.
Does the character tagger work on real photos of people?
No. It recognizes 2,132 popular anime and game characters, and is not made to identify real people.
Are my pictures uploaded to be tagged?
The tagging model runs in your browser tab. For what the site collects and keeps, see our privacy policy.