Tagger Tool
The Tagger Tool applies structured, Danbooru-style tags to your images automatically. It runs a multi-label image tagger over a folder, predicts the tags each image most likely warrants, and writes the confident ones into Bitcrush Studio. This gives you a searchable, structured description of every image without tagging each one by hand.
These tags are the same tags used elsewhere in Studio: they can be filtered and sorted in Gallery View, and they feed the tag-style captions that Studio maintains through Background Sync. See the Captioning page for how tags become captions.
The tool offers two families of multi-label tagger. You select one from the Model dropdown, and the panel beneath it changes to show the controls relevant to that model.
The WD-Tagger v3 family, trained by SmilingWolf, is the default choice for anime, illustration, and Danbooru-style tagging, and generalises reasonably well to photographs. Each variant shares the same tag vocabulary and differs only in the underlying network, trading speed against accuracy:
| Model | Notes |
|---|---|
| WD-ViT-Large v3 | The default. A strong general-purpose choice. |
| WD-ViT v3 | Lighter than ViT-Large. |
| WD-SwinV2 v3 | SwinV2 backbone. |
| WD-ConvNeXt v3 | ConvNeXt backbone. |
| WD-EVA02-Large v3 | Highest accuracy, but the heaviest to run. |
Every WD-Tagger v3 model predicts three kinds of output: general tags, character tags, and a content rating. General and character tags are controlled independently, as described below.
JoyTag (fancyfeast/joytag) is a single-headed tagger covering roughly 5,000 Danbooru tags. It does not separate general and character tags, so it exposes a single threshold rather than two. It also generalises well to photographs.
Model files are downloaded automatically from Hugging Face the first time a model is used, so the first run of a new model includes a one-off download. Downloads require an internet connection; if Studio is running in offline mode, model downloads are blocked until you allow online features again.
The Tagger Tool works on the folder belonging to your active dataset and the folder currently selected in Gallery View. You do not enter a path manually.
Enable Recursive Scan to include every subfolder beneath the current folder as well. If you have an active search filter, the tool tags only the images that filter matches, rather than the whole folder.
A tagger produces a confidence score for every tag it knows. The threshold controls decide how confident the model must be before a tag is kept.
For the WD-Tagger v3 models you can set two thresholds independently:
-
General Tags: The confidence a general tag must reach to be kept. The default is
0.35. -
Character Tags: The confidence a character tag must reach to be kept. Character tags are held to a stricter default of
0.75, because a wrongly assigned character name is more damaging than a missing descriptive tag.
JoyTag exposes a single Tag Threshold, since it does not split general and character tags. Its default is 0.4.
Each threshold also has a Max control beneath it. Raising the maximum below 1.00 keeps only tags whose confidence falls inside the band between the lower and upper values, which is occasionally useful when you want to capture the model's less certain guesses without also re-adding the tags it is already sure about. Leave the maximum at 1.00 to keep every tag above the lower threshold.
Note: Lowering a threshold captures more tags but admits more mistakes; raising it keeps only the tags the model is most sure of. The right setting depends on the model and the kind of images you are tagging.
The Action panel controls how predicted tags are combined with any tags an image already has.
- Append Tags: Adds the newly predicted tags to the image's existing tags, keeping everything already present.
- Overwrite All: Replaces every written tag on the image with the new prediction. Because this can discard earlier tag work, Studio asks you to confirm before a run using this method.
- Smart Overwrite: Replaces only the tags this model is capable of producing, while preserving any tags that came from elsewhere. This lets you re-tag with fresh predictions without losing tags added by other tools or by hand.
Batch size sets how many images are sent to the model at once. Larger batches are faster but use more GPU memory. If a batch exhausts your VRAM, Studio automatically retries it in smaller pieces, so a batch size that is slightly too large will slow a run down rather than stop it.
Tags are written to the Bitcrush Studio database, which is the source of truth for both tags and captions. The Write to Database switch must be enabled for a run to store its results.
Studio also caches each model's raw predictions. If you run the same model over the same images again — for example to try a different threshold — it reuses the cached predictions instead of re-running the model, so repeated passes are far quicker.
Two optional panels let you tidy the predicted tags before they are written.
-
Tag Blocklist: A comma-separated list of tags to exclude from the output entirely. This is useful for stripping tags you never want, such as
simple backgroundor an over-eager1girl. -
Tag Switches: A set of from → to replacements. Any predicted tag matching the from value is rewritten to the to value before it is stored — for example, rewriting
1girltowoman. Both the blocklist and switches are applied consistently whether a run uses fresh predictions or cached ones.
Studio treats the database as the single record of an image's tags, and builds tag-style captions from it during Background Sync. After a tagging run writes its tags, Studio refreshes the captions for the affected images so the new tags flow through into their tag-style caption lines, provided caption syncing is enabled in your settings.
This means the Tagger Tool and the caption system stay in step: the tags you generate here become the structured, comma-separated caption that sits alongside your natural-language descriptions. See the Captioning page for more on caption styles and Background Sync.
Select a model, adjust the thresholds and options to suit your dataset, and select Start Tagging. The status panel reports progress and a running count of tags added, and a results list shows recently tagged images with the tags each received.
Tagging runs as a background job, so you can move to other parts of Studio while it works and return later; the run continues and the screen rebuilds its progress when you come back. Select Stop at any time to end a run early — the tags already written are kept, and because predictions are cached, resuming later picks up from where the run left off.