Pruning and blacklisting tags: what to keep and what to cut

A WD14 tagger at its default settings writes down everything it is reasonably sure it can see. That is exactly what you want from a tagger, but not everything it sees belongs in a training caption. Pruning is where you decide which of those claims help your goal and which work against it.

This guide covers what to cut, what to keep, and how to do both in ImgTagger. If you are new to captioning, start with how to caption a LoRA dataset, which explains the "caption what should vary" rule this page builds on.

Blacklist vs. removing tags afterwards

The ImgTagger tag editor gives you two ways to get rid of a tag, and they behave differently.

A few consequences are worth knowing before you rely on either one:

• Spelling is forgiving. The blacklist and Remove Tags treat simple_background, Simple Background and simple background as the same tag, so type whichever form you like.
• The blacklist leaves existing tags alone. Tags loaded from a .txt file, added by an earlier run or listed in Tags to Prepend are not filtered.
• Changing the blacklist doesn't undo a run. To apply a new blacklist, clear the tags and run Auto Tag again. ImgTagger keeps each image's raw scores in memory for every model it has run, so the second run skips the model and finishes quickly. Clearing also wipes hand edits, so settle the blacklist before you start editing.
• Re-running Auto Tag brings pruned tags back. It keeps what is there and adds any missing model tags. Anything you removed by hand returns unless it is blacklisted.
• The trash button is literal. If a batch contains both long hair and long_hair, the Filter panel lists them as two entries, and each trash button removes only its own spelling. Remove Tags clears both at once. Our guide to caption .txt files covers keeping one spelling style.

A good rhythm: blacklist what you know you will never want, auto-tag, then prune what you discover in the Filter panel.

Redundant parent and child tags

Danbooru uses tag implications: some specific tags automatically add a more general one. Every post tagged very_long_hair also gets long_hair. WD14 taggers learned from those posts, so they tend to output both tags of a pair together.

The post counts come from the tag list of ImgTagger's default model, and they show how much broader the parent is. Only about one in five long_hair posts was also very_long_hair, and only about a quarter of animal_ears posts were cat_ears. The parent stands for a whole family of looks, and the base model has seen it many more times than any one child.

Whether to keep both is a trade-off, not a rule:

• Keep both when you train on an anime model that learned from booru-style captions. That is the pairing it saw in training, and you can prompt with either tag later.
• Cut the parent when captions are getting long, or when the base model wasn't trained on booru tags. The specific tag carries the more useful information.

Don't use the blacklist for this. Blacklisting long hair removes it from every image, including the ones with long hair that isn't very long. Remove the parent only where the child is present. In the Filter panel, check very long hair in Any mode, click Select All to select the visible images, then remove long hair from the selected ones.

Cut tags that describe your concept

The most important cuts are the tags for whatever you are training. In a character LoRA, that means the features the character always has, such as hair color, eye color and a signature accessory. In a style LoRA, it means the tags for the look itself, such as traditional_media, sketch or limited_palette. Leave them in, and the model has somewhere other than your trigger word to put the concept.

To find them, sort the Filter panel by count, high to low. A tag on nearly every image is either part of your concept or a coincidence of your dataset. If it is part of the concept, cut it. If it is a coincidence (say, almost every image happens to be outdoors), keep it, so the LoRA doesn't learn "outdoors" as part of the concept.

Low-value and misleading tags

Some tags are accurate but unhelpful, or helpful only for some goals. The right call depends on what you want the model to learn.

The watermark row is the one people most often get backwards. Blacklisting watermark doesn't remove a watermark from the image. It only hides it from the trainer, which then has to explain the mark some other way, usually with your trigger word.

ImgTagger already holds a few of these to a stricter standard: monochrome, greyscale, sketch and blurry need 0.9 confidence, and the three background tags need 0.5. When they do appear, they are usually right. The confidence threshold guide explains these overrides.

Audit for one-off noise

Most tagger mistakes are rare tags on one or two images. They are hard to see in a single caption and easy to see across the batch:

1. Open the Filter panel and set Sort tags by to Count (low → high).
2. Click the first tag to show only the images that have it (Any mode), and look at them.
3. If the tag is wrong, click its trash button. For a tag on one image, that removes it from just that image.
4. If it is right, leave it. A rare but correct tag is exactly the kind of variable detail a caption should record. If you aren't sure what a tag means, the booru tag glossary explains the common ones.
5. Uncheck it and move on to the next tag.

The search box helps with families of tags. Search ears to see animal ears, cat ears and fox ears together, or background to see every background tag at once. Hand-typed tags benefit too: a typo shows up as a count-of-one entry right next to the correct spelling.

Sample blacklists to start from

These are starting points, not finished lists. Paste one into Never add these tags box in the Auto-tag panel, then adjust it after your first look at the results. They use spaces to match ImgTagger's output, but underscores work too.

Any project: no rating tags

general, sensitive, questionable, explicit

Character LoRA

For a character who always has grey hair, green eyes and a hairclip. Swap in your own character's fixed features, and leave outfit, pose and background tags alone.

general, sensitive, questionable, explicit, grey hair, green eyes, hairclip, hair ornament

Style LoRA

For a traditional ink-and-watercolor look. List only tags that describe the style on every image.

general, sensitive, questionable, explicit, traditional media, watercolor (medium), limited palette

Images already cleaned of marks

Only once signatures and watermarks are really gone from the pixels. This stops stray guesses on clean images.

signature, watermark, artist name, dated, twitter username