← Back to Blog · 2026.07.31
A Guide to AI Image Tagger Models: Understanding General, Character, and Rating Tags
The image tagger AI model that powers AnimeTag-AI was trained on millions of subculture illustrations from Danbooru. Given an image, it calculates, for each of several thousand tags, "the probability that this tag is present in the image." This post walks through the three categories you'll see in your results and how to interpret the Confidence Score.
1. General Tags — The Richest Category
General is the category that covers nearly every visual attribute you can observe in an image —
hair color, eye color, expression, pose, outfit, background, art style, and more. The vast
majority of tags fall under General.
1girl, blue_hair, school_uniform,
looking_at_viewer, and simple_background are all examples. With so
many tags, the raw results can look like a jumble of information — which is exactly why
AnimeTag-AI's "Category View" feature exists: to break these General tags back down into
appearance, outfit, pose, background, and more.
2. Character Tags — Character & Series Recognition
The Character category identifies a specific character or series, like hatsune_miku.
Recognition accuracy here depends heavily on how often a character appeared in the training data
— well-known characters are recognized reliably, while minor or original characters often aren't
recognized at all. An empty Character category isn't an error; it usually just means the model
hasn't seen enough examples of that character.
3. Rating Tags — Content Rating Classification
Rating is a special category with only four possible values: general,
sensitive, questionable, and explicit. It automatically
assesses how suggestive or explicit an image is, so you can filter or sort your dataset safely.
The four probabilities are normalized to sum to roughly 1, so whichever rating has the highest
value is the model's "best guess."
4. How Should You Read the Confidence Score?
Every tag comes back with a Confidence Score between 0 and 1. Think of it less as a strict "probability" and more as "how strongly the model believes this tag applies." A score of 0.9 or higher is close to certain; 0.35–0.6 is ambiguous but worth a look; anything under 0.2 is usually closer to noise. AnimeTag-AI's Threshold slider filters tags based on exactly this value — see our Threshold guide for recommended settings.
Try It Yourself
Nothing beats trying it on a real image. Upload an illustration to AnimeTag-AI to see the General and Character tags and Rating described above in action, and switch to "Category View" to watch those General tags automatically sort themselves into appearance, outfit, pose, background, and more.