Why use tags instead of natural language for anime models?

Models like NovelAI, Counterfeit, and Anything were trained on danbooru and e621 images labeled with discrete tags. They respond far more reliably to that exact vocabulary — for example 1girl, blue_eyes, smile — than to descriptive English sentences.

What do the quality tags like masterpiece and best quality do?

These were common high-rating labels in the training data, so prompting them biases the model toward its more polished, detailed outputs. They are conventions rather than guarantees, but they reliably nudge anime models toward cleaner results.

What does the rating tag control?

Booru datasets tag images with safety ratings — typically safe, questionable, or explicit. Including a rating tag steers the model's content level. This builder defaults to and emphasizes the safe rating for general-audience output.

Do tags use spaces or underscores?

Booru tags are stored with underscores (long_hair) but most modern WebUIs accept either spaces or underscores and tokenize them similarly. This builder outputs spaces for readability, which the common samplers handle correctly.

Booru-Style Image Tagging Prompt Builder

Booru-style image tagging prompt builder

Anime and illustration Stable Diffusion models — NovelAI Diffusion, Counterfeit, Anything v5, and their many merges — were trained on imageboard datasets where every picture is labeled with discrete tags rather than sentences. To get the best out of them you speak their language: 1girl, long_hair, smile, masterpiece. This builder assembles a clean, correctly ordered tag prompt from plain inputs.

How it works

The tool follows the conventional ordering these models expect: quality tags first, then the subject count and character traits, then expression, then setting, and finally a safety rating. Each field maps to real danbooru/e621 vocabulary, so the output is something the model has actually seen during training. Quality boosters like masterpiece and best quality bias the model toward its cleaner outputs, while the rating tag steers the content level — this builder centers on the safe rating for general-audience work. The result is a comma-separated string you paste straight into the positive prompt box.

Tips and examples

Lead with quality tags. masterpiece, best quality at the front sets the bar before any subject detail lands.
Be specific with traits. Concrete booru tags (blue_eyes, twin_braids, school_uniform) outperform vague descriptions every time.
Keep the count tag. 1girl, 2boys, or solo strongly shape composition — always include one.
Pair with a negative prompt. These models lean on negatives heavily; pair this positive prompt with common quality negatives like lowres, bad anatomy, worst quality in the negative field.