Tag Prompts vs Natural Language Prompts (Which to Use?)

Should you write prompts as tags or natural language?

Match how your tool was trained. Anime-focused Stable Diffusion models and NovelAI prefer tags; Midjourney, FLUX and ChatGPT prefer sentences.

An image model understands prompts best when they resemble the image descriptions it was trained on. Many anime models were trained on images described with Danbooru tags, so they are good with tags. Models trained on images described in sentences are better with sentences.

What is a tag-based prompt?

Each word stands for one thing, and you add or remove tags to decide the parts of the picture.

In tags, one word (or short phrase) means one thing: twintails is twin tails, smile is a smile, and each tag is a part of the picture. Adding or removing a tag changes only that part, which makes fine adjustments easy.

• One tag, one meaning. Add it and it appears; remove it and it goes
• Tags written earlier work more strongly (depending on the model)
• Many tools let you change strength with weights such as (tag:1.2)
• Relationships such as "who is holding what" are hard to express

1girl, solo, long hair, black hair, red eyes, hair ribbon, school uniform, smile, holding book, library, sunlight, upper body

What is a natural language prompt?

You describe the whole scene in sentences, so the links between who is doing what and where come across well.

A natural language prompt describes the whole scene in sentences. Something like "a girl with a red ribbon smiles while holding a book by the window of the school library" carries who, what and where together. On the other hand, strengthening just one word precisely can be harder than with tags.

• The links within the scene (who is doing what) come across well
• Positions (on the left, behind) and held items come across well
• Negatives ("do not ...") often do not work well
• Put the important things in the first half

A girl with long black hair, red eyes and a red ribbon smiles while holding a book by the window of a quiet school library, sunlight falling across the shelves, upper body, anime style illustration.

How do you rewrite tags as natural language?

Make the number of people the subject, join the look with "with", clothing with "wearing", and turn actions into verbs.

Steps: tags to natural language

1. Turn the number-of-people tag (1girl and so on) into the subject (A girl)
2. Join the appearance tags (hair, eyes) with "with"
3. Join the clothing tags with "wearing"
4. Turn expressions, actions and held items into verb phrases (smiles, holds a book)
5. Join place and light with in, at or with
6. Drop the quality tags, or keep just one art style word at the end

Illustration usage rules

Steps to rewrite natural language as tags

Steps: natural language to tags

1. Decide the number-of-people tags (1girl, solo) from the subject of the sentence
2. Convert the look and outfit to Danbooru tag names
3. Split actions into tags such as smile and holding something
4. Turn place, light and composition into tags
5. Look up any words whose tag names you do not know in a tag list
6. Add quality tags at the end (following your model's conventions)

For how to look up Danbooru tags, see How to Look Up Danbooru Tags (Finding the Correct Tag Name) →

Tags or sentences not working? (Fixes)

The usual cause is a mismatch between the tool's prompt type and how the prompt is written.

Combining both when a model reads either

Some newer models can read both tags and sentences. With those, use tags for the visual parts and add only the relationships or actions that tags struggle with as a short sentence. That gets you the best of both.

1girl, solo, long hair, black hair, red eyes, hair ribbon, school uniform, library, sunlight, upper body, She is smiling while holding an open book close to her chest.

For the basics of tag prompts, see Prompt Writing Basics: Tag Order, Emphasis, and Tips →

For tips on natural language prompts in FLUX, see How to Write FLUX Prompts (in Natural English) →

For NovelAI, which mixes tags and short phrases, see How to Write NovelAI Prompts (Emphasis, Quality Tags, UC) →