PROMPT GUIDE
What Is the 75 Token Limit in Prompts? (Length and Chunks)
Have you noticed a number like "75/75" or "150" next to the prompt box? That is the length of your prompt counted in tokens. This guide explains what tokens are and how to trim and order a long prompt so it still works.
Last updated: approx. 3 min read
The short answer
Stable Diffusion prompts are counted in tokens, not letters, and read in chunks of 75 tokens. Most interfaces let you go past 75, but the more chunks there are, the weaker the tags near the end become, so put important tags first and cut tags that repeat the same meaning.
What is the 75 token limit in prompts?
It is the size of one chunk the AI reads at a time; in Stable Diffusion models one chunk is 75 tokens.
A token is a small unit the AI splits text into. A short English word is usually one token, while long words or symbols can split into two or three. Commas count as tokens too. Many interfaces read 75 tokens as one chunk and, past that, read the next chunk and join them (the exact behavior differs by interface and version).
What happens when you go over 75 tokens?
Most interfaces let you keep writing, but each tag tends to lose strength as the number of chunks grows.
Anything past 75 tokens is read as a following chunk and joined on. You can still write it, but the more tags there are, the more the AI attention is spread out. When background or prop tags near the end are in the prompt but do not appear, this is often why. Some services stop input at a limit or ignore the end, so check the documentation of the service you use.
How do you trim a long prompt?
Cut overlapping tags first, then tags that have no effect, then stacked quality tags; this keeps the image from breaking.
Steps to shorten a long prompt
- Check the current token count (the number by the prompt box, or a token counting tool)
- Merge tags with the same meaning into one (smile and happy, long hair and very long hair, and so on)
- Remove tags one at a time and find the ones whose removal does not change the image
- Check that you are not stacking quality tags. Two or three are often enough
- Reorder so important tags (person, hair, outfit, framing) come first and background and props come later
How should you place the chunk boundaries?
Group tags by meaning, such as person, outfit and background, so they mix less even across a boundary.
When color tags straddle a chunk boundary, colors tend to bleed, for example hair color moving onto clothes. Order tags by meaning, such as character look, then outfit, then framing, then background, so a group is not cut in the middle. For inserting your own boundary with BREAK, see the weight syntax guide.
Rinka's tips
- Start with about 20-30 tags and add only what is missing, so the prompt does not grow too long
- Token counting differs slightly by model and interface. Treat the number as a guide
- Other services such as NovelAI and Midjourney also have length limits and different behavior, so check each one documentation
For how to use BREAK and the (tag:1.2) weight syntax: The Complete Guide to Prompt Weight Syntax and BREAK →
For the basics of tag order: Prompt Writing Basics: Tag Order, Emphasis, and Tips →
For using quality tags without overdoing them: How to Use Quality Tags (masterpiece / best quality) →
Copy-and-paste templates (short examples)
The first is a short prompt grouped by meaning; the second includes a background but stays a reasonable length.
1girl, solo, long hair, red eyes, smile, white blouse, navy skirt, upper body, looking at viewer, classroom, window, sunlight, masterpiece, best quality1girl, solo, short hair, blue eyes, hoodie, jeans, full body, standing, park, autumn leaves, soft lighting, best qualityFAQ
Frequently Asked Questions
Is it OK to go over 75 tokens?
Most interfaces let you keep writing, but each tag tends to lose strength as chunks increase. Put important tags first and cut tags with overlapping meanings.
How are tokens counted?
A short English word is usually one token, while long words and symbols can split into two or three. Commas count too. Check the exact number in the prompt box display or with a token counting tool.
Does writing in a non-English language make the prompt longer?
Languages such as Japanese often use one or more tokens per character, so they get longer than English. On tag-based models, English tags are shorter and work better.
What length is about right?
Start with about 20-30 tags and add only what is missing. If adding tags changes nothing, reviewing the order often helps more than length.
RINKA WORLD
Play in Rinka World
MORE
