How to write prompts for Stable Diffusion?
Write Stable Diffusion prompts as a comma-separated list of specifics: subject, then details, style, lighting, and quality words, ordered by importance because earlier words carry more weight. Add a negative prompt for what to avoid. Example: 'a red fox in a snowy forest, golden hour, photorealistic, sharp focus, 50mm lens'. Iterate one change at a time.
Why — the first-principles explanation
A prompt is a set of instructions the model weighs by relevance and position. Stable Diffusion doesn't read sentences like a human; it reads a bag of concepts and tries to satisfy all of them at once. So clear, concrete nouns and adjectives work far better than vague poetry.
Word order matters. Terms near the front generally get more emphasis. Lead with your main subject, then add supporting details. If you bury 'cat' at the end of a long description, the model may drown it in other concepts.
Specificity beats length. 'A dog' gives the model total freedom, so results vary wildly. 'A golden retriever puppy sitting on a wooden porch, soft morning light' constrains it toward what you actually want. Good prompts usually name the subject, its attributes, the setting, the lighting, the art style, and a couple of quality cues.
Two levers fine-tune this. Emphasis syntax like (word:1.3) increases a term's weight, and (word:0.7) decreases it, letting you dial concepts up or down. The negative prompt removes unwanted traits such as 'blurry, deformed, extra limbs, watermark'. Since results are partly random, the real technique is iteration: change one element, regenerate, and compare, rather than rewriting everything at once.
An example that makes it click
Think of ordering a custom cake. Saying 'a cake' gets you whatever the baker feels like. Saying 'a two-layer chocolate cake, vanilla frosting, ten red roses on top, gold script that says Happy Birthday' gets you exactly what you pictured.
Stable Diffusion is the baker. The more concrete details you list, and the more you put the important stuff first, the closer the result. And if the baker keeps adding sprinkles you hate, your negative prompt is the note that says 'no sprinkles'.
How to do it
- Start with your main subject in a few words, placed at the front of the prompt.
- Add descriptive details: colors, clothing, pose, age, materials.
- Add the setting and lighting, such as 'in a misty forest, golden hour'.
- Add style and quality cues, such as 'photorealistic, sharp focus, 50mm lens' or 'watercolor illustration'.
- Separate terms with commas and order them by importance.
- Write a negative prompt for defects to avoid, like 'blurry, deformed hands, extra fingers, watermark'.
- Use emphasis syntax like (word:1.3) to strengthen a term, then change one element at a time and regenerate.
Key facts
- Stable Diffusion weights prompt terms partly by position, so earlier words tend to carry more influence.
- Prompts are typically written as comma-separated keywords rather than full sentences.
- Emphasis syntax (word:1.3) increases a term's weight and (word:0.7) decreases it in most interfaces.
- A negative prompt lists traits to suppress, commonly 'blurry, deformed, extra fingers, watermark'.
- Because output is partly random, iterating one change at a time is the standard refinement method.
The open-source image model you can run on your own hardware.
Affiliate link — we may earn a commission at no cost to you.▶ The 60-second explainer (script)
Here's how to write great Stable Diffusion prompts. The model doesn't read sentences like a person. It reads a list of concepts and tries to satisfy all of them, so be specific and use commas. Start with your main subject at the very front, because earlier words carry more weight. Then stack on details: colors, clothing, pose. Add the setting and lighting, like 'misty forest, golden hour'. Then add style and quality words, like 'photorealistic, sharp focus, 50 millimeter lens', or 'watercolor illustration'. So instead of 'a fox', write 'a red fox in a snowy forest, golden hour, photorealistic, sharp focus'. Next, add a negative prompt for what you don't want, like 'blurry, deformed hands, watermark'. Want a concept stronger? Wrap it like this: open parenthesis, word, colon, one point three. Finally, don't rewrite everything at once. Change one detail, generate again, and compare. That's how pros dial in the perfect image.
What authoritative sources say
People also ask
Should prompts be sentences or keywords?
Keywords separated by commas usually work best. Full sentences can work but keyword lists give you clearer control over each element.
How long should a prompt be?
Long enough to be specific, but not padded. Most interfaces weight the first 75 tokens most heavily, so front-load the important terms.
How do I make one word stronger?
Use weighting syntax like (word:1.3) to emphasize it, or (word:0.7) to reduce it, in interfaces that support it.
Why do I get a different image each time?
The seed changes randomly by default. Lock the seed to keep the same base image while you refine the prompt.