How to Write Stable Diffusion Prompts That Work

To write stable diffusion prompts that give you the image you mean, start with a clear subject, add style and detail cues, then refine with weighting and negative prompts while testing each change. This approach turns a vague idea into a reliable recipe the model can follow. Below you’ll find the exact steps, common pitfalls to avoid, and a worked example you can adapt.

Core Prompt Building Blocks

Every usable prompt contains three layers: subject, style, and details. The subject names what you want to see, a person, an object, a scene. The style tells the model how to render it, such as “photorealistic”, “watercolor”, or “low poly”. Details add texture, lighting, mood, or specific features like “wet cobblestones” or “soft morning light”. Keep each layer separated by commas so the model can parse them easily.
A simple prompt might read: “a wooden cabin, photorealistic, pine forest background, golden hour light, detailed bark texture”. Notice how each piece adds information without overlapping. If you leave out the subject, the model has nothing to anchor to. If you skip style, the output may default to an unintended look. Missing details often lead to generic results that lack the feel you want.
When you’re just starting, write each layer on its own line, then combine them. This habit makes it easier to spot missing pieces later.

Using Weighting to Emphasize Parts

Sometimes you need the model to pay extra attention to one element, like the face in a portrait or the texture of a fabric. Most front-ends for stable diffusion let you weight terms with parentheses and a number. A weight of 1.0 is normal; values above 1.0 increase emphasis, values below 1.0 decrease it.
For example, to make sure the eyes are sharp in a portrait you could write: “(sharp eyes:1.3), a woman’s face, soft lighting, detailed skin”. The model will try harder to render sharp eyes while still considering the other cues.
If you want to tone down something that tends to dominate, use a weight below 1.0. A background that keeps pulling focus can be eased: “a bustling market, (muted colors:0.8), detailed stalls, crowd blur”.
When you apply weights, change only one value at a time and note the result. This makes it clear what each adjustment did.

Adding Negative Prompts to Remove Unwanted Elements

Even a well-crafted positive prompt can produce things you don’t want, extra limbs, blurry edges, or a style clash. Negative prompting tells the model what to suppress. Most interfaces have a separate field for negative terms, or you can add them after a double hyphen or the word “no”.
Common negatives include: blurry, low resolution, extra fingers, deformed face, jpeg artifacts, oversaturated, cartoon, watermark.
Start with a short list, test, then add more only if the problem persists. Overloading the negative side can make the image look flat or brittle, so keep it focused on the specific flaw you see.
A practical example: you keep getting a plastic look on skin. Add “plastic skin, airbrushed, overly smooth” to the negative field. Regenerate and see if the skin gains a more natural texture.

Structuring Prompts for Consistency

If you plan to create a series of images that share a look, say, a set of product mockups, write a base prompt that contains the fixed elements, then add variable parts for each item.
Base: “a ceramic mug, studio lighting, soft shadow, pastel glaze, detailed handle”.
Variable: for a blue version add “blue glaze”; for a red version add “red glaze with white stripe”.
Saving the base as a template lets you swap only the changing piece, which reduces errors and keeps the style uniform across the set.

Saving and Organizing Your Prompts

Keep a simple log of every prompt you try, the model version, any parameters you changed (steps, cfg, sampler), and the seed if you used one. A spreadsheet or a notes app works fine.
Record the date, a short description of the goal, the full prompt, negative prompts, settings, and a thumbnail of the result. When you later search for “portrait with warm light” you’ll find the exact combination that gave you the best skin tone.
Reviewing this log weekly helps you spot patterns, like which weights consistently improve sharpness, and prevents you from repeating the same missteps.

Common Mistakes and How to Avoid Them

One frequent error is stuffing too many unrelated adjectives into a single prompt. The model then struggles to prioritize, and the output feels muddled. Keep each prompt focused on a handful of clear ideas.
Another mistake is ignoring the impact of sampling steps and cfg scale. Low steps can leave the image noisy; a very high cfg can introduce artifacts while trying to obey the prompt too strictly. Start with moderate values, around 25-30 steps and cfg 7-9, and adjust only after you’ve checked the prompt wording.
A third pitfall is using vague terms like “nice” or “beautiful”. Replace them with concrete descriptors: “nice” becomes “soft pastel palette”, “beautiful” becomes “symmetrical composition, gentle rim light”.
Finally, many people forget to test one change at a time. When you tweak weighting, negative prompts, and steps all together, you can’t tell which tweak helped. Change one variable, run a test, note the outcome, then move to the next.

Worked Example: From Idea to Final Prompt

Imagine you want an illustration of a reading nook for a blog post.

1. Start with subject: “a cozy reading nook”.
2. Add style: “watercolor illustration”.
3. Add details: “window seat, stacked books, warm lamp, soft blanket, afternoon light through leaves”.
   Combine: “a cozy reading nook, watercolor illustration, window seat, stacked books, warm lamp, soft blanket, afternoon light through leaves”.
4. Test: The first result looks pleasant but the light is too flat.
5. Add weighting to light: “(afternoon light through leaves:1.2), a cozy reading nook, watercolor illustration, window seat, stacked books, warm lamp, soft blanket”.
6. Check for flaws: The blanket looks plasticky.
7. Add negative prompt: “plastic texture, overly smooth”.
8. Final prompt: “(afternoon light through leaves:1.2), a cozy reading nook, watercolor illustration, window seat, stacked books, warm lamp, soft blanket --no plastic texture, overly smooth”.
9. Run with 30 steps, cfg 8 and save the prompt with the settings that gave the soft, inviting look you wanted.

Common questions

How do I know which weighting values to start with?

Begin with 1.0 for every term. If a particular element feels weak, raise its weight in small steps, try 1.1, then 1.2, while keeping the others at 1.0. If something is too strong, lower it to 0.9 or 0.8. Test each adjustment individually to see the effect.

Can I use the same prompt across different stable diffusion front-ends?

The core wording will work, but each front-end may interpret weighting and negative prompt syntax differently. Check the documentation for the specific interface you’re using, then copy the subject-style-details part and adapt the weighting or negative fields to match its format.

What should I do if my prompt keeps producing the same unwanted artifact?

First, add a clear negative term for that artifact, for example, “jpeg artifacts” or “extra limbs”. If the artifact persists, lower the cfg scale a bit and increase the sampling steps by 5-10. Sometimes the model needs more steps to smooth out the glitch. Keep a note of what combination finally removed it so you can reuse the fix.

Is it worth saving prompts that didn’t work out?

Yes. A failed prompt is still useful data. Record what you tried, what went wrong, and any settings you changed. Later you can compare it with successful prompts to spot patterns, like which negative terms consistently fix a certain flaw, turning mistakes into a reference library.