Stable Diffusion Prompts and Negative Prompts: A Practical Guide
Order keywords, write targeted negative prompts, use weights and pick the right settings for SD 1.5, SDXL and newer checkpoints.
Stable Diffusion gives you more control than any hosted image generator โ and more ways to get it wrong. Good results come from three things working together: a well-ordered positive prompt, a targeted negative prompt and sensible generation settings. This guide covers all three for SD 1.5, SDXL and newer checkpoints.
Anatomy of a positive prompt
Stable Diffusion's text encoder pays most attention to the start of the prompt. Order keywords from most to least important:
- Subject โ what the image is about: a red fox sitting in snow
- Medium โ photograph, digital painting, watercolor, 3D render
- Style โ cinematic, art nouveau, studio ghibli style, film noir
- Artist or influence (optional) โ use sparingly and responsibly
- Details โ clothing, props, expressions, textures
- Color and lighting โ golden hour, rim light, teal and orange palette
- Quality tags โ highly detailed, sharp focus, 8k
Example:
a red fox sitting in fresh snow, wildlife photograph, telephoto lens,
early morning mist, soft backlight, frost on fur, highly detailed, sharp focus
SD 1.5 vs SDXL and newer models
- SD 1.5 checkpoints respond well to comma-separated tags and quality boosters such as masterpiece, best quality.
- SDXL and later models understand natural sentences much better. You can write "A red fox sits in fresh snow while morning mist drifts behind it" and drop most quality boosters.
- Anime checkpoints trained on booru tags (e.g. 1girl, solo, long hair) expect that tag style โ check the model card.
Negative prompts
The negative prompt lists things the sampler should steer away from. It is one of Stable Diffusion's most powerful features and one of the most abused.
A sensible general-purpose negative prompt
lowres, blurry, jpeg artifacts, watermark, signature, text, cropped,
out of frame, worst quality, low quality
Add targeted terms for the problem you actually see
| Problem | Add to negative prompt |
|---|---|
| Bad hands | bad hands, extra fingers, fused fingers, missing fingers |
| Distorted anatomy | bad anatomy, extra limbs, deformed, disfigured |
| Unwanted style | cartoon, anime, 3d render (for photos) or photo, realistic (for illustration) |
| Oversaturated look | oversaturated, overexposed |
| Unwanted elements | the specific object, e.g. people, cars, trees |
Keep negative prompts short and specific. A 200-word negative prompt copied from the internet can suppress things you actually want and make images look flat.
Negative embeddings
Textual inversions such as EasyNegative or bad-hands-5 pack many negative concepts into one token. They must be installed in your UI and usually only work with the base model family they were trained on (e.g. SD 1.5). Type the name exactly โ a changed hyphen or space means it will not load.
Prompt weighting
In AUTOMATIC1111, Forge and ComfyUI you can change how strongly a word is applied:
(word)increases weight by about 1.1ร((word))increases it about 1.21ร(word:1.4)sets an exact weight[word]or(word:0.8)reduces weight
Stay between roughly 0.5 and 1.5. Higher weights tend to produce artifacts or burn the concept into every part of the image.
Generation settings that matter
| Setting | What it controls | Good starting point |
|---|---|---|
| Sampling steps | Refinement passes | 20โ30 (more rarely helps) |
| CFG scale | How strictly the prompt is followed | 5โ7 for SDXL, 7 for SD 1.5 |
| Sampler | Denoising algorithm | DPM++ 2M Karras or Euler a |
| Resolution | Output size | 512ร512 for SD 1.5, 1024ร1024 for SDXL |
| Seed | Starting noise | Fix it while tweaking the prompt |
Generating at the model's native resolution and then upscaling gives much better results than generating a large image directly.
A workflow for consistent results
- Write a short prompt with just the subject and medium. Generate a batch of 4.
- Pick the best result and fix its seed.
- Add style, lighting and details one group at a time, checking each change.
- Add negative terms only for problems you see.
- Upscale and use inpainting to fix small areas such as hands or faces.
Examples
Portrait photo
Prompt: portrait of an elderly woman laughing, natural window light,
85mm lens, shallow depth of field, film grain, warm tones
Negative prompt: cartoon, 3d render, blurry, deformed, watermark, text
Fantasy illustration
Prompt: a floating castle above the clouds at sunset, digital painting,
epic fantasy, volumetric light, intricate details, vibrant colors
Negative prompt: photo, lowres, blurry, text, watermark
Product shot
Prompt: a glass perfume bottle on wet black stone, water droplets,
studio lighting, reflections, product photography, sharp focus
Negative prompt: people, text, logo, lowres, overexposed
Generate a prompt and negative prompt together
The Stable Diffusion Prompt Generator returns both a positive and a negative prompt from a one-line idea. Copy them into your UI, fix a seed and refine from there.
Try the Stable Diffusion Prompt Generator
Positive and negative prompts for SD and SDXL. Free, no sign-up.