How Stable Diffusion reads a prompt
Stable Diffusion is an open model family you can run locally or through many hosted interfaces such as AUTOMATIC1111, ComfyUI, Forge or online services. Because it is open, prompt behavior depends heavily on the checkpoint you load. Many community checkpoints respond well to comma-separated descriptive phrases, while newer base models like SDXL and SD 3.x understand fuller natural-language sentences. Start with the subject, then add style, lighting and camera terms.
Stable Diffusion gives you more direct control than most tools. You can raise or lower the influence of individual words with weighting syntax, add a separate negative prompt, and adjust CFG scale, sampling steps, sampler and seed. Extensions like LoRA and ControlNet add styles and pose control. Exact syntax varies between interfaces, so the examples here follow the widely used AUTOMATIC1111 conventions; check your tool's documentation for differences.
The prompt formula
[subject], [key details], [style/medium], [lighting], [camera/lens], [quality terms] | negative: [flaws] | CFG, steps, sampler, seed
Example
portrait of a red-haired woman in a green wool coat, (freckles:1.2), autumn park background, film photography, soft overcast light, 85mm, shallow depth of field, detailed skin texture | negative: blurry, deformed, extra fingers, watermark | CFG 6, 30 steps, DPM++ 2M Karras, seed 1001
Stable Diffusion parameters and controls
| Control | Syntax | What it does |
|---|---|---|
| Weighting | (word:1.2) or (word) | Increases a term's influence. Values around 1.1-1.4 are typical; much higher can distort the image. Syntax shown is AUTOMATIC1111-style.e.g. (golden hour:1.3) |
| De-emphasis | [word] or (word:0.8) | Reduces a term's influence in interfaces that support it.e.g. (background people:0.6) |
| Negative prompt | Separate negative prompt field | Lists concepts to steer away from. One of the most effective tools in Stable Diffusion.e.g. lowres, blurry, extra fingers, watermark, text |
| CFG scale | CFG scale slider | Controls how strongly the image follows the prompt. Around 5-8 is common; too high looks harsh and oversaturated.e.g. CFG 7 |
| Sampling steps | Steps slider | Number of denoising steps. 20-40 is typical; gains diminish beyond that for most samplers.e.g. 30 steps |
| Sampler | Sampler dropdown | The algorithm used to denoise. Different samplers trade speed for detail and style.e.g. DPM++ 2M Karras or Euler a |
| Seed | Seed field (-1 for random) | Fixes randomness so you can reproduce results and test single changes.e.g. seed 424242 |
| LoRA | <lora:name:0.8> | Loads a small add-on model for a style, character or concept at a chosen strength. AUTOMATIC1111-style syntax; ComfyUI uses nodes.e.g. <lora:watercolor_style:0.7> |
| Denoising strength | Denoising strength slider (img2img) | In image-to-image, sets how much the original image changes. Low keeps structure; high follows the prompt.e.g. 0.45 to restyle a photo |
Do this
- Match prompt style to the checkpoint: tag lists for many community models, full sentences for SDXL and SD 3.x bases.
- Use weighting sparingly on one or two key terms rather than weighting everything.
- Keep a reusable negative prompt per checkpoint and adjust it when specific flaws appear.
- Start at CFG 6-7 and 25-30 steps, then tune one variable at a time with a fixed seed.
- Use ControlNet for poses or layouts instead of trying to describe exact positions in words.
- Generate at the model's native resolution, then upscale, to avoid duplicated subjects.
Avoid this
- Using weights like (word:2.0) that break the image instead of emphasizing the term.
- Leaving CFG very high, producing burnt colors and over-sharpened edges.
- Generating far above native resolution and getting repeated heads or bodies.
- Copying negative prompts with embeddings or LoRAs you do not have installed.
Stable Diffusion prompt examples
Photoreal portrait
close-up portrait of an elderly man with a gray beard, (weathered skin:1.2), knitted sweater, warm window light, 85mm lens, f/1.8, film grain, natural colors, highly detailed | negative: blurry, plastic skin, deformed, extra fingers, watermark
Anime key visual
anime key visual, girl with short blue hair on a rooftop at sunset, (wind-blown hair:1.1), city skyline, cel shading, vibrant colors, detailed background | negative: lowres, bad anatomy, extra limbs, text, signature
Architecture
modern concrete and glass house on a hillside, cantilevered terrace, pine forest, dramatic dusk sky, architectural photography, wide-angle 24mm, warm interior lights | negative: distorted lines, people, cars, blurry, oversaturated
Fantasy art
ancient dragon sleeping on a pile of gold coins in a vast cave, (volumetric light:1.2) through a crack in the ceiling, epic fantasy painting, highly detailed scales | negative: lowres, cartoon, blurry, cropped, watermark
Stable Diffusion prompt FAQ
What does (word:1.2) mean in Stable Diffusion?
It is weighting syntax used by AUTOMATIC1111 and many compatible interfaces. The number multiplies the term's influence, so 1.2 means roughly twenty percent more emphasis and 0.8 means less. Plain parentheses add a small boost. Keep weights between about 0.6 and 1.4; extreme values tend to distort the image rather than strengthen the concept.
What is a good CFG scale?
For most Stable Diffusion checkpoints, a CFG scale between 5 and 8 balances prompt adherence and natural-looking images. Lower values give the model more freedom and softer results; higher values follow the prompt more strictly but can look oversaturated or harsh. Some distilled or turbo models are designed for very low CFG, so check the model card.
Which sampler should I use?
DPM++ 2M Karras and Euler a are popular starting points: the first is reliable and detailed, the second is fast with a slightly softer, more varied look. Differences between modern samplers are often subtle at 25-30 steps. Pick one, keep it fixed while you tune your prompt, and only compare samplers once the prompt is working.
Do I always need a negative prompt?
It is not required, but it is one of Stable Diffusion's most useful controls. A short negative prompt removing blur, watermarks, extra fingers and low resolution helps most checkpoints. Avoid enormous copied lists, which can flatten the image. Some newer base models need fewer negatives, so test with and without to see the difference.