CraftPrompts

Stable Diffusion prompt generator

Weighted positive prompt, a matching negative prompt and the settings to go with them, sized correctly for the model you run.

Style or medium
Lighting
Camera and composition
Mood
Colour

Negative prompt

Include sets (see the negative prompts guide)

Settings

Stable Diffusion prompt
subjectdetailsstylelightcameranegative
Fill in a few fields and your prompt is set here, line by line.

Negative prompt

lowres, blurry, jpeg artifacts, worst quality, low quality, watermark, signature, text, cropped, out of frame
0 words · 0 characters

Settings: Model: SDXL · Size: 1024 × 1024 · CFG: 6 · Steps: 30 · Sampler: DPM++ 2M Karras

Expand into tags with AI

Turns your idea into ordered SD tags plus a short negative prompt.

Prompting Stable Diffusion

Stable Diffusion reads prompts through a CLIP text encoder (SDXL uses two), which treats your text more like a bag of weighted tags than a sentence. That is why comma-separated phrases work so well and why order matters: the first phrases pull hardest.

The negative prompt is not a list of “don’ts” the model understands. It is a second prompt the sampler steers away from at every step, using the same CFG mechanism. The negative prompts guide has an interactive explainer of exactly how that works, plus copy-ready sets.

Questions

What does (word:1.2) mean in a Stable Diffusion prompt?

It is attention weighting, supported by AUTOMATIC1111, Forge, ComfyUI and most SD front ends. 1.0 is normal; 1.2 gives that term about 20% more influence; values below 1 weaken it. Keep weights between roughly 0.5 and 1.5, beyond that images tend to break.

Do I need a negative prompt?

For SD 1.5 and SDXL, a short negative prompt noticeably reduces blur, artefacts and anatomy errors. Newer models (SD 3.5, Flux) rely on it less. Keep it short and relevant; a 200-word negative prompt dilutes itself.

What CFG scale should I use?

CFG (classifier-free guidance) sets how strictly the image follows the prompt. For SDXL, 5–8 is a good range; SD 1.5 often likes 7–9. Too high and colours burn and details get crunchy; too low and the prompt is ignored.

Why does the resolution change with the model?

Each model was trained at a native size: 512×512 for SD 1.5, about 1024×1024 for SDXL and SD 3.5. Generating far from that pixel count causes duplicated subjects or mush, so the builder keeps the total pixel count near native and rounds to multiples of 64.

Does word order matter?

Yes. Earlier tokens tend to have more influence, and the CLIP text encoder only reads about 75 tokens per chunk, so put the subject and the most important style words first.

More tools in the studio