Stable Diffusion prompts you can copy and paste

How to prompt Stable Diffusion

Stable Diffusion is an open image model family from Stability AI. You can run it free on your own computer (with a good graphics card) through apps such as ComfyUI or Forge, or use it on many websites. Community checkpoints tune it for styles such as photo, anime or illustration.

Best for

Free, unlimited images on your own computer · Anime and illustration styles via community checkpoints · Fine control with negative prompts and settings · Product and interior scenes you edit afterwards

Prompt formula

Subject + setting + light + camera or art style + quality words, then a negative prompt and settings

Tips that make the biggest difference

• Use a negative prompt. List what you do not want, such as "extra fingers, blurry, watermark, text". It removes common problems without making the main prompt longer.
• Match the size to the model. Newer models are trained at about 1 megapixel, for example 1024x1024 or 1344x768. Very small or very large sizes cause repeated or stretched subjects.
• Keep CFG moderate. A CFG of about 5 to 7 follows the prompt without burning the colours. Very high values make skin waxy and colours harsh.
• Pick the right checkpoint. A photo checkpoint and an anime checkpoint give very different results from the same words. Choose the checkpoint first, then write the prompt.
• Put the most important words first. Words near the start usually carry more weight, especially with short, tag-style prompts.

Weak vs strong prompt

a beautiful woman, best quality, masterpiece

Photo of a woman in her 30s with short curly hair in a sunlit cafe, looking off camera, soft window light, 85mm lens, shallow depth of field, natural skin texture, warm muted colours.
Negative prompt: cartoon, 3d render, plastic skin, extra fingers, watermark, text
Settings: 896x1152, 30 steps, CFG 5

Common mistakes

• No negative prompt, so hands and text come out broken
• Asking for readable words or logos
• Odd image sizes the model was not trained on
• CFG turned all the way up

Guide checked against the current Stable Diffusion release in September 2026.

Controls Stable Diffusion gives you

• A negative prompt box. A separate field for what to keep out. Chat apps have nothing like it, and it is where much of the quality comes from.
• The seed. Keep the same seed and change one word to see exactly what that word does. Change the seed to get a new composition.
• Checkpoints and LoRAs. A checkpoint is the base model; a LoRA is a small add-on for a style or a character. Both change the result more than any wording.
• Inpainting and img2img. Paint over one area to redo only that part, or start from your own picture. A low denoising strength keeps the original layout; a high one changes more.

If Stable Diffusion gets it wrong

• Hands or faces are broken. Add them to the negative prompt, generate at the model's own size, then repaint the small area with inpainting.
• Two heads or repeated subjects. The image size is too far from what the model was trained on. Go back to about 1 megapixel (for example 1024x1024) and upscale afterwards.
• Colours look burnt and harsh. Lower the CFG scale to around 5 to 7.
• The words seem to be ignored. Check which checkpoint is loaded. An anime checkpoint will not give a realistic photo however you word it.