Stable Diffusion gives you more control than most image generators, but that also means more settings to understand. These are the ones that make the biggest difference.
Negative prompts
A negative prompt lists what you don't want in the image. The model steers away from those concepts during generation. Common uses:
- Quality problems:
blurry, low quality, jpeg artifacts, noise - Unwanted elements:
text, watermark, logo, frame - Anatomy issues:
extra fingers, distorted hands, deformed - Wrong style:
cartoon, illustrationwhen you want a photo
Keep negative prompts focused. A huge list of every possible flaw can make images flat and over-smoothed.
CFG scale
CFG (classifier-free guidance) controls how strictly the image follows your prompt. Low values (around 3–5) give the model more freedom and often look more natural; high values (around 9–12) follow the prompt closely but can look harsh or oversaturated. A range of 5–8 is a good starting point for most models.
Sampling steps
Steps are how many times the image is refined. More isn't always better: many modern samplers look good at 20–35 steps, and going much higher mostly costs time.
Resolution
Generate at or near the resolution the model was trained on (for example around 1024×1024 total pixels for SDXL-based models). Very different sizes can cause repeated subjects or strange compositions. Upscale afterwards if you need a bigger image.
Seed
The seed sets the starting noise. Reusing a seed with the same settings recreates an image, which is perfect for testing one change at a time.
A simple workflow
- Write a clear prompt and a short negative prompt.
- Generate a batch at CFG 6–7 and about 30 steps.
- Pick the best seed and adjust one setting at a time.
- Upscale the final image.