Stable Diffusion settings guide

Use SDXL (Stable Diffusion XL) as your base model — it produces significantly better detail and composition than SD 1.5 at the same generation speed.

Always use negative prompts: 'blurry, low quality, bad anatomy, text, watermark, ugly, deformed'.

Set CFG Scale to 7–9 for balanced creativity — below 5 ignores your prompt, above 12 creates oversaturated artifacts.

Use a sampler like DPM++ 2M Karras or Euler a — these are 5–10× faster than DDIM with equal or better quality.

Run locally on Google Colab for free if you don't have an NVIDIA GPU — the T4 GPU handles SDXL at 2–3 seconds per image.

Use ControlNet with a line art input to maintain exact composition while changing style — essential for consistent character design.

Install AUTOMATIC1111 WebUI locally for free — all generation costs nothing, runs indefinitely, private.