← All Cheat Sheets
CHEAT SHEET

Stable Diffusion

The open-source powerhouse. Free, local, infinitely customizable.

Getting Started

  • Run locally with Automatic1111 WebUI, ComfyUI, or Forge
  • Minimum requirements: 8GB VRAM GPU (NVIDIA recommended)
  • Base models: SD 1.5 (lightweight), SDXL (high quality), SD 3.5 (latest)
  • Download models from CivitAI or Hugging Face
  • Use ComfyUI for advanced workflows, Automatic1111 for simplicity

Essential Settings

  • CFG Scale (7-12): how strictly to follow the prompt. Lower = creative, higher = literal
  • Steps (20-50): more steps = more detail but slower. 30 is the sweet spot
  • Sampler: DPM++ 2M Karras is the most popular all-rounder
  • Resolution: 512x512 for SD 1.5, 1024x1024 for SDXL
  • Negative prompt: always include 'blurry, low quality, deformed, watermark'
  • Seed: set a specific seed to reproduce results

Power Features

  • ControlNet -- use pose references, depth maps, and edge detection to guide generation
  • LoRAs -- small add-on models that teach new styles or characters
  • Inpainting -- regenerate just one part of an image
  • Img2Img -- transform an existing image using a prompt
  • Upscaling -- increase resolution with AI upscalers
  • Batch generation -- generate hundreds of variations automatically

Best Prompts (Copy & Use)

Photorealistic
RAW photo of [subject], 8k uhd, dslr, high quality, film grain, Fujifilm XT3
Negative: cartoon, painting, illustration, drawing, blurry, low quality
Anime/Illustration
[Scene description], anime style, detailed illustration, vibrant colors, studio lighting
Negative: realistic, photo, 3d render, blurry, low quality
Concept Art
[Scene], concept art, digital painting, matte painting, cinematic, detailed, artstation
Negative: photo, realistic, blurry, low quality, watermark
With ControlNet Pose
Full body portrait of [description], standing pose, studio lighting, professional photo
+ ControlNet: OpenPose with reference image

Pro Tips

Use Quality LoRAs

CivitAI has thousands of free LoRAs for specific styles, characters, and concepts. Adding one LoRA can transform your results.

Master Negative Prompts

Good negative prompts matter as much as positive ones. Always exclude: 'blurry, bad quality, deformed hands, extra fingers, watermark, text'.

ControlNet is a Game Changer

Use ControlNet with an OpenPose reference to control exact poses. This single feature gives you more control than any closed AI image tool.

Start Simple, Add Detail

Begin with a short, clear prompt. Generate. Then add details and style keywords gradually.

Best For

  • + Unlimited free generation
  • + Maximum creative control
  • + Custom model training
  • + Privacy (runs locally)
  • + Advanced workflows

Not Ideal For

  • - Non-technical users (requires setup)
  • - Quick one-off images (use Midjourney)
  • - People without a capable GPU

Pricing

Completely free. Open source. Run locally on your own hardware. Cloud options available (RunPod, etc.) for ~$0.50/hour.

Full Stable Diffusion ReviewStable Diffusion Alternatives

More Cheat Sheets

ChatGPT

The all-rounder. Does everything well, nothing poorly.

Claude

The writer and thinker. Best for long, nuanced, careful work.

Midjourney

The artist. Creates the most beautiful AI images, period.

Cursor

The AI code editor. Your entire codebase, understood.

Perplexity

The researcher. Every answer comes with receipts.