Advanced

Best Practices

Optimization tips, recommended hardware, ethical guidelines, and solutions to common problems when working with Stable Diffusion.

Hardware Recommendations

  • Minimum: NVIDIA GPU with 4GB VRAM (GTX 1650) — SD 1.5 at 512x512 with optimizations
  • Recommended: NVIDIA GPU with 8GB VRAM (RTX 3060/4060) — SD 1.5 and SDXL comfortably
  • Ideal: NVIDIA GPU with 12-24GB VRAM (RTX 4070 Ti/4090) — all models, ControlNet, high resolution
  • CPU/Apple Silicon: Possible but slow. Apple M1/M2/M3 work via MPS backend

Optimization Tips

  • Use half precision (FP16) to halve VRAM usage with negligible quality loss
  • Enable xformers or Flash Attention for faster generation and lower memory
  • Use VAE tiling for generating images larger than your VRAM allows
  • Lower inference steps to 20-25 (diminishing returns beyond 30 for most schedulers)
  • Use fast schedulers like DPM++ 2M Karras or Euler a for good results in fewer steps
  • Generate at native resolution (512 for SD1.5, 1024 for SDXL), then upscale

Ethical Considerations

⚠
Respect copyright: Be thoughtful about using artist names in prompts. The training data includes copyrighted work, and mimicking specific artists' styles raises ethical questions.
⚠
Transparency: Label AI-generated images when sharing them publicly. Do not present AI-generated images as photographs or original artwork without disclosure.
⚠
Deepfakes: Never use these tools to create non-consensual content of real people. Many jurisdictions have laws against this.

Troubleshooting

  • CUDA out of memory: Reduce resolution, enable FP16, use xformers, or lower batch size
  • Blurry or distorted faces: Add face-specific quality terms to prompt, use a face restoration model (CodeFormer, GFPGAN), or use ADetailer extension
  • Wrong composition: Use ControlNet for structural guidance, or try img2img with a rough sketch
  • Oversaturated colors: Lower CFG scale (try 5-7), or add "natural colors" to prompt
  • Repeated patterns/artifacts: Change the scheduler, adjust step count, or try a different seed

Workflow Tips

  1. Start with low steps (15-20) and small batches to iterate on prompts quickly
  2. Once you find a good prompt, increase steps and generate multiple seeds
  3. Use img2img to refine promising results
  4. Use inpainting to fix specific areas rather than regenerating everything
  5. Save your favorite prompts, seeds, and settings for reproducibility
  6. Upscale final images with a dedicated upscaler (Real-ESRGAN, 4x-UltraSharp)
✅
Congratulations! You have completed the Stable Diffusion course. You now understand how diffusion works, can craft effective prompts, use ControlNet, fine-tune custom models, and work with professional tools.
← Previous Tools & UIs