Introduction
Running Stable Diffusion locally gives you unlimited free generation and maximum control. These 15 tips will help you optimise your setup and improve image quality.
Setup and Optimisation
1. Choose the Right UI
Select based on experience:
- Fooocus: Simplest, Midjourney-like
- A1111: Most features, moderate complexity
- ComfyUI: Node-based, maximum control
- Start with Fooocus or A1111
2. Optimise VRAM Usage
Run efficiently on limited VRAM:
- Enable xformers or SDP attention
- Use --medvram or --lowvram flags
- Generate at lower resolution, upscale after
- Close other GPU applications
3. Use SDXL for Quality
Upgrade from SD 1.5:
- SDXL produces better images
- Requires 8GB+ VRAM
- Use SDXL-specific models
- Worth the upgrade if hardware allows
Model Selection
4. Choose Appropriate Base Models
Match model to output:
- Photorealistic: Juggernaut, RealVisXL
- Anime: Anything, CounterfeitXL
- General: SDXL base, Pony
- Test multiple for your style
5. Use LoRAs for Style
Add specific styles efficiently:
- Download from Civitai
- Apply at 0.5-0.8 weight typically
- Stack multiple LoRAs
- Can train custom LoRAs
6. Understand Samplers
Different samplers for different results:
- DPM++ 2M Karras: Good default
- Euler a: Creative variation
- DDIM: Faster, less variation
- Experiment to find preferences
Generation Techniques
7. Master CFG Scale
Control prompt adherence:
- 7-12: Typical range
- Lower: More creative interpretation
- Higher: Closer to prompt (can oversaturate)
- Start at 7, adjust as needed
8. Use Appropriate Steps
Balance quality and speed:
- 20-30: Good for most images
- 40+: Diminishing returns
- Lower for drafts
- More isn't always better
9. Upscale for Final Images
Get high resolution efficiently:
- Generate at native resolution
- Use img2img or dedicated upscalers
- Add detail during upscale
- 4x upscalers work well
ControlNet and Advanced
10. Use ControlNet for Precision
Control generation with references:
- Canny: Edge detection
- OpenPose: Character poses
- Depth: 3D spatial control
- Essential for consistent output
11. Inpaint for Corrections
Fix problems in specific areas:
- Mask problem areas
- Describe replacement
- Use inpaint-specific models
- Fix hands, faces, details
12. Use Negative Prompts Effectively
Remove unwanted elements:
- Standard negatives: "low quality, blurry"
- Specific removals: "no text, no watermark"
- Use embeddings for common negatives
- Don't overload negative prompts
Workflow Tips
13. Use Seeds for Iteration
Refine specific generations:
- Note seeds of good images
- Lock seed, vary prompt
- Helps achieve consistency
- Useful for character work
14. Generate in Batches
Work efficiently:
- Generate 4-8 images at once
- Select best, refine
- Faster than one-at-a-time
- More options to choose from
15. Build Personal Workflows
Save time with automation:
- Save favourite prompt templates
- Create workflows in ComfyUI
- Use styles presets
- Document what works for you
Conclusion
Local Stable Diffusion generation rewards investment in learning. These techniques compound over time—the more you experiment, the better your results become.
