Advanced Prompting

Advanced techniques for complex and high-resolution images

Fine-tuning the Negative Prompt

A well-structured negative prompt is crucial for high-quality images. With targeted negative prompts, you can avoid common problems:

Quality Issues

blurry, low quality, pixelated, jpeg artifacts, compression artifacts, bad composition, deformed, mutated, signature, watermark

Anatomy Issues

bad anatomy, deformed hands, extra fingers, missing fingers, extra limbs, distorted face, unrealistic proportions, cross-eyed, assymetrical, disfigured

Unwanted Art Style

cartoon style, anime, 3d render, painting, drawing, sketch, illustration, low detail face

Impact of the Negative Prompt

Negative Prompt:

Empty

Negative Prompt:

bad quality, blurry, bad anatomy, worst quality, low quality, lowres, extra fingers, ...

Tips for effective negative prompts:

Adapt project-specifically: Focus on anatomy errors for portraits, quality issues for landscapes
Don't overload: Extremely long negative prompts can confuse the model
Test run: Compare results with and without the negative prompt
Prioritize: Place the most important avoidance terms at the beginning of the negative prompt

Multi-Prompt / Composable Diffusion

Multi-prompt or Composable Diffusion is an advanced technique that allows using multiple separate prompts simultaneously to generate complex scenes.

How it works:

1. Prompt Decomposition

Break down your complex scene into individual elements and create separate prompts for each element.

2. Weighted Combination

Assign a weight to each prompt indicating its importance in the overall composition.

3. Connection

Connect the prompts using specific syntax depending on the UI used.

Example: Landscape with multiple elements

Component 1:

a majestic castle on top of a hill

Weight: 1.2

Component 2:

a beautiful lake with reflections

Weight: 0.8

Common Styles:

sunset, cinematic lighting, detailed, 8k

For all components

Implementation Syntax:

A1111 WebUI:

a majestic castle on top of a hill :1.2 AND a beautiful lake with reflections :0.8, sunset, cinematic lighting, detailed, 8k

ComfyUI:

// Connected in separate prompt nodes with corresponding weights

This technique is particularly useful for:

Complex scenes with multiple main elements
Combining different styles
Precise control over the balance of elements in the image

Hires-Fix & Upscale Loops

For high-resolution images with fine details, special techniques are necessary, as diffusion models often struggle at higher resolutions.

Hires-Fix

This technique first generates an image at low resolution and then refines it in a second pass:

Initial Generation

Generate image at low resolution (e.g., 512x512 px)

→

Upscaling

Enlarge image to target size (e.g., 1024x1024 px)

→

Detail Refinement

Rediffusion with fewer steps for detail enhancement

Typical Settings:

Denoising Strength: 0.3 - 0.6

Lower values = retain more of the original

Upscaler: ESRGAN, Latent

Different upscalers have different style impacts

Second-pass Steps: 15-25

Fewer than in the initial generation

Upscale Loops / Progressive Growing

For particularly large images (2K+), a multi-stage process can work better:

512×512

↓

Initial Generation

→

768×768

↓

First Upscale + Fix
(Denoising: 0.5)

→

1024×1024

↓

Second Upscale + Fix
(Denoising: 0.4)

→

1536×1536+

↓

Final Upscale + Fix
(Denoising: 0.3)

Advantages of this method:

Retains overall composition and important elements
Adds finer details at each step
Avoids artifacts and inconsistencies at very high resolutions
Allows more control over the level of detail in each step

ControlNet & LoRAs: How they work

ControlNet

ControlNet allows you to guide the generation process using an input image. Instead of just working with text, the model can consider outlines, poses, depth maps, or other structural information.

Main Applications:

Canny/Scribble: Convert outlines and sketches into finished images
Pose: Control body postures and gestures
Depth/Normal: Maintain 3D information and spatial structure

LoRAs (Low-Rank Adaptations)

LoRAs are small, trained model add-ons that teach a base model new styles, characters, or concepts without needing to retrain the entire model.

Main Applications:

Styles: Specific art styles or visual aesthetics
Characters: Consistent generation of specific people or figures
Concepts: Special objects or visual elements

Integration into the Workflow

Both technologies can be combined with the prompting techniques learned so far:

🎨

1. Create Base Prompt

Define the main subject and style as usual

🔧

2. Add ControlNet / LoRAs

Select appropriate control inputs and model extensions

⚖️

3. Adjust Weights

Find the balance between text prompt and other inputs

🔄

4. Refine Iteratively

Find the optimal settings through testing

Further Tutorials:

These topics are complex and are covered in separate, specialized tutorials:

Understanding & Applying LoRAs Understanding ControlNet

Summary

You now have a comprehensive overview of advanced prompting techniques:

Use Negative Prompts specifically and tailor them to the use case
Create complex scenes with multiple elements using Multi-Prompt Techniques
Utilize Hires-Fix and Upscale Loops for high-resolution, detailed images
Understand the basic concepts of ControlNet and LoRAs and integrate them into the workflow

Next Steps

To further deepen your skills, we recommend:

Experiment with the learned techniques and document your results
Create your own library of negative prompts for different use cases
Learn specific tools like ControlNet or Image-to-Image in more detail
Share your results and experiences with the community

Course Completion:

Congratulations! You have completed all tutorial modules and now have a solid foundation of knowledge about Stable Diffusion and advanced prompting techniques.