If you’ve dabbled with AI art in Stable Diffusion, you’ve probably come across the term CFG Scale. It’s one of those mysterious sliders that can make or break your image generation, leaving you wondering: “What does it actually do?” Fear not—this guide will demystify CFG (Classifier-Free Guidance), explain how it works in simple terms, and show you how to use it like a pro.

What Is CFG?

At its core, CFG Scale is a tool that helps Stable Diffusion balance between two competing forces:

  1. Following your prompt: This is where CFG pushes the model to stay true to your description.
  2. Creative freedom: This is where CFG allows the model to add its own artistic flair, filling in the blanks with randomness.

Think of CFG as a “creativity knob” for your AI model. Turn it one way, and the model becomes a strict rule-follower, sticking rigidly to your instructions. Turn it the other way, and the model becomes a free spirit, taking liberties to add unexpected details.

How Does CFG Work?

To understand CFG, let’s break it down step-by-step:

  1. Two Paths: When generating an image, Stable Diffusion takes two parallel paths:
    • Prompt Path: The model interprets your text prompt and predicts what the image should look like.
    • Unconditioned Path: The model imagines the image without any guidance (essentially random noise).
  2. Guidance: CFG combines these two paths, amplifying the influence of your prompt while suppressing the randomness from the unconditioned path.
  3. The Scale: The CFG Scale determines how much weight to give your prompt’s influence:
    • Low CFG: The model has more freedom, leading to loose interpretations of your prompt.
    • High CFG: The model adheres strictly to your prompt but may lose creativity or naturalness.

How to Use CFG Effectively

CFG isn’t a one-size-fits-all setting. The optimal value depends on what you’re trying to create. Here’s how to approach it:

1. Start with the Defaults

Most workflows default to a CFG Scale of 7-12. This range strikes a good balance for general use, ensuring your prompt is followed while allowing the model some creative freedom.

2. Fine-Tune Based on Your Goal

Purpose

CFG Setting

Description

For High Adherence

15+

Increase CFG to 15+ if you need the image to match your prompt exactly. Great for detailed prompts or specific instructions (e.g., “a red sports car with black stripes”).

For Artistic Exploration

5 or below

Lower CFG to 5 or below if you’re open to the model interpreting your prompt more loosely. Ideal for abstract art or when you want the AI to surprise you.

3. Avoid the Extremes

Description

Result

Too Low (CFG < 4)

The model can wander too far from your prompt, producing results that don’t resemble what you envisioned.

Too High (CFG > 20)

The output may become overly rigid, looking unnatural or repetitive as the model tries too hard to stick to your description.

CFG in Action: A Real-World Example

Imagine your prompt is: “a futuristic cityscape at night, glowing neon lights, with flying cars.”

CFG = 3

The output might feature a night scene, but the neon lights are muted, and the flying cars might be missing altogether. The AI had too much freedom, leading to a vague interpretation.

CFG = 10

The city is glowing with neon colors, and flying cars zip through the sky. The AI follows your prompt while adding dynamic lighting and realistic textures.

CFG = 20

The image becomes overly sharp and cluttered. Every detail from your prompt is crammed into the scene, making it feel forced or artificial.

CFG and Other Parameters

CFG doesn’t work in isolation—it interacts with other settings in your workflow. Here’s how to leverage it alongside key parameters:

1. Sampling Steps

  • Fewer Steps (10-20): Lower CFG slightly to give the AI more room to explore.
  • More Steps (50+): Higher CFG ensures the model stays on track during extended iterations.

2. Prompts

  • Simple Prompts: Use higher CFG to anchor the model’s interpretation.
  • Complex Prompts: Lower CFG slightly to let the AI balance multiple elements naturally.

3. LoRAs and Embeddings

  • If using a LoRA or embedding, start with a moderate CFG (7-10) to ensure the model integrates the LoRA’s influence effectively without overpowering the base prompt.

Pro Tips for Mastering CFG

  • Test in Batches Use ComfyUI’s batch processing to test different CFG values for the same prompt. This allows you to compare outputs side by side and find the sweet spot.
  • Combine with Style Modifiers Pair a moderate CFG (7-10) with style modifiers like “watercolor” or “cyberpunk” to maintain balance between prompt adherence and artistic flair.
  • Use for Troubleshooting If your output isn’t matching your expectations:
    • Raise CFG if the image feels too vague or random.
    • Lower CFG if the image feels over-detailed or stiff.
  • Experiment with Extremes Occasionally crank CFG up to 20+ or down to 2-3 for artistic exploration. You might stumble upon unexpected and intriguing results.

Final Thoughts: CFG as Your Creative Compass

Classifier-Free Guidance is more than just a slider—it’s a powerful tool for shaping the creative direction of your AI-generated art. By understanding how CFG works and experimenting with its values, you can strike the perfect balance between precision and creativity.

So, the next time you’re tweaking your settings in ComfyUI, remember: CFG is your compass. Whether you’re charting a strict path or wandering into the unknown, it’s there to guide your AI’s imagination. Happy creating!