The Most Important Setting You Might Be Ignoring
You've crafted the perfect prompt. The words are precise, the style is defined, you hit generate... and the result is either a blurry mess or a crispy, oversaturated nightmare. What went wrong?
Often, the culprit is CFG Scale – one of the most important yet misunderstood parameters in AI image generation. Understanding this single setting can dramatically improve your results.
What Is CFG Scale?
CFG stands for Classifier-Free Guidance. It's a parameter that controls how strongly the AI follows your text prompt versus generating more "free-form" creative output.
Think of it as a dial between two extremes:
- Low CFG (1-5): The AI has more creative freedom. Results may drift from your prompt but often look more natural and organic
- High CFG (15-30): The AI strictly adheres to your prompt. Results match your description closely but may look artificial or oversaturated
A Simple Analogy
Imagine giving instructions to an artist:
- Low CFG: "Paint me a landscape. Feel free to interpret that however you want."
- Medium CFG: "Paint me a sunset landscape with mountains. Make sure those elements are there, but you can add your own touches."
- High CFG: "Paint me exactly this: a sunset landscape with snow-capped mountains, three pine trees on the left, a lake reflecting the orange sky. Every element must be exactly as I describe."
How CFG Scale Works (Technical Explanation)
For those who want to understand the mechanics:
The Denoising Process
AI image generation works by starting with random noise and gradually "denoising" it into an image. At each step, the model makes predictions about what the final image should look like.
Conditional vs. Unconditional
The model actually makes two predictions at each step:
- Unconditional prediction: "What would a generic image look like?"
- Conditional prediction: "What would an image matching this specific prompt look like?"
The CFG Formula
CFG Scale determines how to blend these predictions:
Final = Unconditional + CFG × (Conditional − Unconditional)
- CFG = 1: Only uses the conditional prediction (prompt-guided but weak)
- CFG = 7: Strongly pushes toward prompt-matching content
- CFG = 20: Aggressively enforces prompt adherence (often too much)
Higher CFG means the difference between "what you asked for" and "generic image" is amplified more strongly.
Finding the Sweet Spot
The Common Range: 5-15
Most AI image generators work best in this range. The exact sweet spot depends on:
- The specific model you're using
- The complexity of your prompt
- The style you're targeting
- Personal preference
Model-Specific Recommendations
Stable Diffusion (SD 1.5, SDXL):
- General use: 7-8
- Photorealistic: 5-7
- Artistic/stylized: 8-12
- Maximum prompt adherence: 12-15
Flux Models:
- Flux Schnell: 1-4 (designed for low CFG)
- Flux Dev: 3-5
- Flux Pro: 2-4
Midjourney:
- Uses "stylize" parameter instead (similar concept)
- Lower = more literal, higher = more artistic
DALL-E:
- CFG is handled internally, not user-adjustable
Effects of Different CFG Values
Very Low (1-3)
Characteristics:
- Soft, dreamy quality
- Colors are muted and natural
- Prompt adherence is loose
- May ignore specific details
- Can feel unfocused or random
When to use:
- Abstract or surreal art
- When you want AI creativity
- Soft, atmospheric images
- With Flux models (designed for low CFG)
Low-Medium (4-6)
Characteristics:
- Natural-looking images
- Good balance of prompt following and creativity
- Soft lighting, realistic colors
- Minor details may vary from prompt
When to use:
- Photorealistic images
- Portraits and people
- Natural scenes
- When realism matters more than precision
Medium (7-9)
Characteristics:
- Strong prompt adherence
- Balanced saturation and contrast
- Clear subject definition
- Good detail reproduction
When to use:
- General-purpose generation
- When you need reliable results
- Commercial and product images
- Most Stable Diffusion workflows
Medium-High (10-14)
Characteristics:
- Very strong prompt adherence
- Increased saturation
- Higher contrast
- Details become more pronounced
- Starting to look "AI-generated"
When to use:
- When prompt precision is critical
- Stylized or graphic art
- When specific elements must appear
- Text rendering attempts
High (15-20+)
Characteristics:
- Maximum prompt adherence
- Oversaturated colors
- Harsh, artificial look
- Artifacts and distortions common
- "Crispy" or "fried" appearance
When to use:
- Rarely – usually indicates the prompt needs work instead
- Specific artistic effects
- Testing prompt effectiveness
- Some abstract or glitch art styles
Common Problems and Solutions
Problem: Images Look Blurry or Unfocused
Likely cause: CFG too low
Solution: Increase CFG by 2-3 points. If using Flux, try going from 2 to 4.
Problem: Images Look Oversaturated or Artificial
Likely cause: CFG too high
Solution: Lower CFG by 2-3 points. Most cases do well between 6-8.
Problem: AI Ignores Parts of the Prompt
Likely cause: CFG may be too low, but often the prompt itself needs work
Solution: Try increasing CFG slightly. If that doesn't help, restructure your prompt to emphasize important elements.
Problem: Strange Artifacts or "Deep Fried" Look
Likely cause: CFG significantly too high
Solution: Lower CFG to 7-10 range. The crispy artifacts are a classic sign of excessive guidance.
Problem: Faces Look Distorted
Likely cause: CFG interacting poorly with face generation
Solution: For portraits, lower CFG to 5-7 range. Faces are sensitive to high CFG values.
CFG Scale Strategies
Strategy 1: The Bracketing Approach
When unsure, generate the same prompt at multiple CFG values:
- Generate at CFG 5, 7, 9, 11
- Compare results
- Fine-tune around your favorite
This quickly shows you the optimal range for your specific prompt.
Strategy 2: Match CFG to Content
- Realistic photos: Lower CFG (5-7)
- Illustrations: Medium CFG (7-10)
- Graphic art: Higher CFG (9-12)
- Abstract: Variable (experiment!)
Strategy 3: Adjust for Prompt Complexity
- Simple prompts: Can handle lower CFG
- Complex prompts: May need higher CFG to include all elements
- Very specific prompts: Higher CFG but watch for artifacts
CFG and Other Parameters
CFG vs. Steps
These interact significantly:
- Higher CFG often benefits from more steps to resolve details
- Lower CFG can often use fewer steps without quality loss
- If increasing CFG, consider increasing steps slightly too
CFG vs. Sampler
Different samplers have different CFG sensitivities:
- Euler: Standard CFG response
- DPM++ 2M: Works well with moderate CFG
- DDIM: Can handle higher CFG with less artifacting
CFG vs. Model
Each model has its own optimal CFG range:
- Read model documentation for recommendations
- Custom fine-tuned models may have specific CFG needs
- When switching models, don't assume your usual CFG will work
Advanced: Dynamic CFG
Some advanced workflows use varying CFG throughout generation:
- High CFG early: Establishes composition and key elements
- Lower CFG later: Allows natural detail development
This can produce images that are both prompt-accurate and natural-looking. Tools like ComfyUI support this through custom nodes.
Practical Examples
Portrait Photography
Prompt: "Professional headshot of a business woman, studio lighting, neutral background"
- CFG 5: Soft, natural lighting, slight prompt variation
- CFG 7: Clear studio lighting, accurate to prompt
- CFG 12: Harsh lighting, possibly unnatural skin tones
- Best choice: 5-7
Fantasy Illustration
Prompt: "Epic dragon perched on a crystal mountain, sunset, fantasy art style"
- CFG 5: Atmospheric but details may be vague
- CFG 8: Clear dragon and mountain, good balance
- CFG 12: Very defined elements, heightened colors
- Best choice: 7-10
Product Shot
Prompt: "White sneaker on white background, product photography, clean lighting"
- CFG 5: May not achieve the clean product look
- CFG 8: Clean, professional appearance
- CFG 12: Risk of over-sharpening and artifacts
- Best choice: 7-9
Conclusion
CFG Scale is your most powerful lever for controlling AI image generation quality. Too low and your images drift from your vision; too high and they become artificial and harsh.
The key insights:
- 7-8 is a safe starting point for most models and prompts
- Adjust based on content type – realistic = lower, stylized = higher
- Watch for telltale signs – blurry means too low, crispy means too high
- Different models have different sweet spots – always check documentation
- When in doubt, bracket – test multiple values and compare
Master CFG Scale, and you'll have much more control over your AI-generated images. It's the difference between fighting the AI and collaborating with it.