AI Glossary

The definitive dictionary for ComfyUI and generative AI. Understand exactly what every node, parameter, and slider actually does.

CFG Scale (Classifier Free Guidance)

Determines how strictly the AI should follow your text prompt.

A high CFG (e.g., 7-10) forces the model to strictly adhere to your prompt, but can cause 'deep fried' artifacts. A low CFG (e.g., 2-4) gives the model more creative freedom and generally produces more realistic images, especially with newer models like SDXL or FLUX.

Rule of Thumb

FLUX.1 Dev usually works best with CFG 3.5. SD1.5 prefers CFG 7.

Sampling Steps

The number of iterations the model takes to denoise the image from pure static.

More steps generally mean more detail and higher quality, but it takes longer to generate. However, after a certain point (usually 30-40 steps), the image stops changing significantly.

Rule of Thumb

Standard generation: 20-30 steps. LCM/Turbo models: 4-8 steps.

Sampler (e.g., Euler, DPM++ 2M)

The specific mathematical algorithm used to remove noise during each step.

Different samplers produce slightly different textures and details. 'Euler a' adds new noise every step (never truly settles), while 'DPM++ 2M' is highly deterministic and converges quickly on a clean image.

Rule of Thumb

DPM++ 2M is the community favorite for SDXL.

Scheduler (e.g., Karras, Normal)

The curve or schedule that determines how much noise is removed at each step.

A Karras scheduler heavily favors removing noise at the very beginning of the generation, then makes tiny, fine-tuning adjustments at the very end.

Rule of Thumb

Pairing 'DPM++ 2M' with 'Karras' is the gold standard for crisp realism.

VAE (Variational Autoencoder)

The translator that converts the AI's internal 'latent' data into the final visible pixels.

Without a VAE, your images will look like highly saturated, washed-out messes or completely gray static. Most modern models (like SDXL) have the VAE built-in, but older models required you to load them separately.

Rule of Thumb

If your image looks like a deep-fried meme, check your VAE node.

Denoising Strength

Used in Image-to-Image. Determines how much the original image is altered.

0.0 means the original image is untouched. 1.0 means the original image is completely destroyed and replaced by a brand new generation. 0.3-0.5 is the sweet spot for keeping the structure but changing the style.

Rule of Thumb

Set Denoising to 0.4 to turn a photo of yourself into a cyberpunk painting.