CFG Scale
A key parameter during diffusion model inference that amplifies the influence of the prompt by computing both conditional and unconditional predictions and extrapolating according to a formula. Higher values adhere more closely to the prompt, but excessively high values cause oversaturation and rigid composition.
Core Facts
Timeline (last 90 days)
Among generation parameters, the sampling steps determine the number of denoising iterations—too few steps make the image rough, while too many yield diminishing returns; CFG Scale controls how closely the model follows the prompt—too high causes oversaturated colors and rigid composition; the random seed determines the initial noise, and a fixed seed reproduces results under the same prompt.
The formula for the classifier-free guidance scale (CFG Scale) is: final prediction = unconditional prediction + CFG × (conditional prediction - unconditional prediction).
Turbo models require the CFG Scale to be set to 1; a high CFG will cause image corruption.
The CFG for standard SD models is typically set around 7.
All Facts (5)
CFG Scale (Classifier-Free Guidance scale) controls the degree to which the generated result adheres to the prompt.
65%UnverifiedAmong generation parameters, the sampling steps determine the number of denoising iterations—too few steps make the image rough, while too many yield diminishing returns; CFG Scale controls how closely the model follows the prompt—too high causes oversaturated colors and rigid composition; the random seed determines the initial noise, and a fixed seed reproduces results under the same prompt.
50%UnverifiedThe formula for the classifier-free guidance scale (CFG Scale) is: final prediction = unconditional prediction + CFG × (conditional prediction - unconditional prediction).
50%UnverifiedTurbo models require the CFG Scale to be set to 1; a high CFG will cause image corruption.
50%UnverifiedThe CFG for standard SD models is typically set around 7.
50%