What is classifier-free guidance? Understanding strength and image changes
Understand conditional and unconditional predictions, prompt alignment, quality tradeoffs, negative prompts, structural controls, and model-specific scales.
Understand the concept
Classifier-free guidance typically combines conditional and unconditional diffusion predictions to steer sampling without a separate classifier. Strength is not a quality score and larger is not universally better. Models and distilled pipelines can differ in guidance mechanisms, so verify meanings and support rather than compare equal numbers across systems.
Conditional and unconditional predictions steer sampling
Text can be a condition with an unconditional reference, combined to steer generation. The original paper discusses quality-diversity tradeoffs rather than universal gains. Guidance changes sampling, not training or factual verification.
Hugging Face provides guiders with differing conventions, and some pipelines use other approximations. A field named guidance does not establish the same original algorithm. No guidance experiment was run.
Inspect prompt alignment together with image quality
For a red cup, stronger guidance can affect color alignment and possibly saturation or artifacts, depending on the model. Inspect subjects and details rather than equate vividness with quality.
CFG does not guarantee counts or geometry and differs from ControlNet-style structural conditioning. Pose errors may need other controls. Negative prompts also have implementation requirements and are not interchangeable with strength.
Compare matched settings and retain full configuration
For a supported pipeline, fix model and inputs and inspect alignment and artifacts across a few settings. This is test design without images. Record seeds and samplers to avoid confounding effects.
Unsupported fields do not become effective merely by adding them to requests. Check documentation and warnings. No universal best value is prescribed; inspect outputs for intended use.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.