guidance scale

**Guidance scale** is the **numeric factor in classifier-free guidance that sets the strength of conditional steering during denoising** - it is one of the most sensitive controls for prompt fidelity versus visual realism. **What Is Guidance scale?** - **Definition**: Multiplies the difference between conditional and unconditional model predictions. - **Low Values**: Produce more natural and diverse images but weaker prompt compliance. - **High Values**: Increase instruction adherence while raising risk of artifacts or oversaturation. - **Context Dependence**: Optimal scale depends on model checkpoint, sampler, and step budget. **Why Guidance scale Matters** - **Quality Tradeoff**: Directly governs realism-alignment balance in generated outputs. - **User Control**: Simple parameter gives non-experts practical control over generation style. - **Serving Consistency**: Preset tuning improves predictability across repeated runs. - **Failure Prevention**: Incorrect scale settings are a common source of degraded images. - **Benchmark Relevance**: Comparisons across models are only fair when guidance settings are aligned. **How It Is Used in Practice** - **Preset Curves**: Set guidance defaults per sampler and resolution, not as a global constant. - **Prompt Classes**: Use lower scales for portraits and higher scales for dense technical prompts. - **Monitoring**: Track artifact rates and prompt hit rates after changing guidance policies. Guidance scale is **a primary control knob for diffusion inference behavior** - guidance scale should be tuned jointly with sampler settings to avoid unstable outputs.