Guidance Scale

The guidance scale is a scalar multiplier applied during the sampling step of a conditional diffusion model. It amplifies the direction in latent space pointed to by the conditioning signal. The core function is to sharpen the probability distribution of the generated output, increasing the likelihood that the final sample aligns with the provided condition (e.g., a text prompt) at the potential cost of reduced sample diversity.

• Mathematically, in classifier-free guidance, the adjusted noise prediction is: ε_adj = ε_uncond + w * (ε_cond - ε_uncond), where w is the guidance scale.
• A scale of 1.0 applies no guidance, using only the conditional model's prediction.
• Scales greater than 1.0 increasingly bias the generation toward the condition.

Adjusting the guidance scale directly mediates a fundamental trade-off in conditional generation.

• Low Guidance (w ≈ 1.0 - 3.0): Produces more diverse, creative, and sometimes unexpected outputs. The model has greater freedom to explore the data manifold, which can lead to interesting variations but may ignore specific prompt details.
• High Guidance (w ≈ 7.5 - 20.0): Enforces strict adherence to the conditioning prompt. This typically improves image-prompt alignment and sharpens details, but can reduce diversity, cause over-saturated colors, and introduce artifacts if pushed too high. The model's outputs become more deterministic and less varied.

Classifier-Free Guidance is the predominant technique that utilizes the guidance scale. It eliminates the need for a separate, often unstable, classifier model by training a single diffusion model to perform both conditional and unconditional generation.

• During training, the condition (e.g., text) is randomly dropped (replaced with a null token) with a fixed probability (e.g., 10%).
• This yields two capabilities in one model: ε_cond(x_t, t, c) and ε_uncond(x_t, t).
• At sampling, the guidance scale w amplifies the difference between these two predictions: ε = ε_uncond + w * (ε_cond - ε_uncond).
• This vector points more strongly toward data regions consistent with condition c.

Optimal guidance scale values are model-dependent but fall within established ranges. Exceeding these ranges leads to degradation.

• Stable Diffusion (v1.x/v2.x): Common effective range is 5.0 to 15.0. 7.5 is a standard default.
• Zero-shot, uncurated models: May require higher scales (10.0-20.0) to maintain prompt coherence.
• Overguidance (w >> 15.0): Causes saturated colors, unnatural contrast, repetitive details, and 'burned' or grimy artifacts. The sample diversity collapses.
• Underguidance (w ~ 1.0): Outputs are often blurry, lack detail, and show poor prompt adherence, resembling an unconditional model.

The guidance scale does not operate in isolation; its effect is intertwined with other sampling parameters.

• Sampling Steps: Higher guidance can sometimes allow for fewer sampling steps to achieve coherent results, but too-high guidance with few steps increases artifact risk.
• Sampler Choice: Deterministic samplers (like DDIM) are more sensitive to high guidance scales than stochastic samplers (like Euler Ancestral).
• Negative Prompting: The negative prompt acts as a secondary condition. The guidance scale amplifies movement away from the negative concept as defined by ε_cond(neg) - ε_uncond. This makes prompt engineering more precise.

Beyond basic prompt adherence, the guidance scale enables sophisticated control techniques.

• Per-Prompt Optimization: Automated systems can search for an optimal w per prompt to maximize a metric like CLIP score.
• Dynamic Guidance: The scale can be scheduled to change over the sampling steps (e.g., lower early for composition, higher later for detail).
• Multi-Concept Scaling: Different scales can be applied to different parts of a composed prompt using attention-aware techniques, allowing fine-grained control over specific attributes.
• Seed Consistency: At very high guidance, the influence of the initial random noise seed diminishes, leading to more deterministic outputs for a fixed prompt.