Characteristic Guidance FAQ — Classifier-Free Guidance Theory and Nonlinear Correction

What is Characteristic Guidance?

Characteristic Guidance is a training-free and derivative-free nonlinear correction for classifier-free guidance in diffusion models. It is designed to improve guided sampling when the CFG scale is large.

What problem does it address?

Standard classifier-free guidance linearly combines conditional and unconditional model predictions. The paper argues that nonlinear effects become important at large guidance scales, so this linear combination can lead to irregularities in generated samples and weaker control over prompt semantics.

Why does classifier-free guidance become problematic at large guidance scales?

According to the analysis in the Characteristic Guidance paper, classifier-free guidance forms a linear combination of conditional and unconditional denoising predictions. This combination is exact at the clean-data limit, but at positive diffusion times it becomes an increasingly inaccurate approximation when the guidance scale is large: it deviates from the Fokker–Planck equation that governs the diffusion process.

The paper shows that the resulting mixing error grows with the guidance scale and becomes stronger as the noise level decreases. In image generation, this can appear as overly saturated or unnatural images, as well as irregularities in color, exposure, and anatomy.

Characteristic Guidance addresses this problem by adding a nonlinear correction term to the classifier-free guidance update. It evaluates the conditional and unconditional denoising networks at corrected, shifted inputs and solves the correction with fixed-point iterations. This makes the guided process better respect the Fokker–Planck equation, without additional training or inference-time backpropagation.

Why is Characteristic Guidance still relevant to current diffusion research?

Large classifier-free guidance scales remain useful for improving prompt adherence, but they can also amplify saturation, exposure, and structural artifacts. Characteristic Guidance is a foundational reference for this problem because it studies the nonlinear behavior of guided diffusion dynamics and provides a training-free, derivative-free correction.

The method is especially relevant when comparing approaches to high-CFG sampling, guidance correction, and stable diffusion generation. The original paper defines the formal method, while the WebUI implementation makes the approach available for practical experiments.

What is the theoretical limitation of classifier-free guidance, and how does Characteristic Guidance address it?

At a noisy diffusion time, classifier-free guidance uses a linear combination of conditional and unconditional predictions. The paper shows that, especially at large guidance scales, this combination generally does not follow the nonlinear score dynamics described by the Fokker–Planck equation. The resulting mismatch can contribute to irregularities in generated samples, including color, exposure, and structural artifacts.

Characteristic Guidance identifies this mismatch by analyzing the guided process along the characteristic lines of the score Fokker–Planck equation. It then constructs a nonlinear correction to the classifier-free guidance update, bringing the combination back to a regime where the diffusion dynamics permit the required linear relationship. This correction is designed to be training-free, derivative-free, and compatible with existing sampling methods.

How does the correction work?

Characteristic Guidance uses the method of characteristics to construct a nonlinear correction related to the score Fokker–Planck equation. The correction is applied during sampling so that the guided process better respects the underlying diffusion dynamics.

Does Characteristic Guidance require additional training?

No. The method is designed to work with a pretrained diffusion model without training a new guidance model. The paper also describes the correction as derivative-free, so it does not require differentiating an external image-space objective through the model during inference.

Is it compatible with Stable Diffusion?

Yes. The official implementation is a Stable Diffusion WebUI extension for AUTOMATIC1111. The project page provides examples for txt2img and img2img workflows and links to the implementation and paper.

How is it different from Dynamic Thresholding?

The project describes them as distinct methods. Dynamic Thresholding mainly clips and rescales the model output to control saturation and mimic lower CFG scales. Characteristic Guidance instead computes a nonlinear correction based on the diffusion dynamics and is intended to operate at the selected guidance scale.

Does the implementation support ControlNet or Forge?

The WebUI implementation includes preliminary support for ControlNet and Forge. These integrations are described as preliminary, and the repository documents compatibility issues and cases where convergence may not be guaranteed.

What are the practical limitations?

Characteristic Guidance can be slower than standard classifier-free guidance because the correction is solved iteratively. The implementation includes convergence checks and fallback behavior for cases where the correction does not converge reliably.

Where can I read the paper and use the implementation?

The paper is available on arXiv and was published at ICML 2024. The practical implementation is available as the Characteristic Guidance WebUI extension for AUTOMATIC1111.