Classifier-Free Guidance (D-CFG)

• Classifier-Free Guidance (D-CFG) is a family of inference-time modifications that refines standard CFG by integrating an iterative, Gibbs-like sampling process to address issues like diversity loss and over-sharpening.
• It employs alternating cycles of noise injection and CFG-guided denoising to correct the bias from missing Rényi divergence components, thereby aligning samples with the target distribution.
• Empirical results on image and audio benchmarks demonstrate significant improvements in metrics such as FID, precision, and semantic alignment over traditional CFG approaches.

Classifier-Free Guidance (D-CFG) is a family of inference-time modifications to conditional diffusion models that adapt, optimize, or refine the guidance mechanism originally introduced in classifier-free guidance (CFG). While standard CFG boosts sample fidelity and prompt alignment by linearly combining conditional and unconditional model outputs at a fixed “guidance scale”, D-CFG proposals address its critical limitations—such as loss of diversity, over-sharpening, inefficient sampling, and lack of theoretical consistency—by leveraging prompt- or feedback-aware scheduling, non-linear updates, geometric corrections, fixed-point iteration, or Gibbs-like MCMC alternations. D-CFG now encompasses a broad set of methodologies essential for state-of-the-art image- and audio-conditional generative modeling.

1. Foundations of Classifier-Free Guidance and Its Limitations

Standard classifier-free guidance (CFG) combines the unconditional denoiser $D_\sigma(x)$ and the prompt-conditional denoiser $D_\sigma^c(x)$ as

$$D_\sigma^{\,c;w}[{\rm cfg}](x) = w\,D_\sigma^c(x) + (1-w) D_\sigma(x)$$

with $w>1$ (Moufad et al., 27 May 2025). This amounts, in score form, to

$$s_\sigma^{c;w}[{\rm cfg}](x) = w\,s_\sigma^c(x) + (1-w) s_\sigma(x),$$

where the guided model approximates sampling from a “tilted” marginal

$$p_\sigma^{\,c;w}[{\rm cfg}](x) \propto [p_\sigma(x|c)]^w p_\sigma(x).$$

However, this does not, in general, yield samples from the correct family of marginals for a well-defined denoising diffusion model (DDM). The paper demonstrates, via Gaussian analysis, that the trajectory induced by CFG does not match the convolved marginals $$\pi_\sigma(x) = \int \mathcal{N}(x; x_0, \sigma^2 I) \pi_0(x_0) dx_0$$ for any data distribution $\pi_0$: the variance and mean at $\sigma \to 0$ generally do not align (Moufad et al., 27 May 2025). This mismatch manifests as excessive contraction (mode collapse) or loss of diversity beyond the guidance-induced sharpening.

2. Theoretical Analysis: The Missing Rényi Correction

Standard CFG is missing a critical term in its effective score function. The true score of the desired tilted target (the power-weighted conditional distribution) includes a Rényi divergence term: $$\nabla \log \pi_\sigma^{c;w}(x) = \nabla \log p_\sigma^{c;w}[{\rm cfg}](x) + (w-1) \nabla R_\sigma(x, c; w)$$ where

$D_\sigma^c(x)$0

This repulsive correction—vanishing as $D_\sigma^c(x)$1—prevents over-contraction and preserves proper tail behavior in guided samples at intermediate noise levels (Moufad et al., 27 May 2025). Neglecting $D_\sigma^c(x)$2 means the CFG ODE over-concentrates on conditional modes, particularly at strong guidance, leading to sample collapse and loss of global diversity.

3. Classifier-Free Gibbs-Like Guidance (D-CFG) Algorithm

The D-CFG approach introduces a theoretically principled, sampling-based correction. Instead of a single forward CFG pass, it alternates Gibbs-like cycles of noise injection and CFG-guided denoising:

1. Initialization: Sample $D_\sigma^c(x)$3 using a conditional diffusion model with no or weak guidance $D_\sigma^c(x)$4.
2. Iterative refinement: For $D_\sigma^c(x)$5,
   • Noising: $D_\sigma^c(x)$6
   • CFG ODE denoising: Track $D_\sigma^c(x)$7 via a probability flow ODE from $D_\sigma^c(x)$8 to 0 with $D_\sigma^c(x)$9
3. Output: Use $$D_\sigma^{\,c;w}[{\rm cfg}](x) = w\,D_\sigma^c(x) + (1-w) D_\sigma(x)$$0 as the final sample.

By iterating, the sampler injects necessary stochasticity (mixing across diverse modes) and applies sharpening without excessive over-concentration. As shown in the paper, the chain is irreducible and admits the targeted power-weighted distribution as stationary for suitable choices of $$D_\sigma^{\,c;w}[{\rm cfg}](x) = w\,D_\sigma^c(x) + (1-w) D_\sigma(x)$$1 and $$D_\sigma^{\,c;w}[{\rm cfg}](x) = w\,D_\sigma^c(x) + (1-w) D_\sigma(x)$$2 (Moufad et al., 27 May 2025).

4. Practical Implementation and Algorithmic Insights

The D-CFG Gibbs-like procedure is minimal and universal, requiring only the ability to (a) sample noise, and (b) run a standard CFG-enabled probability-flow ODE with arbitrary guidance scale $$D_\sigma^{\,c;w}[{\rm cfg}](x) = w\,D_\sigma^c(x) + (1-w) D_\sigma(x)$$3. A representative pseudocode: $w>1$4

• Hyperparameters typically are: total steps $$D_\sigma^{\,c;w}[{\rm cfg}](x) = w\,D_\sigma^c(x) + (1-w) D_\sigma(x)$$4, initial steps $$D_\sigma^{\,c;w}[{\rm cfg}](x) = w\,D_\sigma^c(x) + (1-w) D_\sigma(x)$$5, weak initial guidance $$D_\sigma^{\,c;w}[{\rm cfg}](x) = w\,D_\sigma^c(x) + (1-w) D_\sigma(x)$$6, strong guidance $$D_\sigma^{\,c;w}[{\rm cfg}](x) = w\,D_\sigma^c(x) + (1-w) D_\sigma(x)$$7, $$D_\sigma^{\,c;w}[{\rm cfg}](x) = w\,D_\sigma^c(x) + (1-w) D_\sigma(x)$$8, repetitions $$D_\sigma^{\,c;w}[{\rm cfg}](x) = w\,D_\sigma^c(x) + (1-w) D_\sigma(x)$$9.
• Each loop injects entropy lost by over-sharpening and restores the missing repulsive (diversity-preserving) effect of the absent Rényi divergence term.

The method is parameter-agnostic: as $w>1$0, the injected bias vanishes, and the samples converge to the correct conditional power law target. Finite $w>1$1 accelerates mixing but may introduce an $w>1$2 bias.

5. Empirical Performance and Comparative Results

Experiments were conducted on both image (ImageNet-512, EDM2) and text-to-audio (AudioCaps, AudioLDM2) conditional generation tasks (Moufad et al., 27 May 2025). Main findings:

• On ImageNet-512 with EDM2-S (32 NFE), CFG yields FID=2.29, D-CFG achieves strictly better FID (full tabulated results in the paper), and stronger wins across FD, Precision, and human-perceived alignment.
• In text-to-audio, D-CFG brings improved caption-embedding similarity, coverage-utility, and perceptual measures, yielding clearer timbre, fewer artifacts, and stronger semantic cues.
• In all benchmarks, D-CFG closes the trade-off gap between sample fidelity (sharpness, alignment), and global and fine-grained diversity, with significant wins in compositional or ambiguous prompts.

6. Theoretical and Practical Implications

The D-CFG Gibbs-like approach demonstrates that vanilla CFG, while effective for conditional alignment, cannot achieve exact power-law density sampling across all noise levels due to the missing Rényi-divergence gradient component. The iterative, alternating D-CFG regime corrects this, yielding not only consistency with the desired target distribution but also practical improvements in sample quality and trade-offs.

A key theoretical implication is that as the noise schedule becomes fine-grained ($w>1$3), D-CFG is asymptotically unbiased. For practical models, only modest additional computational overhead is incurred, and no retraining or architecture change is required, rendering the approach broadly applicable to existing modern generative backbones. This methodology is particularly suitable when balancing sample fidelity, trade-off curves, and diversity are all imperative.

7. Connections to Broader Guidance Research

D-CFG’s MCMC-like construction stands in contrast to approaches such as prompt-aware utility-maximization (Zhang et al., 25 Sep 2025), time- or schedule-dependent guidance scaling (Malarz et al., 14 Feb 2025), region-wise or semantic-aware rescaling (Shen et al., 2024), or predictor-corrector guidance and geometric (manifold-based) correction (Jia et al., 12 Mar 2026). All share an overarching goal of mitigating the sharp diversity/fidelity trade-off in default CFG, but D-CFG’s thermodynamically-motivated alternation and exact tail control provide orthogonal and theoretically grounded advantages. For text-to-audio, image, and other modalities, the method generalizes directly.

---

References:

• "Conditional Diffusion Models with Classifier-Free Gibbs-like Guidance" (Moufad et al., 27 May 2025)
• "Prompt-aware classifier free guidance for diffusion models" (Zhang et al., 25 Sep 2025)
• "Classifier-free Guidance with Adaptive Scaling" (Malarz et al., 14 Feb 2025)
• "Rethinking the Spatial Inconsistency in Classifier-Free Diffusion Guidance" (Shen et al., 2024)
• "Manifold-Optimal Guidance: A Unified Riemannian Control View of Diffusion Guidance" (Jia et al., 12 Mar 2026)