Analytic Distribution of Classifier-Free Guidance for Schedule Design

Technical Summary: Analytic Distribution of Classifier-Free Guidance for Schedule Design

Problem Statement
Classifier-Free Guidance (CFG) is the standard mechanism for conditional generation in diffusion models, blending conditional and unconditional score functions via a guidance strength parameter $\omega$. While empirically successful, a fundamental theoretical question remains unresolved: what distribution does the deterministic CFG reverse process actually sample from? A common heuristic assumes CFG samples from a normalized product distribution $\tilde{r}_0 \propto p_0^\omega q_0^{1-\omega}$. However, this pointwise interpretation fails to capture the cumulative effect of time-dependent score perturbations along the entire sampling trajectory. Existing theoretical works explain specific failure modes (e.g., mode collapse, oversaturation) or propose heuristics but do not provide an exact characterization of the induced distribution under the deterministic probability flow ODE. This lack of an exact analytic form hinders the principled design of time-dependent guidance schedules.

Methodology
The authors analyze CFG within the framework of the deterministic probability flow Ordinary Differential Equation (ODE). Their approach involves three main stages:

1. Theoretical Derivation:

   • They formulate the guided reverse process as an ODE where the score is replaced by the affine combination $s_\omega^t(x) = \omega \nabla \log p_t(x) + (1-\omega) \nabla \log q_t(x)$.
   • By deriving the Fokker–Planck equation for the normalized product reference distribution $\tilde{r}_t$, they identify an "interaction potential" term arising from the discrepancy between the conditional and reference scores.
   • Using characteristic-based analysis, they solve for the exact analytic path-integral representation of the induced distribution $\hat{r}_{t_0}$ at a stopping time $t_0$. This derivation covers both constant guidance ($\omega(t) \equiv \omega$) and time-dependent guidance ($\omega(t)$).

2. Schedule Design (DG-CFG):

   • Analyzing the derived path-integral formula, the authors observe that the effective weight of guidance at each timestep is modulated by the noise schedule coefficient $\beta(t)$ (in the VPSDE setting) and the guidance deviation $\omega(t)-1$.
   • They identify that standard constant CFG implicitly overweights the high-noise regime because $\beta(t)$ varies by orders of magnitude, leading to saturation and diversity loss.
   • To correct this, they propose Distribution-Guided CFG (DG-CFG), a schedule designed to balance the path-integral contribution. The schedule is constructed by:
     • Compensating for the non-uniformity of $\beta(t)$.
     • Weighting by signal content ($\sqrt{\bar{\alpha}_t}$) to focus on timesteps with meaningful signal.
     • Mitigating score-approximation error amplification in the low-noise regime (where $\sigma_t \to 0$) by scaling with $(1-\bar{\alpha}_t)$.
   • The resulting formula is: $\omega(t) = 1 + C \cdot (\bar{\omega} - 1) \cdot \frac{(1-\bar{\alpha}_t)\sqrt{\bar{\alpha}_t}}{\beta(t)}$.

3. Experimental Validation:

   • Toy Model: They verify the theory on a discrete distribution on a circle with analytic scores, comparing empirical samples against the theoretical path-integral predictions.
   • Large-Scale Evaluation: They evaluate DG-CFG on Stable Diffusion 1.5 (SD1.5) using the MS-COCO 2017 validation set, comparing it against constant CFG and heuristic baselines (Interval CFG, $\beta$-CFG) across various guidance strengths and sampling step budgets.

Key Contributions

1. Exact Analytic Distribution: The paper derives the first exact analytic path-integral representations for the distribution induced by deterministic CFG under both constant and time-dependent guidance (Theorems 4.3 and 4.4). These formulas reveal that CFG modifies the target distribution via an exponential path-integral correction weighted by $\omega(t)-1$, rather than simply sampling from the product distribution.
2. Principled Schedule Design: Based on the path-integral characterization, the authors propose DG-CFG. This schedule explicitly balances the temporal contributions of guidance, correcting for noise schedule non-uniformity, signal content variation, and low-noise error amplification.
3. Empirical Verification: The theory is rigorously verified on a toy model where the predicted distributions match empirical samples. On SD1.5, DG-CFG demonstrates improved robustness, better control over color saturation, and a superior diversity–fidelity trade-off, particularly at high guidance strengths where baselines degrade.

Results

• Toy Model: Empirical distributions from DDIM sampling closely matched the theoretical predictions (Total Variation distance $\approx 1-2\%$). Time-dependent guidance significantly improved mode balance (Min-Max ratio) compared to constant CFG, especially at high guidance strengths ($\omega=9$), preventing the collapse of peripheral modes.
• Stable Diffusion 1.5:
  • High Guidance Performance: At high guidance strengths ($\bar{\omega} \ge 15$), constant and heuristic baselines exhibited severe color oversaturation and quality degradation (negative ImageReward scores). DG-CFG maintained positive ImageReward scores and competitive FID/CLIP scores.
  • Saturation Control: DG-CFG kept saturation metrics significantly closer to real-image references compared to baselines, which often exceeded reference levels by wide margins.
  • Efficiency: In fixed-quality analyses, DG-CFG required fewer sampling steps (NFE) to reach specific ImageReward, CLIP, and FID targets compared to all baselines. For instance, to achieve an ImageReward $\ge 0.2$ at $\bar{\omega}=23$, DG-CFG required 12 steps, while the strongest baseline required 16.

Significance and Claims
The paper claims that its primary significance lies in providing a theoretical foundation for understanding CFG. By moving beyond the "product-distribution heuristic" to an exact path-integral characterization, the authors clarify why constant guidance fails (temporal imbalance in the path integral) and how to correct it.

The authors state that DG-CFG is not merely a heuristic but a schedule derived directly from the induced distributional representation. The results suggest that this principled approach allows for stronger guidance strengths without the typical trade-offs of saturation and diversity loss, effectively reducing the computational cost (sampling steps) required to achieve high-quality conditional generation. The work positions itself as complementary to existing studies on failure modes, offering a unified analytic framework that explains accumulation of score discrepancies and guides schedule design.