acceptodds
Under review as a conference paper at ICLR 2027

Distribution-Guided Classifier-Free Guidance

Abstract

Classifier-free guidance (CFG) is central to conditional generation with diffusion models, but using a constant guidance strength overlooks how the diffusion rate, signal content, and score reliability vary throughout denoising. We introduce Distribution-Guided CFG (DG-CFG), a training-free schedule that reallocates a fixed guidance budget across timesteps to improve the balance among generation quality, diversity, and stability. A path-integral analysis of deterministic CFG identifies the temporal weight that governs how guidance accumulates along sampling trajectories. Combining this principle with signal-content weighting and error-motivated low-noise attenuation yields a middle-focused schedule that emphasizes informative and reliable stages of denoising. Controlled experiments with analytic scores support the path-integral predictions and illustrate the effect of temporal reweighting. Across different backbones and sampling methods, DG-CFG provides a favorable diversity–fidelity trade-off and improves generation quality under strong guidance while controlling saturation. Component ablations and sampling-budget evaluations further support the schedule design and robustness.

Primary area:
generative models
Keywords:
Classifier-Free Guidance, Diffusion Models, Probability Flow ODE

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2nj8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DG-CFG, a training-free guidance schedule motivated by a path-integral characterization of deterministic CFG. It combines inverse diffusion-rate weighting, signal-content weighting, and low-noise attenuation while matching an integrated guidance budget. Experiments include analytic toy distributions and Stable Diffusion models with different samplers and sampling budgets.

Critical strengths

The path-integral characterization provides a useful perspective on time-dependent CFG. The method is easy to implement, and the analytic experiments, component ablations, and NFE sweeps provide evidence for improved stability under strong guidance.

Critical weaknesses
  1. The connection between the analysis and the proposed schedule remains partly heuristic. Equation (17) contains a state-dependent term evaluated along a guidance-dependent trajectory. Canceling its explicit diffusion-rate coefficient does not equalize the actual distributional contribution across time. The additional signal and attenuation factors are plausible, but the analysis does not establish why this particular combination should offer a better trade-off than other schedules.

  2. The experiments establish robustness at high guidance more clearly than superiority over well-tuned alternatives. At moderate guidance, DG-CFG can have worse FID; on SDXL at guidance 7, it obtains 76.79 versus 72.86 for constant CFG and 72.29 for β–CFG. Interval boundaries and the beta profile are fixed rather than systematically tuned. Comparing the best attainable quality–diversity trade-offs would be more informative than emphasizing equal nominal guidance. Evaluation on 1,000 COCO prompts without uncertainty estimates also makes small differences difficult to assess.

Questions for authors
  1. Can the analysis justify the specific schedule beyond balancing the explicit coefficient, or should it be presented as a theory-motivated heuristic?
  2. How do the methods compare after independently tuning guidance strength and schedule parameters under equal tuning budgets? Can uncertainty estimates be provided for the main FID and diversity comparisons?
1 comment

Q1. Theoretical justification of the specific schedule

Thank you for your careful review and this important question about the scope of the theoretical justification. We expanded the analysis to derive the specific schedule by maximizing an efficiency lower bound in a two-component Gaussian mixture model, beyond balancing the explicit temporal coefficient.

  1. From the density analysis to an allocation criterion. Theorem 4.3 identifies ht=β(t)(ω(t)−1)h_t=\beta(t)(\omega(t)-1) as the explicit temporal coefficient in the accumulated density correction. We then specify a fixed guidance budget, ∫t0T(ω(t)−1)dt=B\int_{t_0}^T(\omega(t)-1)dt=B, and a common noise-prediction perturbation constraint. Theorem 5.1 establishes an efficiency lower bound balancing distributional response against perturbation risk:
\frac{(\int_{t_0}^T a_th_t\,dt)^2}{\int_{t_0}^T h_t^2/\sigma_t^2\,dt}, \quad a_t=\sqrt{\bar\alpha_t},\quad\sigma_t^2=1-\bar\alpha_t.$$ 2. **Derivation of the full schedule.** By Cauchy-Schwarz, this lower bound is uniquely maximized by $h_t^*=c_Ba_t\sigma_t^2$, where $c_B$ enforces the budget. Substituting the definition of $h_t$ yields DG-CFG: $$\omega(t)=1+C(\bar\omega-1) \frac{\sqrt{\bar\alpha_t}(1-\bar\alpha_t)}{\beta(t)}.$$ The signal and noise factors are thus jointly supported by the efficiency-bound optimization, providing justification beyond balancing the explicit coefficient. 3. **Scope of the claim.** The result maximizes the stated lower bound in the two-component model. It motivates the schedule for practical diffusion models as stated in Sec. 5.3. We retain time balancing, signal-content weighting, and low-noise attenuation as intuitive interpretations of the resulting schedule. **Revision locations:** Secs. 5.1-5.3, pages 5-7; Theorem 5.1 and Eqs. (21)-(22), page 6; proof in Appendix B.3, pages 21-25. Revisions are highlighted in blue. ## Q2. Independently tuned comparisons and uncertainty estimates Thank you for raising these important questions about tuning fairness and statistical uncertainty. **We added comparisons at method-specific guidance strengths and bootstrap standard errors for the main metrics, including FID and diversity.** 1. **Tuning protocol.** We evaluate SDXL with DDIM and Euler at 8 steps. DG-CFG introduces no hyperparameters beyond $\bar\omega$. For baselines, we use parameters recommended in the original papers. When recommendations are unavailable, we tune these parameters by grid search on a small sample subset. This provides broadly comparable tuning budgets across methods. With these parameters fixed, we sweep the same guidance-strength grid and select method-specific configurations by jointly considering ImageReward (IR), CLIP, and FID. 2. **Tuned results.** After method-specific tuning, DG-CFG retains the highest observed perceptual quality (ImageReward) and text alignment (CLIP) among the selected configurations with both samplers. These gains come with slightly higher FID, clarifying DG-CFG's advantage for perceptual quality and prompt alignment. Key comparisons are below. Table 3 includes all baselines. | Sampler | Method | $\bar\omega$ | FID | CLIP | IR | |---|---|---:|---:|---:|---:| | DDIM | $\beta$-CFG | 11 | 70.65 | 32.06 | 0.350 | | DDIM | DG-CFG | 19 | 73.57 | 32.10 | 0.417 | | Euler | C\$^2\$FG | 11 | 70.36 | 31.84 | 0.316 | | Euler | DG-CFG | 19 | 73.34 | 31.91 | 0.402 | 3. **FID and diversity uncertainty.** We now report bootstrap SEs from 200 resamples, using matched real-generated image pairs for FID and per-prompt Vendi scores for diversity. Diversity is averaged over 100 prompts, each with 16 samples from different initializations. **Revision locations:** Sec. 6.2.3, pages 8-9, and Table 3, page 9 (tuned comparison); Appendix D.1, page 28 (uncertainty protocol); Table 2, page 8, and Tables 7-9, pages 29-30 (main FID and diversity results with SEs); Table 10, page 34 (Euler results with SEs). Revisions are highlighted in blue.
Reviewer_PjMXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DG-CFG, a training-free time-dependent schedule for classifier-free guidance. The main theoretical contribution is a path-integral characterization of the distribution that deterministic (PF-ODE) CFG sampling actually produces, for both constant and time-varying guidance. The analysis shows that the guidance at each step enters the final distribution with a weight proportional to the diffusion rate β(t), which the authors argue makes constant CFG implicitly over-weight high-noise steps. DG-CFG compensates by scaling the guidance deviation inversely with β(t), and further multiplies in two heuristic factors (retained signal content and low-noise attenuation), while keeping the total integrated guidance equal to that of constant CFG. The result is a schedule that concentrates guidance at intermediate noise levels.

Experiments on a toy model with analytic scores confirm the path-integral prediction. On SD1.5/2.1/XL with few-step sampling, DG-CFG is compared with constant, Interval and β-CFG (and with C2FG, APG, CFG Rescale on SDXL) at matched guidance budget; the authors report better quality and much lower saturation at strong guidance, with ablations and an NFE sweep supporting the design.

Critical strengths
  1. The path-integral characterization (Theorems 3.2–3.3) gives an exact expression for the distribution that PF-ODE CFG sampling produces under an arbitrary time-dependent schedule. The proofs are correct and the toy experiment verifies the result numerically. This puts guidance scheduling, previously purely empirical, on a footing where different schedules can be compared under one principle.
  2. The resulting schedule is training-free, has no per-model hyperparameters beyond the nominal strength, and preserves the total guidance budget. Its saturation control is consistent across all backbones, samplers and NFE settings tested, which makes it a credible drop-in replacement in practice.
Critical weaknesses
  1. Although the authors contrast DG-CFG with Interval CFG and β\beta-CFG, the proposed schedule is still a middle-focused bump of the same family, and the theoretical core overlaps with prior analyses: the Fokker--Planck correction term of pωq1−ωp^\omega q^{1-\omega} (Lemma B.1) appears in Feynman--Kac Correctors [1], the conclusion that CFG does not sample the product distribution is in the predictor--corrector analysis of CFG [2], and the schedule-analysis angle is in the CFG weight scheduler study [3], which is not cited. The paper does not discuss how its results differ from these, which limits its novelty.

  2. The method is only partly supported by its own analysis. The path integral motivates only the 1/β(t)1/\beta(t) factor, and even that depends on using the DDPM step index as the time clock; the other two factors are heuristic. The ablation shows that 1/β1/\beta alone (Medium 1) performs as well as, or even better than, the full schedule (e.g., in Table 3, Medium 1 achieves better FID and saturation than DG-CFG on both SD1.5 and SD2.1), and the toy model validates a schedule (αˉ/β\sqrt{\bar{\alpha}}/\beta) that is the worst-performing variant on Stable Diffusion.

[1] Feynman-Kac Correctors in Diffusion: Annealing, Guidance, and Product of Experts. ICML 2025.

[2] Classifier-Free Guidance is a Predictor-Corrector. arXiv 2024.

[3] Analysis of Classifier-Free Guidance Weight Schedulers. TMLR 2024.

Questions for authors
  1. Please expand the discussion of related work to make the contribution clearer: how do the theoretical results differ from existing analyses of what CFG actually samples, and how does the proposed schedule differ from existing time-dependent schedules beyond the 1/β(t)1/\beta(t) factor?
  2. The path integral motivates only the 1/β(t)1/\beta(t) factor, and the ablation shows that this factor alone (Medium 1) matches the full schedule on FID and saturation. What evidence supports adding the αˉt\sqrt{\bar{\alpha}_t} and (1−αˉt)(1-\bar{\alpha}_t) factors? Error bars from multiple seeds would help here.
  3. The "time balancing" argument depends on the time parameterization: in log-SNR time λt=log⁡αˉt1−αˉt\lambda_t=\log\frac{\bar{\alpha}_t}{1-\bar{\alpha}_t} one has g2(t) dt=−σt2 dλg^2(t)\,dt=-\sigma_t^2\,d\lambda, so the weight becomes σt2(ω−1)\sigma_t^2(\omega-1) and would lead to a different schedule. Why is the DDPM step index the appropriate clock?
1 comment

Q1. Relation to prior CFG analyses and schedules

Thank you for this important request to clarify our contribution. We expanded the related-work discussion to distinguish our endpoint-density analysis and schedule derivation from [1-3].

  1. Distributional analysis. [2] establishes CFG's discrepancy from the gamma-powered product distribution and a predictor-corrector interpretation in the SDE limit. [1] derives Feynman-Kac corrections for prescribed geometric-average marginals. We acknowledge its centered score-discrepancy potential in our auxiliary lemma. Building on this evolution, we characterize deterministic CFG's endpoint density along characteristics and extend it to time-dependent guidance (Theorems 4.2-4.3). We further derive the schedule by maximizing a response-perturbation efficiency lower bound in a two-component Gaussian mixture model under a fixed budget (Theorem 5.1).

  2. Schedule design beyond inverse-rate weighting. [3] studies heuristic and parameterized schedules with matched guidance. DG-CFG shares the middle-focused profile of interval and beta-shaped schedules, but derives its allocation by maximizing an efficiency lower bound. Theorem 5.1 shows that, under a fixed guidance budget, this bound is uniquely maximized by

ht∗=cBatσt2,h_t^*=c_Ba_t\sigma_t^2,

where cBc_B enforces the budget. This derivation determines both factors without extra schedule-shape hyperparameters.

Revision locations: Sec. 2, page 2; Appendix A.3-A.4, page 14. Revisions are highlighted in blue.

Q2. Evidence for the signal and noise factors

Thank you for examining the ablations and requesting error bars. We added a theoretical justification for the joint factors, clarified their complementary empirical effects, and reported bootstrap standard errors.

  1. Theory. The path integral identifies the temporal coefficient. Theorem 5.1 allocates guidance by maximizing an efficiency lower bound:
\frac{(\int a_th_t\,dt)^2}{\int h_t^2/\sigma_t^2\,dt}.$$ The numerator measures distributional response and the denominator bounds perturbation sensitivity. Cauchy-Schwarz gives the maximizing allocation $h_t\propto a_t\sigma_t^2$, supporting both factors jointly. 2. **Ablations.** Signal-content weighting and low-noise attenuation should act jointly. Constant CFG vs. Medium 4 and Medium 1 vs. DG-CFG show that adding both factors improves CLIP and ImageReward on both backbones under strong guidance, with modest FID changes and controlled saturation. Their combination balances signal content and error sensitivity, consistent with Theorem 5.1. Medium 3 vs. DG-CFG isolates signal weighting, which restores emphasis on informative stages. Medium 2 vs. DG-CFG isolates low-noise attenuation, which limits sensitivity to prediction errors at low noise. Both additions improve FID and ImageReward on SD1.5 and SD2.1. 3. **Uncertainty.** The theoretical analysis and complementary ablations above provide substantial support for adding signal-content weighting and low-noise attenuation. We therefore did not add error bars across independent generation seeds. To quantify metric uncertainty, we added bootstrap SEs from 200 resamples. We hope these estimates address your request on variability. **Revision locations:** Theorem 5.1, page 6; Sec. 6.2.5, pages 9-10; Table 4, page 10; Appendix D.1, page 28, and D.5, pages 35-37. Revisions are highlighted in blue. ## Q3. Dependence on the time parameterization Thank you for this important observation about time parameterization. **We now base the schedule on the efficiency-bound optimization and time balancing serves as an intuitive interpretation of the resulting schedule.** 1. **Consistent coordinate changes preserve the optimization problem.** We first specify the guidance budget and perturbation constraints, then derive the schedule by maximizing an efficiency lower bound. Under a change of time coordinates, transforming the entire problem, including the integration measures, budget, and perturbation constraints, gives an equivalent optimization problem. Its optimizer preserves guidance strength at corresponding noise states. The conclusion is coordinate-independent for the same problem. 2. **Time balancing alone does not uniquely determine the schedule.** We agree that deriving $1/\beta(t)$ solely by equalizing the temporal coefficient depends on the chosen clock. In the revision, time balancing is retained only as an intuitive interpretation of one schedule component. The formal justification is the efficiency-bound optimization in Theorem 5.1, without requiring a uniquely appropriate DDPM clock. **Revision locations:** Secs. 5.1-5.3, pages 5-7. Revisions are highlighted in blue. References: [1] Feynman-Kac Correctors in Diffusion: Annealing, Guidance, and Product of Experts. ICML 2025. [2] Classifier-Free Guidance is a Predictor-Corrector. arXiv 2024. [3] Analysis of Classifier-Free Guidance Weight Schedulers. TMLR 2024.
Reviewer_BG6wOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents DG-CFG, a training-free method that reallocates guidance strength across denoising timesteps. A path-integral analysis characterizes the distribution induced by deterministic CFG and motivates the guidance schedule. Controlled toy experiments support the theoretical predictions, while experiments on Stable Diffusion models demonstrate improved robustness under strong guidance and favorable quality–diversity trade-offs.

Critical strengths

1.The path-integral characterization provides a useful theoretical perspective on how guidance accumulates along sampling trajectories and affects the generated distribution. The agreement between theoretical predictions and empirical results in the controlled toy setting strengthens this contribution. 2.The proposed method is simple to implement and requires no additional training. Its effectiveness in mitigating quality degradation under strong guidance is supported across multiple models and sampling settings.

Critical weaknesses

1.The connection between the theoretical analysis and the final schedule remains incomplete. The final design therefore remains partly heuristic, and the ablations provide limited evidence for the mechanisms attributed to its individual components. 2.The empirical benefits are most evident under strong guidance. Although guidance-strength sweeps are provided, it remains unclear how much improvement persists when each method is independently tuned for its preferred operating point. This limits the evidence for practical advantages beyond robustness to large guidance strengths.

Questions for authors

1.Given the reported guidance-strength sweeps, how does DG-CFG compare with baselines when each method uses its own best guidance strength and relevant schedule parameters? Could the authors provide trade-off curves, such as FID versus CLIP or ImageReward, to make the method’s advantages clearer? 2.The theoretical analysis describes the distribution produced by CFG, but its connection to the specific schedule in Eq. (18) remains unclear to me. Could the authors explain more directly how the analysis supports this design? 3. Table 3 shows that adding signal-content weighting to Medium 1 worsens both FID and ImageReward in Medium 2. Could the authors explain why this happens and how signal-content weighting helps in the final DG-CFG method?

1 comment

Q1. Independently tuned comparisons and trade-off curves

Thank you for your careful review and this important question about fair tuning. We added comparisons at method-specific guidance strengths and full trade-off curves.

  1. Protocol. We evaluate SDXL with DDIM and Euler at 8 steps. DG-CFG introduces no hyperparameters beyond ωˉ\bar\omega. For baselines, we use the parameters recommended in the original papers. When recommendations are unavailable, we tune these parameters by grid search on a small sample subset. This provides broadly comparable tuning budgets across methods. We then sweep the same guidance-strength grid for each method and select configurations by jointly considering IR, CLIP, and FID.

  2. Tuned results. After method-specific tuning, DG-CFG retains the highest observed perceptual quality (ImageReward) and text alignment (CLIP) among the selected configurations with both samplers. These gains come with slightly higher FID, clarifying DG-CFG's advantage for perceptual quality and prompt alignment. Key comparisons are below. Table 3 includes all baselines.

SamplerMethodωˉ\bar\omegaFIDCLIPIR
DDIMβ\beta-CFG1170.6532.060.350
DDIMDG-CFG1973.5732.100.417
EulerC$^2$FG1170.3631.840.316
EulerDG-CFG1973.3431.910.402
  1. Trade-offs. The new IR, CLIP, and saturation versus FID curves show that DG-CFG favors higher IR and CLIP, with slightly higher FID. Beyond the best-performing guidance range, its smaller IR/CLIP declines and FID increase, and restrained saturation growth indicate greater robustness across the evaluated strengths.

Revision locations: Sec. 6.2.3, page 8; Table 3, page 9; Appendix D.4, Fig. 7, page 35. Revisions are highlighted in blue.

Q2. Connection between theory and schedule

Thank you for identifying this important gap. We now derive DG-CFG by maximizing an efficiency lower bound in a two-component Gaussian mixture model, balancing distributional response against noise-prediction perturbation risk.

  1. Allocation variable. Theorem 4.3 identifies ht=β(t)(ω(t)−1)h_t=\beta(t)(\omega(t)-1) as the explicit temporal coefficient in the density correction. We optimize its allocation under ∫t0T(ω(t)−1)dt=B\int_{t_0}^T(\omega(t)-1)dt=B.

  2. Efficiency bound. Theorem 5.1 bounds efficiency, defined as squared Wasserstein guidance response divided by worst-case squared Wasserstein perturbation risk:

\frac{(\int_{t_0}^T a_th_t\,dt)^2}{\int_{t_0}^T h_t^2/\sigma_t^2\,dt}, \quad a_t=\sqrt{\bar\alpha_t},\quad\sigma_t^2=1-\bar\alpha_t.$$ 3. **Schedule.** Cauchy-Schwarz gives the maximizing allocation $h_t\propto a_t\sigma_t^2$, yielding $$\omega(t)=1+C(\bar\omega-1)\frac{\sqrt{\bar\alpha_t}(1-\bar\alpha_t)}{\beta(t)},$$ where $C$ enforces the budget. This is DG-CFG (original Eq. 18; revised Eq. 22). The result maximizes the stated lower bound in the two-component model and motivates the practical schedule. We retain the three-factor explanations as intuitive interpretations. **Revision locations:** Secs. 5.1,5.3, pages 5-6; Theorem 5.1 and Eqs. (21),(22), page 6; proof in Appendix B.3, pages 21-25. Revisions are highlighted in blue. ## Q3. Signal-content weighting in the ablations Thank you for your careful examination and this important observation. **We revised the discussion to explain the Medium 2 degradation and the joint role of signal weighting and low-noise attenuation.** 1. **Why Medium 2 worsens.** Multiplying by $a_t$ shifts normalized guidance toward signal-rich, low-noise stages, where prediction errors are amplified through $\delta s_t=-\delta\epsilon_t/\sigma_t$. Without low-noise attenuation, this can worsen FID and IR. 2. **The factors should act jointly.** Constant CFG vs. Medium 4 and Medium 1 vs. DG-CFG show that jointly adding both factors improves CLIP and IR on both backbones under strong guidance, with modest FID changes and controlled saturation. Their combination balances signal content and error sensitivity, consistent with Theorem 5.1. 3. **How the signal factor helps DG-CFG.** Medium 3 vs. DG-CFG isolates adding $a_t$ to an existing low-noise attenuation schedule: $h_t\propto\sigma_t^2\rightarrow a_t\sigma_t^2$. Signal weighting restores emphasis on informative stages while the noise factor controls low-noise sensitivity, improving FID and IR on both SD1.5 and SD2.1. **Revision locations:** Sec. 6.2.5, pages 9-10; Table 4, page 10 (original Table 3 in review); Appendix D.5, pages 35-37. Revisions are highlighted in blue.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.