acceptodds
Withdrawn from ICLR 2027

CLASSIFIER-FREE GUIDANCE IN FLOW MATCHING: NON-AUTONOMOUS POTENTIALS, OVERSHOOT, AND POSTERIOR-MEAN CONTROL

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Classifier-free guidance (CFG) improves conditional generation in Flow Matching, but strong guidance can distort the generated distribution and reduce diversity. We provide a geometric account of this behavior by viewing Flow Matching as a time-varying gradient flow and characterizing how CFG reshapes its underlying potential. This view explains how stronger alignment can be accompanied by mean displacement and trajectory concentration, and motivates controlling guidance through the model-implied terminal posterior mean. We therefore propose PosteriorMean-Capped CFG (PMC-CFG), a training-free, per-sample method that adaptively retains the strongest feasible guidance without additional network evaluations. Experiments on synthetic and large-scale image-generation benchmarks show that PMC-CFG limits guidance-induced distortion and concentration while improving the alignment–diversity trade-off, with particularly strong benefits when nominal guidance is large.

Primary area:
generative models
Keywords:
flow matching, non-autonomous potentials

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_SssfOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper analyzes classifier-free guidance in linear flow matching through two perspectives: a time-varying potential and the posterior mean. The analysis expresses the conditional–unconditional velocity difference as a condition-posterior force and relates its Jacobian to posterior covariance. From this, the authors argue that strong fixed guidance may improve local conditional alignment and cause displacement or local volume contraction. Then they propose Posterior-Mean-Capped CFG (PMC-CFG), a training-free controller that preserves the original CFG direction but scales it down per sample whenever the model-implied terminal posterior mean exceeds a relative norm bound. Extensive experiments confirm PMC-CFG's effectiveness.

Critical strengths
  1. The paper offers a coherent interpretation of CFG in posterior-mean space and connects the analysis to a simple, state-dependent controller. PMC-CFG is inexpensive and straightforward to implement.

  2. Current experiments support that PMC-CFG improves the generation.

Critical weaknesses
  1. One concern is that the theoretical analysis does not uniquely motivate the proposed constraint. The potential representation and the posterior-force identity follow almost directly from the linear-path Tweedie identity and Bayes' rule. The covariance analysis shows how local contraction can arise, but it does not imply that the appropriate remedy is to bound the absolute norm of the model-implied terminal mean. That bound also depends on where the origin of the latent coordinate system lies, and no general link to data density, distance from the data manifold, or perceptual distortion is established. Thus, the controller remains a plausible heuristic rather than a consequence of the preceding theory.

  2. The main results support robustness to excessive nominal guidance more clearly than a general improvement of the alignment–diversity trade-off. PMC-CFG changes the effective guidance scale, so comparisons at the same nominal scale do not necessarily compare methods at the same achieved alignment.

Questions for authors
  1. Please see weaknesses.

  2. Can the authors clarify whether PMC-CFG improves the Pareto frontier when methods are compared at similar achieved alignment, rather than only at the same nominal guidance scale?

  3. How were the values of hyperparameters selected for ImageNet and SD3.5? What if deploying PMC-CFG to other datasets or base models?

Currently my rating leans to positive. However, I'd like to defer the final judgement after reviewing the rebuttal and other reviewers' comments.

3 comments

Thank you for your constructive feedback and for recognizing the posterior-mean interpretation and the simplicity of the controller. We address your questions about the constraint, coordinate dependence, comparisons at similar achieved alignment, and hyperparameter selection below.

Q1. Theoretical motivation and justification of the norm constraint

We agree that the potential and covariance identities do not uniquely determine a norm constraint. They motivate controlling guidance-induced extrapolation; the relative norm ball is our specific design choice. We have clarified this distinction and added analysis of what this choice guarantees. Section 3.4 (pp. 5–6) shows that condition confidence can increase outward even as data density decreases. Equation (11) in Section 3.2 (p. 4) expresses the posterior force as extrapolation of the implied terminal mean along the conditional–unconditional difference, connecting this failure mechanism to the quantity we control.

We use the conditional posterior mean, the minimum mean-square-error terminal estimate under exact fields, as the reference. Its norm sets a dimensionless amplification budget. The constraint is most restrictive for outward-aligned extrapolation and permits larger displacements in other directions; it therefore targets norm amplification rather than imposing the same displacement budget in every direction. PMC-CFG retains the largest feasible guidance coefficient along the original CFG direction in closed form, using the two standard network evaluations. Section 4.1 (pp. 6–7) explains these design choices.

We have also added direct analysis of the chosen constraint. Proposition 4.2 in Section 4.2 (p. 7) bounds displacement from the conditional prediction:

\leq\sqrt{z^2+\Gamma^2-1}-z, \qquad z=\frac{\langle m_t^{\mathrm c},\Delta_t\rangle} {\|m_t^{\mathrm c}\|\|\Delta_t\|},$$ for nonzero $m_t^{\mathrm c}$ and $\Delta_t$. In the aligned outward case, this reduces to a relative displacement bound of $\Gamma-1$. The proposition also shows that the largest feasible coefficient maximizes the additional instantaneous posterior-growth rate within the budget. Theorem 4.3 in Section 4.3 (p. 7) connects this control to endpoint density under explicit assumptions: exact fields of an equally weighted, equally spaced circular GMM, a single conditioned component, and initialization on its outward ray. The PMC-CFG trajectory lies between conditional sampling and fixed CFG, and its endpoint has conditional density at least as large as the fixed-CFG endpoint. Section 4.3 (pp. 7–8) also shows how feedback reduces local contraction at the mixture center. Proofs are provided in Appendices D.2, D.4, and D.6 (pp. 26, 28–30). For learned image models, the evidence is empirical: Tables 2–3 (p. 9) evaluate distributional quality and perceptual preference through FID and ImageReward. **Q2. Dependence on the latent coordinate system** Thank you for raising this question. We evaluate the constraint in each base model's native latent coordinates and normalization, as described in Section 4.1 (pp. 6–7) and Appendices E.2–E.3 (pp. 31–33). Proposition D.2 in Appendix D.3 (p. 27) characterizes its coordinate behavior: a common positive scalar rescaling or orthogonal transformation preserves the feasible guidance coefficients. Translating the origin or applying an anisotropic scaling can change them. That proposition also describes the anchored form $$\|m_t^{\mathrm{cap}}-r_t\| \leq\Gamma\|m_t^{\mathrm c}-r_t\|.$$ Translating the anchor and both means together leaves both sides unchanged. This form makes the reference origin explicit and provides a translation-covariant extension. The reported experiments use the base models' native latent reference; the anchored extension has not been evaluated here. Thus the current implementation remains coordinate dependent, and our experiments establish its behavior in those native coordinates rather than under arbitrary recentering.

Q3. Comparisons at similar achieved alignment and the empirical Pareto trade-off

Thank you for this suggestion. We agree that comparisons at the same nominal scale alone do not establish a better alignment–diversity trade-off. Appendix F.4 (pp. 37–38) therefore compares methods across scales at similar achieved CLIP, with ImageReward as a complementary preference measure. Figure 7 (p. 38) displays all 22 method–scale evaluations in Table 3 (p. 9): five methods at λ∈{3,5,7,9}\lambda\in\{3,5,7,9\}, plus APG and PMC-CFG at λ=11\lambda=11.

The following reported results illustrate the comparisons:

MethodNominal λ\lambdaCLIP ↑ImageReward ↑FID ↓Recall@3 ↑Within-prompt CLIP diversity
C2FG30.31740.671526.970.43940.134
PMC-CFG70.31740.803325.670.45400.142
APG90.31720.809725.880.44720.137
PMC-CFG90.31780.826225.610.45120.142
APG110.31800.863926.330.43980.134
PMC-CFG110.31830.833325.470.46000.142

At the same reported CLIP of 0.3174, PMC-CFG at λ=7\lambda=7 has lower FID and higher coverage, within-prompt feature variation, and ImageReward than C2FG at λ=3\lambda=3. At similar ImageReward (0.8033 vs. 0.8097) and CLIP (0.3174 vs. 0.3172), it also has lower FID, higher recall, and greater within-prompt variation than APG at λ=9\lambda=9. These cross-scale comparisons directly support improved quality and diversity at matched or similar measured alignment.

The additional λ=11\lambda=11 comparison makes the remaining trade-off explicit. At similar CLIP (0.3183 vs. 0.3180), PMC-CFG has lower FID and higher recall and within-prompt diversity than APG, while APG has higher ImageReward (0.8639 vs. 0.8333). In Figure 7(b), PMC-CFG at λ=11\lambda=11 adds a non-dominated point to the empirical FID–ImageReward frontier and dominates APG at λ=9\lambda=9 on those two metrics; APG retains both lower-FID and higher-reward frontier points. Thus PMC-CFG improves part of the observed frontier, without dominating it throughout. This quality–preference frontier complements the matched-CLIP diversity comparisons above. Our conclusion is limited to the evaluated operating points, not all possible guidance scales. Section 5.1 (p. 8) introduces the diversity metric, and Appendix E.3 (pp. 32–33) defines its protocol.

Q4. Hyperparameter selection and deployment on other datasets or base models

Thank you for asking us to clarify parameter selection. The nominal scale λ\lambda specifies the requested guidance strength, while Γ=1+ϵ\Gamma=1+\epsilon controls the permitted relative increase in posterior-mean norm. We select the reported caps using the sensitivity sweeps in Appendices F.2–F.3 (pp. 35–37). These results show useful performance across several cap values, with tighter caps favoring quality and coverage and looser caps retaining more conditioning.

  • SiT/ImageNet-256: We use Γ=1.10\Gamma=1.10 at λ=1.5\lambda=1.5 to balance conditioning benefit and coverage: at 20 NFE, this reduces FID from 3.58 to 3.03, increases IS from 239.49 to 261.95, and preserves recall at 0.57. At stronger guidance, Γ=1.05\Gamma=1.05 favors FID and recall; at λ=2.5\lambda=2.5, FID improves from 6.22 to 3.82 and recall from 0.43 to 0.51 relative to fixed CFG. Appendix F.2 (pp. 35–36) reports the full cap sweep.
  • SD3.5: We use Γ=1.15\Gamma=1.15 for its balance between conditioning and distributional quality. Appendix F.3 (pp. 36–37) shows that both tested caps, 1.15 and 1.20, maintain stable FID through λ=11\lambda=11. The looser cap gives higher CLIP and ImageReward, while the tighter cap gives lower FID and higher recall throughout the tested sweep.

For a new dataset or backbone, we suggest starting from its usual nominal CFG scale and evaluating a small set of caps close to one, choosing according to the observed quality, coverage, and alignment. Our sweeps provide practical starting values: 1.05, 1.10, and 1.15 on ImageNet, and 1.15 and 1.20 on SD3.5. All three ImageNet caps improve FID over fixed CFG at both λ=1.5\lambda=1.5 and λ=2.5\lambda=2.5. The sensitivity results show useful sampling performance across these cap values, with tighter caps favoring distributional quality and coverage and looser caps retaining more conditioning.

As an additional backbone example, the FLUX.1-dev experiments in Appendix E.4 use the same Γ=1.15\Gamma=1.15 as SD3.5. At external λ=3\lambda=3, PMC-CFG gives FID 26.50 and ImageReward 0.8447, compared with 30.80 and 0.5419 for fixed CFG. This supports using the same cap as a practical starting point on another backbone, without establishing a universally optimal value; at λ=2\lambda=2, PMC-CFG has higher FID than fixed CFG. The opening “Parameter selection” paragraph in Appendix E (p. 31) explains the selection rationale, and Appendices F.2–F.3 (pp. 35–37) give the model-specific sensitivity results.

Summary of revisions

We have clarified the theory-to-method connection and its scope in Sections 3.2, 3.4, and 4.1 (pp. 3–7), with direct displacement and GMM trajectory/volume analysis in Sections 4.2–4.3 (pp. 7–8). Appendix D.3 (pp. 27–28) characterizes coordinate behavior and distinguishes the implemented controller from the anchored extension. Appendix F.4 (pp. 37–38) compares methods at similar achieved alignment or preference, including the additional λ=11\lambda=11 results and their remaining trade-offs. The opening paragraph of Appendix E (p. 31) and Appendices F.2–F.3 (pp. 35–37) explain the cap choices and sensitivity across the tested ranges.

We appreciate the time and thought you have devoted to reviewing our work. Your comments have helped us explain the controller's rationale, the comparisons at similar achieved alignment, and the parameter choices more clearly. We hope the revised manuscript and the responses above address your questions, and thank you again for your constructive suggestions.

Reviewer_hMN2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

In Classifier-freeguidance, strong guidance can distort the generated distribution and reduce diversity. This paper propose Posterior-Mean Capped CFG(PMC-CFG), a training-free, per-sample controller that imposes a pointwise relative bound on the implied terminal meanwhile retaining the strongest feasible guidance, without extra network evaluations. This paper is not well-written, there is no strong connection between theoretical analysis and model formulation. We can not get the basic idea of this paper.

Critical strengths
  1. Provide some thoeretical analysis, showing that the same posterior force that provably increases the local conditional-alignment rate can also accumulate mean displacement and contract trajectory volume, thereby separating why CFG works from why fixed strong CFG fails.
  2. introduce PMC-CFG.
  3. evaluate the predicted mechanisms numerically.
Critical weaknesses
  1. This paper is not well-written, there is no strong connection between theoretical analysis and model formulation. It is better to provide the theoreical analysis of ∣∣mct+βΔt∣∣≤Γ∣∣mct∣∣||m^t_c+\beta\Delta t||\le\Gamma||m^t_c||.
  2. Lack of highlight in Figure 2.
  3. For the numerical experiments, which metric is to measure the diversity? Is it recall? If so, the experimental results can not supprot your claim about reducing diversity.
Questions for authors
  1. Please provide the connection between your theoretical results and model formulation.
  2. It seem it is not useful of the theoretical analysis in Subsection 3.4. Could you give some explationations?
  3. Why to introduce ∣∣mct+βΔt∣∣≤Γ∣∣mct∣∣||m^t_c+\beta\Delta t||\le\Gamma||m^t_c||, your theory is not for this part. The authors also need to claim.
2 comments

Thank you for your careful reading and constructive feedback. We have clarified the theory–method connection, added analysis of the constraint, annotated Figure 2, and expanded the diversity evaluation. We address each concern below.

Q1. Connection between the theoretical analysis and the method formulation

Thank you for highlighting this gap in our explanation. The connection is that the posterior force analyzed in Section 3 induces the terminal-mean extrapolation controlled in Section 4.

Equation (11) in Section 3.2 (p. 4), together with the mean definitions in Section 4.1 (pp. 6–7), gives, for exact fields under the independent linear Gaussian path and 0<t<10<t<1,

=\frac{(1-t)^2}{t}\nabla_x\log p(c\mid X_t=x), \qquad m_t^\lambda=m_t^{\mathrm c}+(\lambda-1)\Delta_t.$$ Thus, increasing guidance extrapolates the terminal mean along the posterior-gradient direction. PMC-CFG bounds this extrapolation through a relative norm constraint while retaining the direction and the largest feasible guidance coefficient. With learned fields, the means are model-implied estimates computed from the two standard network evaluations. We have made this connection explicit in Sections 3.2 and 4.1, and analyzed the controller in Sections 4.2–4.3 (pp. 7–8). **Q2. Role of the analysis in Section 3.4** Thank you for asking us to clarify its purpose. Section 3.4 (pp. 5–6) shows why higher condition confidence alone does not ensure movement toward data-dense regions. In the circular GMM, confidence approaches one along an outward ray while data density approaches zero, and the posterior force adds an outward velocity component. In the centered GMM, this mechanism is already visible at initialization: the conditional terminal mean is $\mu_0$, while fixed CFG extrapolates it to $\lambda\mu_0$. This identifies the outward extrapolation that motivates the constraint in Section 4.1. **Q3. Motivation and theoretical analysis of the relative norm constraint** Thank you for pointing out the need to justify and analyze the constraint. For exact fields, the conditional posterior mean is the minimum mean-square-error terminal estimate given the current state and condition. We use this prediction as the reference and impose $$\|m_t^{\mathrm c}+\beta\Delta_t\| \leq\Gamma\|m_t^{\mathrm c}\|, \qquad 0\leq\beta\leq\lambda-1.$$ The relative bound makes $\Gamma$ dimensionless and scales the budget with the conditional prediction. It is most restrictive for outward-pointing gaps, targeting the amplification identified in Section 3.4. PMC-CFG selects the largest feasible coefficient, preserving nominal CFG whenever it satisfies the constraint. We have added direct analysis of this constraint. Proposition 4.2 in Section 4.2 (p. 7) defines $$R=\frac{\|m_t^{\mathrm{cap}}-m_t^{\mathrm c}\|}{\|m_t^{\mathrm c}\|}, \qquad z=\frac{\langle m_t^{\mathrm c},\Delta_t\rangle} {\|m_t^{\mathrm c}\|\|\Delta_t\|},$$ for nonzero $m_t^{\mathrm c}$ and $\Delta_t$, and establishes $$R\leq\sqrt{z^2+\Gamma^2-1}-z.$$ For aligned outward extrapolation, $z=1$, this gives $R\leq\Gamma-1$; other directions follow the angle-dependent bound above. For exact fields, Proposition 4.2 also shows that the largest feasible coefficient maximizes the additional instantaneous posterior-growth rate at the current time and state within this budget. Theorem 4.3 (pp. 7–8) further establishes a trajectory-level result for exact circular-GMM fields with equally weighted, equally spaced components: when conditioning on one component and starting on its outward ray, the PMC-CFG trajectory lies between the conditional and fixed-CFG trajectories started at the same point. Its endpoint has at least as high a target conditional density as the fixed-CFG endpoint. Section 4.3 also establishes reduced local contraction at the mixture center for $1<\Gamma<\lambda$. Proofs are in Appendices D.2, D.4, and D.6 (pp. 26, 28–30). **Q4. Highlighting the main message of Figure 2** Thank you for this suggestion. Figure 2 (p. 6) now includes radial arrows, dashed rays marking the conditioned sector, explicit time labels, and a shared posterior colorbar. The annotations highlight that the posterior increases outward along the middle ray at fixed time, while the near-one region extends inward across the displayed times. The revised caption and Section 3.4 explain these observations and their connection to the terminal-mean gap.

Q5. Diversity metrics and support for the empirical claims Thank you for raising this point. Recall measures coverage of the real-data distribution. The revised evaluation also measures within-condition diversity directly, as distinguished in Section 5.1 (p. 8):

  • GMM: Table 1 (p. 8) reports within-branch variance and branch occupancy. For the singleton condition, PMC-CFG raises the variance ratio from fixed CFG's 0.135 to 0.348. For three equally weighted conditioned components, it reduces the central branch's occupancy from 68.63% to 41.38%, toward the target 33.33%. These results directly show reduced within-branch contraction and mode concentration.
  • ImageNet-256: Table 2 (p. 9) reports classwise mean Vendi Score using DINOv2 features, with 10 images per class across 1,000 classes. At 20 NFE, fixed CFG's Vendi decreases from 5.4326 to 4.7291 to 4.2867 as λ\lambda increases from 1.5 to 2 to 2.5. At λ=2.5\lambda=2.5, PMC-CFG raises Vendi to 4.6128, improves FID from 6.22 to 3.82, and increases recall from 0.43 to 0.51. This supports reduced diversity loss under strong guidance. Appendix E.2 (pp. 31–32) gives the protocol.
  • SD3.5: Table 3 (p. 9) reports within-prompt CLIP diversity: the mean pairwise cosine distance between normalized image embeddings, averaged over 100 prompts with 10 images each (Appendix E.3, pp. 32–33). Excessively high fixed CFG causes severe image-quality degradation: increasing λ\lambda from 5 to 9 worsens FID from 29.20 to 40.83 and lowers ImageReward from 0.7101 to −0.0295, while CLIP diversity rises from 0.131 to 0.181. In this degraded regime, feature distances can reflect uncontrolled variation in image content, making the diversity increase difficult to interpret. PMC-CFG preserves both image quality and within-prompt diversity as guidance increases: across λ∈{5,7,9}\lambda\in\{5,7,9\}, FID stays at 25.65, 25.67, and 25.61, ImageReward rises from 0.7538 to 0.8262, and CLIP diversity remains nearly unchanged at 0.143, 0.142, and 0.142. At the same reported text–image CLIP of 0.3174, PMC-CFG at λ=7\lambda=7 also gives greater within-prompt diversity than C2FG at λ=3\lambda=3 (0.142 vs. 0.134), lower FID (25.67 vs. 26.97), and higher ImageReward (0.8033 vs. 0.6715). Appendix F.4 (pp. 37–38) discusses these comparisons at similar measured alignment or preference.

Thank you again for the time and care you devoted to reviewing our work. Your comments helped us recognize where our explanation was incomplete and prompted us to strengthen the connection between the theory and the method, analyze the constraint directly, and clarify the diversity evidence. We appreciate the opportunity to address these concerns and hope the revised manuscript and responses make our reasoning clearer.

Reviewer_qcG7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies classifier-free guidance (CFG) in linear Flow Matching using non-autonomous potential and posterior-mean analyses. It introduces Posterior-Mean-Capped CFG (PMC-CFG), a training-free, per-sample controller that limits the relative norm of the implied terminal mean while retaining the largest feasible guidance coefficient. The coefficient admits a quadratic closed-form solution with no extra network evaluations. Experiments cover analytic Gaussian mixtures, ImageNet-256 (SiT-XL/2), and Stable Diffusion 3.5 Medium.

Critical strengths
  1. The paper connects CFG extrapolation, endpoint displacement, and local volume change through flow geometry. Its per-sample control rule has a closed-form solution and requires neither retraining nor additional network evaluations.
  2. The gains at guidance are substantial in some settings. At 20 NFE and λ = 2.5, ImageNet FID improves from 6.22 to 3.82 and recall from 0.43 to 0.51. On SD3.5 at λ = 9, FID falls from 40.83 to 25.61.
Critical weaknesses
  1. Theorem 4.1 proves feasibility of the latent-norm cap, not that it limits displacement from the conditional prediction or preserves high-density regions.
  2. The extra cap Γ must be hand-tuned jointly with λ, Γ = 1.10 for ImageNet at λ = 1.5, Γ = 1.05 at stronger scales, and Γ = 1.15 for SD3.5. Appendix F.2–F.3 shows meaningful trade-offs but no transferable parameter-selection principle or comparison of tuning costs.
  3. The novelty relative to recent adaptive CFG work remains unclear. Direct matched-budget comparisons are needed with CFG-Ctrl (CVPR 2026), Dynamic Classifier-Free Diffusion Guidance via Online Feedback (ICLR 2026), and Navigating with Annealing Guidance Scale in Diffusion Space (SIGGRAPH Asia 2025), rather than relying mainly on fixed CFG and APG/C²FG.
Questions for authors
  1. Theorem 4.1 establishes the feasibility of the proposed constraint and identifies the maximal feasible guidance coefficient, but does not demonstrate how the constraint relates to the quality of the final generated distribution. Please provide further theoretical justification for this constraint and clarify under what conditions it can effectively mitigate overshoot and distributional shrinkage.
  2. PMC-CFG still requires manual selection of the cap parameter Γ, with different values used across models and nominal guidance scales. Despite the sensitivity analyses provided, a unified principle for selecting this parameter and an evaluation of its cross-model generalizability are still lacking.
  3. The authors compare PMC-CFG with fixed CFG at different guidance scales, APG, and C2FG. However, comparisons with recent advanced methods based on dynamic guidance scheduling and feedback control remain insufficient. Please further validate the advantages of PMC-CFG under matched models, sampling budgets, and comparable hyperparameter-tuning conditions.
  4. Appendix Table 4 shows that PMC-CFG achieves only marginal FID improvements of 0.03–0.04 over PMC-Mean in several experimental settings. Please further analyze the independent contribution of per-sample feedback beyond the time-only guidance schedule and assess the reliability of these small improvements through repeated experiments or confidence intervals.
  5. Existing studies, such as CFG-Ctrl and Adaptive Diffusion Guidance via Stochastic Optimal Control, have already explored feedback-based or state-dependent guidance control. The authors should further clarify the distinctive contributions of PMC-CFG and provide direct comparisons with these closely related adaptive guidance methods.
  6. The current image-generation experiments primarily focus on SiT-XL/2 and SD3.5 Medium, which may be insufficient to establish the generalizability of PMC-CFG across different generative model architectures. Please provide additional quantitative and qualitative evaluations on more representative generative models.
  7. Figure 4 mainly demonstrates that PMC-CFG better preserves the unguided composition, but composition similarity does not necessarily imply higher generation quality. Please provide additional qualitative comparisons with APG, C2FG, and other advanced adaptive guidance methods and random seeds, together with quantitative evaluations.
  8. On pp. 11–12, please correct “Imagenet” to “ImageNet,” “gans” to “GANs,” “Ieee” to “IEEE,” and “PmLR” to “PMLR.”
4 comments

Thank you for your constructive feedback. Your comments helped us strengthen the theoretical justification, expand the experimental comparisons, and clarify the contribution of per-sample feedback. We address your concerns point by point below.

Q1. Theoretical justification of the constraint and its effects on overshoot and distributional shrinkage

Thank you for suggesting further analysis of the constraint. We have added three results in Sections 4.2–4.3 (pp. 7–8) that connect the norm cap to displacement, endpoint density, and local volume change.

  1. Displacement control. Proposition 4.2 (p. 7) relates the norm cap directly to displacement from the conditional prediction. For nonzero mtcm_t^{\mathrm c} and Δt\Delta_t, let z=cos⁡∠(mtc,Δt)z=\cos\angle(m_t^{\mathrm c},\Delta_t). If z≥κ≥0z\geq\kappa\geq0, then

    {\|m_t^{\mathrm c}\|} \leq\sqrt{\kappa^2+\Gamma^2-1}-\kappa.$$ For aligned outward extrapolation, this bound becomes $\Gamma-1$. Thus, under the stated alignment condition, the norm budget bounds displacement from the conditional prediction.
  2. Overshoot and endpoint density. Theorem 4.3 (p. 7) considers exact FM fields for an equally weighted, equally spaced circular GMM with at least three components, conditioned on one component. From a common initial point x0=s0u0x_0=s_0u_0, s0≥0s_0\geq0, with λ,Γ≥1\lambda,\Gamma\geq1, the radial trajectories satisfy

    \qquad 0\leq t\leq1.$$ Consequently, PMC-CFG has no greater outward displacement from the matched conditional trajectory than fixed CFG, and its endpoint satisfies $$p_1(x_1^{\mathrm{cap}}\mid c=0) \geq p_1(x_1^\lambda\mid c=0).$$
  3. Local contraction. For the same GMM, at the mixture center and 1<Γ<λ1<\Gamma<\lambda, Eq. (34) (p. 8) establishes

    \qquad 0<t<1.$$ At this point, feedback compensates the extra radial contraction and reduces the extra tangential contraction coefficient from $\lambda-1$ to $\Gamma-1$.

Complete proofs appear in Appendices D.2, D.4, and D.6 (pp. 26, 28–30). Together, these results explain how the constraint controls displacement and mitigates overshoot and local contraction under the stated geometric conditions. For learned image-generation fields, the pointwise bound applies directly to the model-implied terminal mean, and the image experiments evaluate its effects on quality and coverage.

Q2. Selection of the cap parameter, transferability, and calibration cost

Thank you for asking us to clarify how to select Γ\Gamma. The opening “Parameter selection” paragraph in Appendix E (p. 31) now explains its interpretation as Γ=1+ϵ\Gamma=1+\epsilon, where ϵ≥0\epsilon\geq0 specifies the permitted relative increase in posterior-mean norm. In the aligned outward case, it also bounds relative displacement from the conditional prediction. This gives the parameter a common geometric interpretation across models.

The sensitivity experiments in Appendices F.2–F.3 (pp. 35–37) identify useful operating ranges. On ImageNet, the tested values {1.05,1.10,1.15}\{1.05,1.10,1.15\} shift the balance between FID/recall and IS/precision. On SD3.5, both Γ=1.15\Gamma=1.15 and 1.201.20 maintain stable distributional quality through λ=11\lambda=11, with the former favoring FID and coverage and the latter favoring CLIP and ImageReward.

A practical procedure suggested by these sweeps is to start from the base model's usual nominal guidance scale, compare a small set of cap values near one, and select according to the desired balance between conditioning and coverage. We used the reported sensitivity sweeps to select the caps in our experiments. The geometric interpretation is shared across models, while the preferred value depends on the target trade-off.

We also evaluated the same cap, Γ=1.15\Gamma=1.15, on the additional FLUX.1-dev backbone (Appendix E.4, Table 4). At external scale 3, PMC-CFG improves FID from 30.80 to 26.50 and ImageReward from 0.5419 to 0.8447 relative to fixed CFG. This extends the evidence for using the same cap across SD3.5 and FLUX.1-dev. Q6 discusses the full results across scales.

Calibration costs consist mainly of generating and evaluating samples for the candidate settings. Within each setting, PMC-CFG reuses CFG's conditional and unconditional predictions, adding only norms, an inner product, and a closed-form coefficient calculation, with no additional network evaluations. The revised appendix documents the tested cap sets and evaluated configurations. PMC-CFG introduces only one additional tunable parameter, Γ\Gamma, beyond the nominal guidance scale.

Q3. Comparisons with recent adaptive guidance methods under matched conditions

Thank you for this suggestion. We have added direct comparisons with CFG-Ctrl on SiT-XL/2 and SD3.5 Medium using matched checkpoints, initial noise, solvers, and sampling budgets. ImageNet evaluations use 50,000 samples at 20 and 50 NFE; SD3.5 evaluations use 5,000 COCO prompts at 10 NFE (Tables 2–3, p. 9). The new FLUX.1-dev benchmark also compares fixed CFG, APG, C2FG, CFG-Ctrl, and PMC-CFG on 5,000 COCO prompts at 10 NFE (Appendix E.4, Table 4).

Appendix E specifies the evaluated configurations and parameter-selection procedures. CFG-Ctrl uses (λctrl,k)=(5,0.003)(\lambda_{\mathrm{ctrl}},k)=(5,0.003) on SiT-XL/2, selected through a small-scale pilot evaluation, and (5,0.2)(5,0.2) on SD3.5 and FLUX.1-dev. Here λctrl\lambda_{\mathrm{ctrl}} is a controller parameter distinct from the nominal CFG scale. All five methods share the common nominal-scale sweep within each benchmark; additional evaluations of APG and PMC-CFG at scale 11 on SD3.5 and scale 4 on FLUX are identified separately. The experiments match sampling budgets.

PMC-CFG achieves the lowest FID among the evaluated configurations at both ImageNet sampling budgets. On SD3.5, APG and PMC-CFG offer complementary quality–alignment trade-offs, while CFG-Ctrl achieves higher CLIP at the larger scales. On FLUX at scale 3, PMC-CFG has lower FID (26.50 versus 29.41) and higher ImageReward (0.8447 versus 0.7642) than CFG-Ctrl, with slightly lower CLIP (0.3151 versus 0.3164).

The other two suggested methods involve learned components: Dynamic CFG via Online Feedback uses trained latent-space evaluators, and Annealing Guidance learns a scheduling policy, including a Flow Matching extension. Evaluating them on our backbones requires adapting and validating these components. Given the time constraints of the revision, we have focused on direct CFG-Ctrl comparisons across the three backbones and leave these additional evaluations for future work.

Q4. Independent contribution of per-sample feedback and reliability of small improvements

Thank you for asking us to distinguish these contributions more clearly. We have expanded Appendix F.1 (pp. 34–35) to explain the role of the average temporal profile and added an SD3.5 ablation in Table 6 (p. 35), complementing the ImageNet ablation in Table 5 (p. 35).

PMC-Mean averages PMC-CFG's effective guidance scales over calibration samples and freezes the resulting curve. It retains the controller's average temporal profile. Its strong performance shows that this profile already accounts for much of the improvement: PMC-Mean lowers FID relative to fixed CFG in all six ImageNet settings. Per-sample feedback additionally adapts to each sample's current posterior-mean geometry and enforces the cap pointwise; a shared time-only scale need not satisfy that constraint for every sample.

On ImageNet, the reported PMC-CFG point estimates have lower FID in five of six settings, with a tie after rounding in the remaining setting, and higher IS in all six. The magnitude depends on the operating point. At 20 NFE and nominal scale 1.5, FID changes from 3.23 to 3.03 and IS from 252.97 to 261.95. The 0.03–0.04 FID differences at other settings are small observed differences over an already effective schedule.

The new SD3.5 ablation uses 10 NFE and nominal scale 9. PMC-Mean is constructed from 32 calibration samples and evaluated on 5,000 COCO prompts. It reduces fixed CFG's FID from 40.83 to 25.70. Per-sample PMC-CFG further improves FID to 25.61, ImageReward from 0.8142 to 0.8262, and Recall@3 from 0.4414 to 0.4512, at the same reported CLIP of 0.3178.

We agree that repeated runs or confidence intervals would help assess the reliability of the smallest differences. The current evaluations report point estimates from 50,000 images on ImageNet and 5,000 prompts on SD3.5; due to time constraints, repeated-run estimates and confidence intervals are left for further evaluation. We have clarified the interpretation in Appendix F.1: the average temporal profile accounts for much of the measured gain, while per-sample feedback enforces the cap for each sample and provides modest, setting-dependent improvements in the reported metrics.

Q5. Distinctive contribution relative to existing feedback and state-dependent guidance

Thank you for highlighting these closely related studies. Feedback-based and state-dependent guidance have been explored in prior work. CFG-Ctrl introduces a sliding-mode controller based on the conditional–unconditional discrepancy, while Adaptive Diffusion Guidance via Stochastic Optimal Control formulates guidance selection as a control objective and optimizes a guidance policy.

Our contribution is the specific posterior-mean constraint and its solution for linear Flow Matching. The posterior-force identity connects CFG to extrapolation of the implied terminal mean. PMC-CFG then selects the largest coefficient within the nominal guidance range that satisfies a relative norm bound on this mean. The solution is available in closed form for each sample, preserves the CFG direction, retains nominal CFG whenever it is feasible, and requires no auxiliary training or additional network evaluations. Sections 3.2, 3.4, and 4.1 (pp. 3–7) clarify this connection, and Sections 4.2–4.3 (pp. 7–8) add the displacement, endpoint-density, and local-volume analyses described in Q1.

We have also added quantitative comparisons with CFG-Ctrl on all three image-generation backbones (Tables 2–4) and qualitative comparisons on SiT-XL/2 and SD3.5 (Figure 4, p. 10), using the sampling conditions described in Q3. A direct comparison with the stochastic-optimal-control method would require optimizing its guidance policy for the evaluated setting. We appreciate this suggestion and leave that additional evaluation for future work.

Q6. Evaluation on more representative generative models

Thank you for suggesting a broader evaluation. We have added FLUX.1-dev, a guidance-distilled rectified-flow backbone, alongside SiT-XL/2 and SD3.5 Medium. Section 5.2 (p. 10) summarizes the findings, and Appendix E.4 (pp. 33–34) provides the protocol and full results in Table 4 (p. 34).

The benchmark uses 5,000 COCO 2017 validation prompts, 512×512 resolution, and 10 NFE. We compare fixed CFG, APG, C2FG, CFG-Ctrl, and PMC-CFG at external CFG scales 2 and 3, with an additional comparison between APG and PMC-CFG at scale 4. The model's embedded guidance input is held at 1, so the swept parameter is the external conditional–unconditional extrapolation scale. PMC-CFG uses Γ=1.15\Gamma=1.15, the same numerical cap as on SD3.5.

At external scale 3, PMC-CFG achieves the lowest FID (26.50) and highest ImageReward (0.8447) among the five evaluated methods. Relative to fixed CFG, FID decreases from 30.80 to 26.50, ImageReward increases from 0.5419 to 0.8447, and CLIP increases from 0.3132 to 0.3151. At scale 4, PMC-CFG obtains FID 26.14 and ImageReward 0.8933, compared with APG's 27.06 and 0.8538.

The full table also shows the dependence on guidance scale. Increasing the scale from 2 to 3 improves both FID and ImageReward for PMC-CFG, consistent with its role in controlling excessive extrapolation at stronger guidance.

This revision adds both quantitative and qualitative evidence on FLUX.1-dev, and it expands the qualitative comparisons on SiT-XL/2 and SD3.5 as detailed in Q7. Appendix F.8 (Figure 12) now includes matched FLUX.1-dev examples for a bedroom prompt and a cat-holding-a-sign prompt: eight samples per prompt, each row sharing the prompt and initial latent seed across no guidance, fixed CFG, and PMC-CFG at nominal scale 3. Fixed CFG renders the sign legibly but repaints the wall painting, alters the cat's markings, and drops the straw hat, at higher saturation (0.376 versus 0.273, Table 4). PMC-CFG keeps the unguided scene, sharpens it, and improves sign legibility.

Q7. Additional qualitative comparisons, random seeds, and quantitative diversity evaluation

Thank you for suggesting broader qualitative comparisons and quantitative evaluation. We have expanded Figure 4 (p. 10) to include APG, C2FG, and CFG-Ctrl alongside no guidance, fixed CFG, and PMC-CFG on both ImageNet-256 and SD3.5. Each row uses the same class or prompt and initial seed across methods, and guided methods use the same nominal scale within each benchmark. Appendix F.8 (pp. 40–43) provides additional matched examples across classes, prompt categories, and seeds for no guidance, fixed CFG, and PMC-CFG, including FLUX.1-dev examples for a bedroom prompt and a cat-holding-a-sign prompt (Figure 12).

We have also expanded the quantitative evaluation with measures of variation under a fixed condition:

  • ImageNet-256: Table 2 (p. 9) now reports classwise mean Vendi Score using DINOv2 features on a balanced 10,000-image subset. This complements the existing FID, sFID, IS, precision, and recall evaluations.
  • SD3.5: Table 3 (p. 9) now reports within-prompt CLIP diversity using 100 prompts and 10 seeds per prompt, totaling 1,000 images per method and guidance scale. We compute the average pairwise cosine distance between image embeddings within each prompt and then average over prompts.

Section 5.1 (p. 8) introduces these metrics, and Appendices E.2–E.3 (pp. 31–33) explain the evaluation protocols and distinguish within-condition variation from distributional coverage measured by recall. We interpret diversity jointly with FID, CLIP, and ImageReward, since greater feature variation can also accompany degraded samples. Appendix F.4 (pp. 37–38) further compares quality and diversity at similar achieved CLIP and ImageReward. For example, at the same reported CLIP of 0.3174, PMC-CFG at nominal scale 7 improves ImageReward from C2FG's 0.6715 at scale 3 to 0.8033, while also improving FID, recall, and within-prompt diversity. These evaluations complement the visual comparisons with measures of quality, alignment, and sample variation.

Q8. Capitalization corrections

Thank you for pointing out these errors. We have corrected “Imagenet,” “gans,” “Ieee,” and “PmLR” to “ImageNet,” “GANs,” “IEEE,” and “PMLR,” respectively, in the reference list (pp. 12–13).

Summary of revisions

The revision adds displacement control and GMM trajectory/volume analyses (Sections 4.2–4.3), explains cap selection (Appendix E), and expands the empirical evaluation to FLUX.1-dev with CFG-Ctrl comparisons across all three image-generation backbones (Tables 2–4). We have also added within-condition diversity metrics and comparisons at similar achieved alignment (Tables 2–3 and Appendix F.4), expanded the qualitative comparisons (Figure 4 and the FLUX.1-dev examples in Appendix F.8), extended the PMC-Mean ablation to SD3.5 and clarified the contribution of per-sample feedback (Appendix F.1), and corrected the reference capitalization.

We thank the reviewer again for the thoughtful comments and constructive suggestions, which have helped us improve the clarity and completeness of the paper.

Comments on the paper

1 comment

Dear Reviewers,

The authors have provided their rebuttal, but no follow-up comments have been posted yet. As the author-reviewer discussion period is ending soon, please review and acknowledge the rebuttal, and engage in further discussion if any concerns remain unresolved.

AC

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

Related papers

Discussion (0)

Sign in to comment.
CLASSIFIER-FREE GUIDANCE IN FLOW MATCHING: NON-AUTONOMOUS POTENTIALS, OVERSHOOT, AND POSTERIOR-MEAN CONTROL | acceptodds