- 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
-
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.
-
Current experiments support that PMC-CFG improves the generation.
Critical weaknesses
-
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.
-
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
-
Please see weaknesses.
-
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?
-
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
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 , plus APG and PMC-CFG at .
The following reported results illustrate the comparisons:
| Method | Nominal | CLIP ↑ | ImageReward ↑ | FID ↓ | Recall@3 ↑ | Within-prompt CLIP diversity |
|---|---|---|---|---|---|---|
| C2FG | 3 | 0.3174 | 0.6715 | 26.97 | 0.4394 | 0.134 |
| PMC-CFG | 7 | 0.3174 | 0.8033 | 25.67 | 0.4540 | 0.142 |
| APG | 9 | 0.3172 | 0.8097 | 25.88 | 0.4472 | 0.137 |
| PMC-CFG | 9 | 0.3178 | 0.8262 | 25.61 | 0.4512 | 0.142 |
| APG | 11 | 0.3180 | 0.8639 | 26.33 | 0.4398 | 0.134 |
| PMC-CFG | 11 | 0.3183 | 0.8333 | 25.47 | 0.4600 | 0.142 |
At the same reported CLIP of 0.3174, PMC-CFG at has lower FID and higher coverage, within-prompt feature variation, and ImageReward than C2FG at . 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 . These cross-scale comparisons directly support improved quality and diversity at matched or similar measured alignment.
The additional 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 adds a non-dominated point to the empirical FID–ImageReward frontier and dominates APG at 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 specifies the requested guidance strength, while 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 at 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, favors FID and recall; at , 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 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 . 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 and . 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 as SD3.5. At external , 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 , 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 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.
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.