- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies classifier-free guidance (CFG) in stochastic and deterministic diffusion samplers from the perspectives of inverse optimal control and probability density evolution. The authors derive control objectives under which prescribed CFG policies are optimal. They also characterize the terminal distributions of guided sampling and explain how initialization and sampling trajectories affect the final distribution. The theoretical analysis connects stochastic and deterministic sampling through a zero-noise limit. Based on these results, the paper introduces coefficient-budget coordinates for guidance schedule selection and Online-K for adaptive guidance. Experiments on Gaussian mixtures, CIFAR-10, and ImageNet validate the theoretical findings and demonstrate improvements in FID.
Critical strengths
-
A systematic theoretical analysis of CFG. The paper provides a unified perspective on CFG through inverse optimal control and terminal density evolution. The theoretical results clarify how guidance affects stochastic and deterministic sampling, offering useful insights into their different distributional behaviors.
-
Meaningful characterization of terminal distributions. The paper derives terminal density representations that account for initialization and accumulated guidance effects. These results help explain why the final sampling distribution can differ from the density suggested by local score extrapolation. The connection between stochastic and deterministic sampling further strengthens the theoretical contribution.
-
Practical extensions supported by experiments. The paper connects its theoretical analysis to guidance schedule selection and adaptive guidance through Online-K. The experiments include analytical Gaussian mixtures and image generation benchmarks. The reported results show that the proposed approaches can improve FID over the evaluated baselines.
Critical weaknesses
-
Limited novelty and practical significance of the inverse optimal control formulation. The paper constructs control objectives under which prescribed CFG policies become optimal. However, this approach appears closely related to standard inverse optimal control techniques, where the objective is designed to match a given policy. It remains unclear what new insights this formulation provides beyond existing theoretical analyses of CFG. In particular, the derived optimality does not directly establish that CFG is optimal for generation quality or provide a principled way to select guidance schedules. This limits the significance of the theoretical contribution.
-
Insufficient empirical validation of the proposed adaptive guidance mechanism. The theoretical analysis assumes exact scores and continuous-time dynamics, while Online-K relies on estimated conditional evidence and finite-step sampling. More importantly, the experiments do not clearly separate the effect of evidence-based feedback from changes in guidance strength and timing. Therefore, the reported FID improvements cannot be directly attributed to the proposed feedback mechanism. Additional ablations with matched guidance schedules and computational budgets would help establish its effectiveness.
Questions for authors
How does the proposed framework differ from existing time-varying CFG methods? Several recent studies, including -CFG (Classifier-free Guidance with Adaptive Scaling) and C$^2$FG (Control Classifier-Free Guidance via Score Discrepancy Analysis, CVPR 2026), have investigated adaptive guidance through different theoretical motivations and scheduling strategies. Could the authors further discuss the conceptual and theoretical differences between these methods and the proposed framework? In particular, it would be helpful to clarify what additional insights or practical advantages the inverse optimal control formulation and Online-K provide beyond existing time-dependent guidance approaches.
2 comments
In Author Response 3, the performance on the ImageNet dataset does not seem to demonstrate the method's advantage, and the comparisons are not conducted under the standard conditions for SDE-250 inference (FID-50K). Therefore, I prefer to maintain the score.
Response to Reviewer 8jvh.
Thank you for your constructive comments. We address your concerns below.
1. Relation to beta-CFG, C2FG and What the inverse-control characterization adds
Guidance signals and methodological scope.
β-CFG combines a beta-shaped temporal envelope with current conditional--unconditional prediction-gap normalization, yielding sample-dependent scales (Malarz et al., arXiv:2502.10574, Eq.17). C$^{2}FGusesscore−discrepancyboundstomotivateasharedexponentialtimeschedule(Gaoetal.,arXiv:2603.08155,Section 3.2).Online−K$ uses an accumulated estimate motivated by conditional density-ratio dynamics. They differ in scale-setting information and policy characterization, beyond time variation/sample dependence.
What the control and terminal-law characterization adds.
For a prescribed policy, we give an explicit compatible objective, its optimality gap and a sampler-dependent terminal law, exposing initialization and visited-state effects. Stochastic density corrections average over endpoint-conditioned reference paths; deterministic corrections follow a guided characteristic. Identical schedules can thus yield different terminal laws.
For a smooth time-only schedule, Ft=a(t)Kt gives Rtra=a˙K+(a2−a)E. This analyzes boundary/trajectory corrections for exponential or smooth beta-shaped envelopes. Timing and encountered score separation prevent coefficient budgets from determining terminal laws. For evidence response a=ϕK, the residual also contains ∂KaE, missed by substituting a varying scale into the constant-scale expression. These results use the stated exact-score assumptions; they neither claim the induced objective measures generation quality nor imply that β-CFG's full norm-dependent rule is always representable as a(t,Kt).
What distinguishes the Online-K feedback signal.
Under exact scores, Kt(x)=log[pt(x∣c)/pt(x)] is a state function whose gradient is the conditional--unconditional score difference; its norm does not determine K. Along a guided ODE, dtdKt(Xt)=−Ht(Xt)+(at−1)Et(Xt). Divergence in Ht prevents score-gap energy alone from recovering K. Online-K approximates this evolution for its sigmoid response: equal gap norms at the same timestep can yield different coefficients when accumulated estimates straddle the threshold, unlike time-only or gap-norm rules. With learned scores, zero initialization and finite steps, Kj is history dependent.
2. Bridge from exact-score theory to finite-step experiments
Analytic mixtures test terminal laws and omission of their corrections. Learned CIFAR-10 tests show sampler and timing dependence without reconstructing high-dimensional terminal-density corrections: at 996 steps, cross-sampler 103MMD2 is 0.259 for w=1, 0.706 for w=3; all six equal-coefficient pulse comparisons have Holm-adjusted p=0.020 (Tables L.1, L.3).
Table N.2 isolates quadrature, initialization and trace errors on a common conditional DDIM trajectory with exact mixture scores. At 256 steps, evidence MAE: 3.009×10−5 with exact initialization/divergence, 0.1970 with zero initialization/one probe, and 0.1913 with zero initialization/exact divergence. These validate ideal quadrature; finer grids alone do not remove initialization bias.
3. Matched mean schedule and computational cost
Table M.1 freezes the timestep-wise mean from the original 150K CIFAR-10 trajectories, matching initial noise, labels, model, grid and FP32 arithmetic. Values: mean ± sample SD over three seeds, 50K images/method/seed.
Online-K achieved lower FID than replay in all three seeds (2.5617 vs. 2.6305 on average), with higher mean recall and lower mean precision.
The low-cost energy proxy reuses conditional/unconditional predictions, avoiding endpoint-estimator backward computation. At fixed DDIM-200/50 on CIFAR/ImageNet, sampling time is 1.002×/1.003× fixed CFG, matching steps and approximate cost.
Proxy mean FID improves on fixed CFG but trails PCS/C$^{2}FGonCIFAR/ImageNet.Qualityusesarchivedevaluations;timingwasmeasuredseparately.EachoffivetunedImageNetmethodsscreens12configurationsat1Kimageseachbeforethree50Kevaluations;CIFARuses18candidatesfororiginalbaselines,sixforOnline−K$, twelve for C$^{2}$FG. Records include seed provenance.