acceptodds
Under review as a conference paper at ICLR 2027

Correcting Guided Diffusion Trajectories with Spectral Alignment

Abstract

The practical success of conditional image generation hinges on fine-grained differences in condition alignment and visual fidelity. Classifier-free guidance (CFG) is central to this success, but its lack of an explicit criterion makes it difficult to assess whether the guided trajectory is progressing as intended. To address this gap, we show that spectral alignment provides a principled criterion for understanding guidance behavior and improving guided diffusion sampling through adaptive correction. Our analysis identifies the spectra of intermediate states as an indicator of consistency with the expected spectral evolution of the forward process. Based on this observation, we introduce Spectral Correction Guidance, a method that corrects deviations from an analytic reference spectrum during sampling. The proposed method is training-free and applicable across diffusion backbones and conditional generation tasks without modifying the underlying model. Experiments demonstrate consistent gains in preference-based metrics over baseline guidance methods in text-to-image generation and improved generation quality over CFG on ImageNet. These improvements persist across a range of guidance scales and with fewer denoising steps. Our analyses and ablations provide insight into guidance behavior and how the proposed method affects generation quality.

Primary area:
generative models
Keywords:
Diffusion Models, Classifier-free Guidance, Spectral Correction

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_nMahOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Spectral Correction Guidance, a training-free method for improving classifier-free guidance (CFG) in diffusion models. The core idea is that the power spectra of intermediate denoising states can indicate whether a guided trajectory remains consistent with the forward process. Accordingly, the authors first derive an analytic reference spectrum from a power-law model of natural-image spectra (Proposition 1), then progressively correct the Fourier magnitude of the current latent toward this target spectrum via phase-preserving log-spectral interpolation. The method is applied only at sampling time and requires no model modification or retraining. The authors validate it on text-to-image generation with SDXL, PixArt-α, and SD3.5, and on class-conditional ImageNet generation with DiT-XL/2, covering standard/high/low guidance scales, multiple sampling budgets, stylized generation, human preference, and multi-seed robustness experiments, reporting consistent gains in FID, IS, and preference metrics.

Critical strengths
  1. The method is training-free and plug-and-play, adding only ~1.7–2.6 s per image, and reports consistent gains across four backbones and on both text-to-image and class-conditional tasks. It also includes an output-only correction ablation (which helps isolate the contribution of in-trajectory correction) and a human preference study with randomized left-right placement.
  2. The diagnostic perspective is novel and the analytic formulation is clean and well motivated. Framing guidance quality as alignment between reverse-state spectra and forward-process spectra is a genuinely fresh and interpretable angle compared to timestep-scheduling or geometry-constraint methods; Proposition 1 gives a closed-form target spectrum from the power-law image model, avoiding simulation of forward states; the log-spectral interpolation is simple (per-frequency rescaling by the ratio R_t(ω), phase-preserving), and empirical support for the power-law assumption in latent space is provided with R² > 0.99.
Critical weaknesses
  1. The core diagnostic claim rests solely on correlational, single-setting evidence. The claim that spectral distance indicates guidance quality is the foundation of the paper, yet its only support is Figure 2, where spectral distance is correlated with FID/IS on a single backbone (SDXL) with 1,000 COCO prompts. Moreover, Eq. 6 compares aggregate spectra rather than per-sample distances, and the paper does not test whether reducing spectral distance by other means also improves the metrics (no causal control experiment). The lack of cross-setting validation of the central mechanistic claim directly undermines my confidence in the source of the gains and the generalizability of the method.
  2. The hyperparameter λ is tuned per backbone and spans two orders of magnitude, with no selection rule. λ ranges from 0.005 (SiT) to 0.1 (SD3.5) to 0.5 (JiT), about two orders of magnitude, and the paper provides no validation-based or automated selection protocol. This makes it impossible to distinguish between "spectral alignment drives quality improvement" and "per-backbone tuning fits the benchmarks," weakening my assessment of the method's intrinsic generality.
  3. The relaxation schedule ρ is only described qualitatively, with no formal definition. ρ is mentioned only qualitatively as "relaxed in the late denoising stage" in the main text and Figure 8; no functional form or timestep threshold is given in the main text.
  4. Table 6 and Table 1 differ by an order of magnitude for SDXL and this is unexplained. Table 6 reports FID ≈ 66.6 and sFID ≈ 344, while Table 1 reports FID ≈ 17 and sFID ≈ 48 for the same model, with no explanation of differences in prompt sets, resolutions, or evaluation protocols. This makes me doubt whether the remaining table values are comparable across experimental settings.
  5. The anomalous sFID of LF-CFG on SD3.5 in Table 1 is not discussed. LF-CFG achieves sFID = 12.61, far below all other methods including the proposed one (49.07), yet the text only generically states that TV-CFG and LF-CFG achieve lower metrics. Without discussion, one of the paper's main selling points, the magnitude of FID/sFID improvement, is undermined by an unexplained outlier, which may even suggest that sFID and visual quality are decoupled in this setting.
  6. The timestep distribution in the expectation of Eq. 6 is unspecified (uniform? what weighting?). In Appendix F, the vanishing of the cross-term requires ẑ₀ and ϵ̂ to be uncorrelated at each frequency, but the main text only states this informally as "z₀ ⊥ ϵ and E[ϵ] = 0" without rigorous justification for Fourier coefficients. These gaps in the foundation of the analytic derivation reduce my assessment of the theoretical rigor.
  7. Algorithm 1 introduces ε for numerical stability, but neither the main text nor Appendix E defines it or gives its value. Though a minor issue, omitting implementation details undermines reproducibility for a method that emphasizes "plug-and-play."
  8. Baseline hyperparameters are set by the authors, and tuning budgets are unclear. Baselines such as CFG++ (w/12.5) and the TV-CFG schedule use hyperparameters chosen by the authors; the paper does not report how many configurations were tried for the baselines, nor was any baseline tuned with the same effort as the proposed method. If the baselines are undertuned, the reported relative improvements may be overstated, which directly affects my trust in the strength of the experimental conclusions.
Questions for authors
  1. Does the relationship between spectral distance and FID/IS hold on other backbones (e.g., PixArt-α, DiT) and under different noise schedules? If additional experiments are not possible, can you at least explain why this relationship is expected to be universal?
  2. How were λ (0.005 to 0.5) and the bin count selected for each backbone? Is there a validation-based protocol? If a sensitivity analysis is performed around the optimal λ, how quickly does performance degrade?
  3. Why do the FID/sFID for SDXL in Table 6 differ by roughly an order of magnitude from those in Table 1? How do the two differ in prompt set, resolution, and evaluation protocol?
  4. LF-CFG's sFID on SD3.5 (12.61) is far below that of your method (49.07). How do you interpret this gap? Does it suggest that sFID and visual quality are decoupled in this setting?
  5. What is the value of the numerical stability constant ε in line 8 of Algorithm 1? How is the timestep distribution in the expectation of t in Eq. 6 defined?
4 comments
AuthorsOpenReview ↗

Response to Reviewer nMah [1/3]

1. Theoretical basis and empirical validation of the spectral criterion (W1, Q1)

The spectral criterion is derived from a necessary condition for distributional consistency between the forward and reverse processes (lines 157–159), rather than from an empirical correlation. Proposition 1 derives a closed-form reference for the expected power spectrum of the forward marginal at each timestep. Distributional consistency requires agreement between the expected power spectra of the corresponding forward and reverse marginals. Eq. 6 quantifies this discrepancy at the distribution level by comparing the expected spectra of the corresponding marginals. Figure 2 examines how this discrepancy relates to generation quality. The preference-based evaluations and human study (Appendix H) provide complementary empirical evidence for the practical benefits of applying this criterion through our method, Spectral Correction Guidance (SCG).


1.A. Cross-backbone validation of the spectral criterion

We extend the analysis in Figure 2 to PixArt-α and SD3.5 to examine the relationship between spectral distance and generation quality across backbones. We follow the evaluation protocol of Figure 2 and evaluate CFG at w = 1, w = 4.5, and w = 13.5 using 1k prompts and the spectral-distance definition in Eq. 6.

BackboneCFG scale (w)Spectral distance ↓FID ↓IS ↑
PixArt-α10.00428789.5717.10
4.50.00375576.1125.28
13.50.01643583.7422.52
SD3.510.01447481.5317.48
4.50.00493966.2926.42
13.50.00793970.5523.49

For both backbones, the standard guidance scale of w = 4.5 achieves the lowest spectral distance, lowest FID, and highest IS among the evaluated scales. Increasing guidance to w = 13.5 increases spectral distance and worsens both generation metrics. These results support the practical value of the spectral criterion derived from distributional consistency and extend its empirical validation beyond SDXL.


1.B. Controlled intervention on the spectral correction direction

Next, we examine the causal contribution of reducing spectral distance to generation quality through a controlled intervention on the correction direction. We compare SCG with Anti-SCG using the same spectral reference at fixed CFG scales. Anti-SCG reverses the correction to move the spectrum away from the reference. This intervention extends the evidence beyond the correlations in Figure 2 by testing the effect of correction toward versus away from the reference. We conduct this comparison using 1k prompts on SDXL at w = 5 and SD3.5 at w = 4.5.

ModelCFG scale (w)MethodHPSv3 ↑ImageReward ↑PickScore ↑HPSv2 ↑CLIP-T ↑FID ↓sFID ↓
SDXL5SCG8.4370.73822.4990.28040.265266.818345.093
Anti-SCG6.6010.54822.2100.26020.267469.435342.040
SD3.54.5SCG9.4010.97422.4810.28390.269567.852329.090
Anti-SCG9.0300.93722.3470.28140.271769.030332.434

Compared with Anti-SCG, SCG achieves higher scores on all four preference metrics and lower FID for both backbones. The HPSv3 gains range from 0.371 to 1.836, the ImageReward gains from 0.037 to 0.190, and the FID reductions from 1.178 to 2.617. These preference gains also hold on SDXL, where Anti-SCG achieves a lower sFID. CLIP-T is lower by 0.0022 for both backbones. Reversing the correction at fixed CFG scales therefore worsens all four preference metrics and FID, providing intervention-based evidence for the contribution of correction toward the reference to generation quality. These results support the practical value of the spectral criterion.

AuthorsOpenReview ↗

Response to Reviewer nMah [2/3]


1.C. From distribution-level consistency to individual trajectory correction

We complement the distribution-level diagnostic in Eq. 6 by measuring how SCG changes spectral deviations in individual sampling trajectories. We follow the default SDXL settings in Appendix E and apply spectral correction during steps 1–25 before relaxing the correction during steps 26–50. The analysis uses 1k real COCO images and 50 paired CFG and SCG trajectories generated using the same prompts and seeds.

We encode the real images into VAE latents and forward-noise them to match the sampling noise levels, providing a baseline for natural sample variation. Eq. 6 measures the distance between expected spectra. This analysis instead computes the mean of individual distances to the shared timestep-dependent reference. For each sample and timestep, we compute the mean squared log-power discrepancy across radial bins and then average these distances over samples and timesteps. This captures deviations that can be masked when spectra are averaged before computing the distance. The comparison examines whether SCG reduces the mean magnitude of these individual deviations toward the level observed in real forward states.

Sampling stepsReal-image mean distanceCFG mean distanceSCG mean distanceReduction from CFG
1–25: correction applied0.013700.036140.0154557.25%
26–50: correction relaxed0.073010.174740.0960545.03%
1–50: full trajectory0.043350.105440.0557547.12%

SCG brings the mean spectral deviation of individual sampling trajectories closer to the level observed in forward-noised real images. During steps 1–25, the distance decreases from 0.03614 to 0.01545, approaching the real-image baseline of 0.01370. During steps 26–50, the distance remains 45.03% lower than CFG after correction is relaxed, showing that early correction influences the subsequent evolution of individual samples. Across the full trajectory, SCG brings the mean distance from 0.10544 to 0.05575, closer to the real-image baseline of 0.04335. The distribution-level consistency criterion guides correction of individual sampling trajectories toward the spectral consistency expected from the forward process.


2. Parameter selection and generality across backbones (W2, Q2)

SCG combines a shared spectral criterion with an adjustable degree of correction to accommodate different sampling dynamics and guidance settings. The parameter λ controls the fraction of the log-spectral discrepancy corrected at each step (Eq. 9). Backbone-specific values adjust the degree of correction without changing the criterion or correction formulation. We set λ through coarse ablations using representative correction ratios such as 0.01, 0.05, 0.1, and 0.5. We inspect the generated images to assess the balance between correcting guidance artifacts and preserving sample-specific detail. Figure 7 compares representative correction ratios of λ = 0.005, 0.05, and 0.5. Section 6.2 shows that weak correction leaves outputs close to CFG and excessive correction can introduce artifacts. The bin counts are chosen with respect to the spatial dimensions of the latent representation and specified in Appendix E. Appendix N examines their effect on spectral correction.

The evidence for the shared criterion extends beyond experiments involving backbone-specific λ settings. The cross-backbone CFG analysis in W1/Q1 provides empirical support for the criterion on PixArt-α and SD3.5 without applying SCG or selecting λ. The controlled interventions provide complementary evidence for the contribution of correcting toward the reference at fixed backbone and guidance settings. These experiments support the applicability of the spectral criterion across backbones, alongside the ablations that guide the choice of correction degree.


3. FID/sFID evaluation settings (W4, Q3)

Table 1 uses 10k generated images for the main evaluation (Section 6), and Appendices I and J use 1k images for supplementary analyses and ablations. Table 6 reports the mean and standard deviation across 10 seeds, using 1k images for each evaluation. The image resolution, real-image reference, and sampling parameters are the same across these evaluations. The numerical scale of FID and sFID can vary with the evaluation sample size. Within each experiment, we hold the prompts, sample size, and evaluation protocol fixed across methods. These controls support meaningful comparisons of the methods or components within each experimental setting.

AuthorsOpenReview ↗

Response to Reviewer nMah [3/3]


4. Further analysis and correction of FID/sFID results (W5, Q4)

The FID and sFID entries for LF-CFG on SD3.5 were computed using a separate torchvision-based evaluator. We recomputed the metrics and reference statistics using the OpenAI guided-diffusion evaluator and the same reference preprocessing as CFG and SCG. The corrected FID is 18.156 instead of 19.930 and the corrected sFID is 46.953 instead of 12.614. The other five metrics remain unchanged. We rechecked the remaining evaluations and found no other FID/sFID protocol mismatches. We will correct these two entries in Table 1.

After the correction, TV-CFG and LF-CFG remain among the methods with lower FID and sFID on SD3.5. SCG achieves higher scores on all four preference metrics than both methods. These different rankings show that better distributional scores do not necessarily coincide with higher preference scores. We therefore assess generation quality through seven complementary metrics and human evaluation. We will expand the discussion to explain this metric pattern and the trade-offs exhibited by TV-CFG and LF-CFG.


5. Derivation and implementation details (W3, W6, W7, Q5)

Correction relaxation. Sampling proceeds from 1 to 0 over the interval [0, 1]. SCG applies spectral correction over (ρ, 1].

Timestep averaging. We do not assign additional weights to specific timesteps. Eq. 6 uses a uniform average over the inference timesteps of each backbone’s default scheduler.

Fourier cross-term. The Fourier transform is deterministic, so independent signal and noise remain independent in the Fourier domain. Linearity preserves the zero mean of the noise, so the cross-term vanishes in expectation at each frequency.

Numerical stability. We use ε = 10⁻¹².

We will clarify these points and specify the implementation details in the revised manuscript.


6. Baseline parameter settings and comparison protocol (W8)

The baseline settings follow the cited methods and comparison protocols described in Appendix E. For CFG++, we use λ = w/12.5 based on the approximate scale correspondence reported in the original work [1] and the same conversion used in ADG [2]. This conversion was not introduced for our comparison. ADG also reports negative ImageReward scores for CFG++ on SD3.5 under this conversion.

We also evaluated CFG++ on SD3.5 at λ = 0.28 (w = 3.5) and λ = 0.56 (w = 7), below and above the reported setting of λ = 0.36 (w = 4.5). These values follow the same λ = w/12.5 conversion. Reducing λ to 0.28 improves HPSv3 from −4.826 to −1.587 and ImageReward from −1.082 to −0.524. Increasing λ to 0.56 lowers ImageReward to −1.659. ImageReward remains negative across all three settings.

For TV-CFG, we follow the symmetric schedule used in the original main experiments [3]. The guidance scale peaks at the midpoint of sampling and its average is matched to the constant CFG scale in the corresponding experiment. These choices are documented in Appendix E and follow the published configuration.

Our comparisons evaluate the baselines under settings grounded in prior work. The improvements over standard CFG and the controlled interventions reported above provide additional evidence for SCG that does not depend on the parameter choices for CFG++ or TV-CFG.


Reference

[1] Chung et al. CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models. ICLR 2025.

[2] Jin et al. Angle Domain Guidance: Latent Diffusion Requires Rotation Rather Than Extrapolation. arXiv:2506.11039, 2025.

[3] Jin et al. Stage-wise Dynamics of Classifier-Free Guidance in Diffusion Models. ICLR 2026.

AuthorsOpenReview ↗

Follow-up on our rebuttal

Thank you again for your review.

We extended the spectral distance analysis across backbones to assess its relationship with generation quality beyond a single model. We also added controlled ablations and trajectory analyses to examine the contribution of the proposed correction to generation quality and spectral consistency in individual samples. We clarified the evaluation protocols, baseline comparisons, and implementation details. These additions provide further empirical support for the practical value of the spectral criterion in assessing and correcting guided sampling trajectories.

We would appreciate your feedback on whether these additions address your main concerns.

Reviewer_qEvTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that classifier-free guidance (CFG) lacks an explicit criterion for whether a guided trajectory is progressing as intended, and proposes to use empirical spectral alignment as such a criterion: the power spectra of intermediate states during guided sampling should match the expected spectra of the forward diffusion process. The authors derive an analytic reference spectrum from the power-law statistics of natural images. They then introduce Spectral Correction Guidance (SCG), a training-free method that rescales the Fourier magnitude of the latent toward this reference at each early denoising step. Experiments on multiple generators report improvements over CFG, mainly in preference-based metrics (e.g., HPS, IR etc.), and most clearly at high guidance scales.

Critical strengths
  1. The method needs no extra denoiser evaluations or trainining, it can be applied to various backbones and models.
  2. The idea of using spectrum as a criterion is somehow new and interesting.
Critical weaknesses
  1. My biggest concerns of this paper is : Is the power spectrum a reliable criterion? I am not sure to what extent the spectral criterion is valid and helpful: (a) Ensemble vs. individual images. The power law describes the ensemble-averaged spectrum, not individual images. Even for real COCO images, individual spectra are not expected to follow it closely. Torralba & Oliva (2003) report per-image exponents of α ≈ 2 with a standard deviation of about 0.5–0.6, as well as systematic differences across scene categories. SCG nevertheless pulls every sample toward a single, unconditional mean spectrum. Two observations are consistent with this concern: i) The method is unstable on the pixel-space JiT, where the authors themselves note large sample-level spectral variability; ii) The correction must be relaxed in late steps, where the target is dominated by the image term rather than the noise term.

(b) Latent space. The method operates in VAE latent space, where the natural-image power law has no theoretical justification. Fig. 12 only shows that the average spectrum of encoded COCO images can be fitted (R² > 0.99). This fit is in-sample, with three free parameters on a smooth, radially averaged curve, which makes it a weak test. It remains unclear: whether individual latents, individual channels, or latents of other real-image datasets (e.g., closer to the training distribution) follow this spectrum; how large these natural deviations are compared to the deviations SCG corrects.

Given this, using the spectrum as the criterion needs further verification.

  1. Accordingly, we see some counterintuitive metric pattern. If the correction truly brings samples closer to the spectrum of real images, the most direct benefit should appear in distributional metrics such as FID. Instead, the text-to-image results show a different pattern:

FID gains over CFG are marginal (mostly 0.1–0.6), and sFID is mixed. The gains concentrate in HPS, ImageReward and PickScore. These are learned preference models known to favor vivid, saturated, high-contrast and sharp images.

  1. What does the effective guidance weight look like, and how does it compare to schedulers?

I would like to see the effective guidance weight over time induced by SCG for 30–50 randomly sampled prompts, for example by projecting the correction onto the guidance direction ε_c − ε_u. Since the correction is applied early and relaxed late, I suspect it mostly reshapes the guidance weight into a common curve. If so, SCG would essentially be a particular guidance schedule found through an empirical formula, and would differ little from the guidance-scheduler literature (e.g., Wang et al., TMLR 2024; Kynkäänniemi et al., 2024). Such schedulers are cost-free, and simple linear or cosine schedules generalize reasonably well. They have been evaluated across many settings (CIFAR-10, ImageNet, SD1.5, SDXL, SD3), and they are particularly effective at high guidance scales, which is exactly the regime where SCG shows its largest gains. The only scheduler baseline in the paper, TV-CFG, uses a Λ-shaped schedule.

A direct comparison with these schedulers at default and high w, as well as SCG combined with them, is needed.

Minor: Concurrent work (SPA, arXiv 2607.22091) observes that spectral mismatch exists even without guidance (as exposure bias), that it is channel-wise and timestep-dependent in latent models, and that log-spectra are skewed. Although SPA measures the spectrum of x0^\hat{x_0} rather than z_t (maybe authors should try to measure x0^\hat{x_0} as well to see if they fit the empirical spectrum curve)... A discussion would be helpful.

Questions for authors

See my major concerns.

10 comments
AuthorsOpenReview ↗

Response to Reviewer qEvT [1/5]

1. Scope and empirical validation of the spectral criterion (W1)

The distinction between ensemble and individual spectra is reflected in the design and analysis of our method, Spectral Correction Guidance (SCG). The spectral model characterizes an expected spectrum and does not assume that every image or channel follows the same power law. SCG uses the reference to control spectral deviations during sampling and accommodates sample-specific variation through an adjustable degree of correction and late-stage relaxation. Section 6.2 and Appendix O examine this trade-off, and Appendix M discusses the uneven performance on JiT. These considerations concern how the ensemble reference is applied to individual samples and do not contradict the expected-spectrum relation underlying the criterion.

We agree that natural-image statistics alone do not establish the same spectral form in VAE latent space. The fits in Appendix N support this form as a model of the expected power spectra of clean latents. This characterization is central to SCG because Proposition 1 uses the spectral model at z₀ to provide an explicit reference at each timestep, allowing guided trajectories to be assessed against the expected spectral evolution of the forward process. The additional experiments below examine the applicability of this spectral form across datasets and the practical value of using the reference to correct individual sampling trajectories.


1.A. Criterion validation across datasets

We provide further evidence for the applicability of the spectral form specified in Proposition 1 across datasets in VAE latent space. We estimate the expected radial power spectra from 10k ImageNet validation images encoded using the pretrained VAEs of the T2I backbones. We fit the same functional form in Eq. 7 to these estimated spectra. This evaluates the spectral form on an external dataset that was not chosen to match the training distribution of the T2I backbones.

T2I backboneSDXLPixArt-αSD3.5
R² on ImageNet validation (10k)> 0.99> 0.99> 0.99

These results support modeling ensemble spectra through shared domain-level spectral regularities in latent space. The proposed functional form captures the expected latent spectra of an external dataset across the evaluated T2I backbones. The DiT experiments reported in the manuscript provide complementary evidence for the practical use of an external spectral reference. SCG improves ImageNet class-conditional generation using a reference estimated from COCO, without using ImageNet training or evaluation images for reference fitting (Appendix E, lines 793–797).

AuthorsOpenReview ↗

Response to Reviewer qEvT [2/5]


1.B. Criterion validation through functional form ablation

We examine whether the frequency profile of the proposed spectral reference (Eq. 7) contributes to generation quality. We compare SCG with three alternative spectral shapes. Flat uses one constant across all radial bins. Two-band flat uses separate constants for the low- and high-frequency bands. Linear fits a straight line to the clean spectrum. All three alternatives preserve the total power of the clean reference, and Two-band flat also preserves power within each band. We apply the same forward noise schedule to construct the timestep-dependent reference spectra. The comparisons use 1k prompts on SDXL at w = 5 and SD3.5 at w = 4.5.

ModelCFG scale (w)TargetHPSv3 ↑ImageReward ↑PickScore ↑HPSv2 ↑CLIP-T ↑FID ↓sFID ↓
SDXL5SCG8.4370.73822.4990.28040.265266.818345.093
Flat7.5720.63522.2870.26640.265269.597348.104
Two-band flat7.6060.64622.2560.26790.265470.327350.307
Linear7.6950.65122.2790.26860.265469.991351.603
SD3.54.5SCG9.4010.97422.4810.28390.269567.852329.090
Flat8.1610.88621.8600.27600.280569.646335.721
Two-band flat8.2570.90421.8880.27740.280869.275335.892
Linear8.2580.90321.8870.27720.280669.272335.856

SCG achieves higher scores on all four preference metrics and lower FID and sFID than each alternative target across both backbones. Compared with the best alternative for each metric, SCG improves HPSv3 by 0.742 and ImageReward by 0.087 on SDXL. On SD3.5, the HPSv3 gain is 1.143 and the ImageReward gain is 0.070. FID is lower than the best alternative by 2.779 on SDXL and 1.420 on SD3.5. Two-band flat and Linear achieve higher CLIP-T scores on SDXL, and all three alternatives achieve higher CLIP-T scores on SD3.5. The gains across preference and distributional metrics under matched total power support the proposed spectral form as a suitable reference for correction in the evaluated VAE latent spaces.


1.C. Criterion validation through controlled spectral alignment

We assess the contribution of spectral alignment to generation quality by comparing SCG with Anti-SCG at fixed CFG scales. Anti-SCG reverses the spectral correction to move the spectrum away from the reference, allowing us to test whether this reversal degrades generation quality. We conduct this comparison using 1k prompts at the standard guidance scales of w = 5 for SDXL and w = 4.5 for SD3.5.

ModelCFG scale (w)MethodHPSv3 ↑ImageReward ↑PickScore ↑HPSv2 ↑CLIP-T ↑FID ↓sFID ↓
SDXL5SCG8.4370.73822.4990.28040.265266.818345.093
Anti-SCG6.6010.54822.2100.26020.267469.435342.040
SD3.54.5SCG9.4010.97422.4810.28390.269567.852329.090
Anti-SCG9.0300.93722.3470.28140.271769.030332.434

Compared with Anti-SCG, SCG achieves higher scores on all four preference metrics and lower FID for both backbones. The HPSv3 gains range from 0.371 to 1.836, the ImageReward gains from 0.037 to 0.190, and the FID reductions from 1.178 to 2.617. These preference gains also hold on SDXL, where Anti-SCG achieves a lower sFID. CLIP-T is lower by 0.0022 for both backbones. These results support the role of correcting individual sampling trajectories toward the ensemble spectral reference in improving generation quality across the evaluated backbones.

AuthorsOpenReview ↗

Response to Reviewer qEvT [3/5]


1.D. Quantifying the effect of SCG on spectral deviations

We measure the spectral distances of CFG and SCG trajectories from the shared reference to assess the effect of correction. We follow the default SDXL settings in Appendix E and apply spectral correction during steps 1–25 before relaxing the correction during steps 26–50. The analysis uses 1k real COCO images and 50 paired CFG and SCG trajectories generated using the same prompts and seeds.

We encode the real images into VAE latents and forward-noise them to match the sampling noise levels, providing a baseline for natural sample variation. Eq. 6 measures the distance between expected spectra. This analysis instead computes the mean of individual distances to the shared timestep-dependent reference. For each sample and timestep, we compute the mean squared log-power discrepancy across radial bins and then average these distances over samples and timesteps. This captures deviations that can be masked when spectra are averaged before computing the distance. The comparison examines whether SCG reduces the mean magnitude of these individual deviations toward the level observed in real forward states.

Sampling stepsReal-image mean distanceCFG mean distanceSCG mean distanceReduction from CFG
1–25: correction applied0.013700.036140.0154557.25%
26–50: correction relaxed0.073010.174740.0960545.03%
1–50: full trajectory0.043350.105440.0557547.12%

SCG brings the mean spectral deviation closer to the level observed in forward-noised real images during both sampling intervals. During steps 1–25, the distance decreases from 0.03614 to 0.01545, approaching the real-image baseline of 0.01370. During steps 26–50, the distance remains 45.03% lower than CFG after correction is relaxed, showing that the influence of early correction persists in subsequent denoising steps. Across the full trajectory, SCG brings the mean distance from 0.10544 to 0.05575, closer to the real-image baseline of 0.04335. These measurements provide empirical support for the spectral aspect of distributional consistency targeted by SCG.


2. Interpretation of the metric pattern (W2)

SCG focuses on spectral consistency during the denoising process, whereas FID and sFID evaluate distributional similarity in the feature space of final images. To improve generation quality, SCG uses timestep-dependent reference spectra to correct deviations in intermediate states and accommodates sample-specific variation through adjustable correction and late-stage relaxation. FID and sFID provide meaningful evaluations of the final outputs, but agreement in power spectra does not determine these feature distributions. We therefore assess the practical benefits of SCG through seven complementary metrics covering distributional similarity, text alignment, and preference, together with human evaluation. We also note that the reference dataset was not chosen to match the training distribution of each backbone.

As the reviewer suggests, changes in saturation, contrast, or sharpness may contribute to higher preference scores. Such gains could also arise from spectral correction applied only to the final output. The output-only ablation in Appendix J (Table 7) examines whether such final-output adjustment is sufficient to obtain the benefits of SCG. Output-only correction applies the spectral adjustment only at the end of sampling, whereas SCG corrects intermediate states using the expected spectra of the forward marginals. We recall the reported results below.

MethodHPSv3 ↑ImageReward ↑PickScore ↑HPSv2 ↑CLIP-T ↑FID ↓sFID ↓
CFG8.1040.70822.4730.2750.26767.430345.259
Output-only correction8.2180.71022.4820.2760.26767.557345.246
SCG8.4370.73822.4990.2800.26566.818345.093

Output-only correction improves all four preference scores over CFG despite a higher FID, showing that preference gains do not require a corresponding FID improvement. SCG achieves further gains on all four preference metrics and lower FID and sFID than output-only correction. ImageReward increases from 0.710 to 0.738, and FID decreases from 67.557 to 66.818. CLIP-T is lower by 0.002. These results support the contribution of timestep-dependent trajectory correction to generation quality beyond the gains obtained from final-output spectral adjustment. The human study reported in the manuscript also finds that participants prefer SCG over CFG in both visual quality and prompt alignment, providing complementary evidence for perceptual benefits beyond the learned preference scores (Appendix H).

Reviewer_92PsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Classifier-free guidance is important for conditional image generation, yet it offers little to tell whether the sampling process is going well or not. As a result, its failure often shows at the end with a number of artifacts, over-saturation, lost details. The paper proposes a check at each stage of denoising. Because the ideal reference can’t be observed while generating, the authors derive it from a well-known regularity in how natural photographs distribute detail across scales, which gives a simple formula for what the image should look like at every point in the process. After that, the rebalancing technique at the intermediate image toward that reference at each step. Across several text-to-image and class-conditional models, the generated images show better quality.

Critical strengths
  1. The correction criterion is derived rather than prescribed.
  2. the empirical evidence is rich
Critical weaknesses
  1. The paper’s central claim, that spectral alignment is a criterion for guidance quality, is supported only by a correlation that is partly built in, and the strongest numerical gains appear on the metric most directly coupled to the quantity being optimized. FID and especially sFID are themselves sensitive to spectral statistics, so a method that explicitly matches power spectra to a natural-image reference should be expected to improve them regardless of whether perceptual quality improves. The largest reported gain in the paper is exactly this: sFID 7.42 → 4.92 on ImageNet. The paper never establishes that reducing spectral distance improves generation quality independently of spectrum-sensitive metric, there is no intervention experiment (e.g. degrading spectral alignment in a controlled way and showing quality falls), and the correlation is measured only along the one-dimensional family of CFG scales, where many things covary.

  2. The gains are small relative to the amount of per-backbone tuning, the reference is estimated on the evaluation distribution, and the baseline set omits the closest competitors. On many metrics the improvement is at or near the noise floor — PickScore 22.41 → 22.43 on SDXL, CLIP-T flat or marginally worse in several rows — with real movement confined to HPSv3 and ImageReward.

  3. The method carries per-backbone hyperparameters with no selection rule: λ spans 0.005 to 0.5 across models (a factor of 100), plus a relaxation point ρ and a bin count that Appendix N shows materially changes behaviour. The paper gives no principled

  4. FID/sFID are inconsistent across tables (Table 1: 17.0 / 47.9; Appendices I–J: ~66.6 / ~344) with no explanation of the differing protocol. The sFID $ \approx 344$ figure in particular looks wrong.

  5. The related-work taxonomy (scheduled / geometry-aware / frequency-based) has no slot for methods that correct each step toward a precomputed distributional target, e.g. Representative Guidance (ICLR 2025), which is very close to the idea of the paper.

Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 92Ps [1/3]

1. Derivation and empirical validation of the spectral criterion (W1)

Rather than relying on correlation, we derive our spectral criterion from a necessary condition for distributional consistency between the forward and reverse processes (lines 157–159). Proposition 1 derives a closed-form reference spectrum by propagating the empirical power-law model of natural-image spectra through the forward process. Our method, Spectral Correction Guidance (SCG), uses these timestep-dependent references to correct spectral deviations in intermediate states during denoising.

The DiT and text-to-image experiments provide complementary evidence for the benefits of SCG. The DiT experiments evaluate distributional quality through FID and sFID. To assess perceptual benefits beyond these metrics, we also evaluate SCG in text-to-image generation using multiple preference metrics and a human study. SCG improves several preference scores, and participants prefer SCG over CFG in both visual quality and prompt alignment (Appendix H). These findings provide evidence that the benefits of SCG extend beyond improvements in FID and sFID.

The output-only ablation in Appendix J (Table 7) provides evidence that the benefits of SCG are not confined to FID/sFID or to spectral correction of the final output. Output-only correction applies the spectral adjustment only at the end of sampling, whereas SCG corrects intermediate states using the expected spectra of the forward marginals. We recall the reported results below.

MethodHPSv3 ↑ImageReward ↑PickScore ↑HPSv2 ↑CLIP-T ↑FID ↓sFID ↓
CFG8.1040.70822.4730.2750.26767.430345.259
Output-only correction8.2180.71022.4820.2760.26767.557345.246
SCG8.4370.73822.4990.2800.26566.818345.093

Output-only correction improves all four preference scores over CFG despite a higher FID, showing that the preference gains do not simply track FID improvement. SCG achieves further gains on all four preference metrics and lower FID and sFID than output-only correction. CLIP-T is lower by 0.002. The additional preference gains support the contribution of timestep-dependent trajectory correction beyond spectral adjustment of the final output.

The additional controlled ablations below examine the contribution of the spectral criterion by varying the correction direction and the reference’s frequency profile at fixed CFG scales.


1.A. Controlled intervention on spectral alignment

We assess the contribution of spectral alignment to generation quality by comparing SCG with Anti-SCG at fixed CFG scales. Anti-SCG reverses the spectral correction to move the spectrum away from the reference, allowing us to test whether this reversal degrades generation quality. We conduct this comparison using 1k prompts at the standard guidance scales of w = 5 for SDXL and w = 4.5 for SD3.5.

ModelCFG scale (w)MethodHPSv3 ↑ImageReward ↑PickScore ↑HPSv2 ↑CLIP-T ↑FID ↓sFID ↓
SDXL5SCG8.4370.73822.4990.28040.265266.818345.093
Anti-SCG6.6010.54822.2100.26020.267469.435342.040
SD3.54.5SCG9.4010.97422.4810.28390.269567.852329.090
Anti-SCG9.0300.93722.3470.28140.271769.030332.434

Compared with Anti-SCG, SCG achieves higher scores on all four preference metrics and lower FID for both backbones. The HPSv3 gains range from 0.371 to 1.836, the ImageReward gains from 0.037 to 0.190, and the FID reductions from 1.178 to 2.617. These preference gains also hold on SDXL, where Anti-SCG achieves a lower sFID. CLIP-T is lower by 0.0022 for both backbones. These results support the role of correcting the spectrum toward the reference in improving generation quality across backbones without changing the CFG scale.

AuthorsOpenReview ↗

Response to Reviewer 92Ps [2/3]


1.B. Controlled comparison of reference functional forms

We examine whether the frequency profile of the proposed spectral reference (Eq. 7) contributes to generation quality. We compare SCG with three alternative spectral shapes. Flat uses one constant across all radial bins. Two-band flat uses separate constants for the low- and high-frequency bands. Linear fits a straight line to the clean spectrum. All three alternatives preserve the total power of the clean reference, and Two-band flat also preserves power within each band. We apply the same forward noise schedule to construct the timestep-dependent reference spectra. The comparisons use 1k prompts on SDXL at w = 5 and SD3.5 at w = 4.5.

ModelCFG scale (w)TargetHPSv3 ↑ImageReward ↑PickScore ↑HPSv2 ↑CLIP-T ↑FID ↓sFID ↓
SDXL5SCG8.4370.73822.4990.28040.265266.818345.093
Flat7.5720.63522.2870.26640.265269.597348.104
Two-band flat7.6060.64622.2560.26790.265470.327350.307
Linear7.6950.65122.2790.26860.265469.991351.603
SD3.54.5SCG9.4010.97422.4810.28390.269567.852329.090
Flat8.1610.88621.8600.27600.280569.646335.721
Two-band flat8.2570.90421.8880.27740.280869.275335.892
Linear8.2580.90321.8870.27720.280669.272335.856

SCG achieves higher scores on all four preference metrics and lower FID and sFID than each alternative target across both backbones. Compared with the best alternative for each metric, SCG improves HPSv3 by 0.742 and ImageReward by 0.087 on SDXL. On SD3.5, the HPSv3 gain is 1.143 and the ImageReward gain is 0.070. FID is lower than the best alternative by 2.779 on SDXL and 1.420 on SD3.5. Two-band flat and Linear achieve higher CLIP-T scores on SDXL, and all three alternatives achieve higher CLIP-T scores on SD3.5. The gains across all four preference metrics and both distributional metrics under matched total power support the suitability of the proposed spectral form as a reference for correction.


2. Practical significance of the performance gains (W2)

The multi-seed evaluation and human study support meaningful improvements in generation quality despite small numerical changes in some metrics. Across 10 random seeds on SDXL, the HPSv2 gain of 0.0050 is large relative to the reported standard deviations of 0.0003–0.0004 (Appendix I, Table 6). These results provide evidence of improvements beyond HPSv3 and ImageReward, and the human study provides complementary support for the perceptual benefits of SCG (Appendix H). SCG achieves these gains across multiple preference metrics and different backbones without retraining.

We evaluate SCG using seven quantitative metrics to capture complementary aspects of generation quality. This evaluation design accounts for the different responses of metrics documented in guidance research. TCFG [1] and CFG++ [2] report settings where FID improves without a change in CLIP at the reported precision. TV-CFG [3] shows that preference gains can accompany small changes in FID and CLIP. We therefore assess the practical benefits of SCG through the combined evidence from distributional, text-alignment, and preference metrics.

In the DiT experiments, the spectral reference is estimated from COCO without using ImageNet training or evaluation images for reference fitting (Appendix E). SCG improves ImageNet class-conditional generation with this external reference, providing evidence that its benefits are not restricted to references estimated from the evaluation distribution. For the text-to-image experiments, the reference coefficients are fitted to clean latent spectra from COCO rather than selected using generation metrics (lines 237–240).

AuthorsOpenReview ↗

Response to Reviewer 92Ps [3/3]


3. Parameter choices across backbones (W3)

SCG combines a shared spectral criterion with explicit control over its application during sampling. The bin count, ρ, and λ specify the frequency resolution, timing, and degree of correction. This design allows the same guiding principle to be applied across backbones with different representations and sampling dynamics.

The bin count is chosen with respect to the spatial dimensions of the latent representation. SDXL, PixArt-α, and SD3.5 share 64 bins for their 128 × 128 spatial dimensions despite differences in channel count. DiT uses 32 bins for its 32 × 32 spatial dimensions. The bin count determines the frequency resolution of the radial spectrum. Appendix N examines how coarse and fine binning affect the correction.

The relaxation parameter ρ controls when spectral correction is relaxed during denoising (lines 470–473). Relaxing constraints in the frequency domain allows more freedom for fine-detail synthesis during later stages of image generation [4]. Section 6.2 and Appendix O support an intermediate relaxation point that balances spectral alignment with preserving sample-specific detail. We use ρ = 0.5 across SDXL, SD3.5, SiT, and JiT.

We set λ through coarse ablations using representative correction ratios such as 0.01, 0.05, 0.1, and 0.5. We inspect the generated images to assess the balance between correcting guidance artifacts and preserving sample-specific detail. The parameter controls the fraction of the log-spectral discrepancy corrected at each step (Eq. 9). Figure 7 illustrates this process by comparing λ = 0.005, 0.05, and 0.5 to examine the balance between spectral alignment and preserving sample-specific characteristics. A comparison of Tables 1 and 2 shows that CFG preference scores decline more for SD3.5 than for SDXL and PixArt-α when moving from default to high guidance under the evaluated settings. Figure 5 also shows over-saturated and over-simplified outputs from CFG at high guidance. This sensitivity motivates a greater degree of correction for SD3.5, for which we use λ = 0.1 at default guidance compared with λ = 0.05 for SDXL and PixArt-α. For PixArt-α, we increase λ from 0.05 at w = 4.5 to 0.1 at w = 9 to correct a larger fraction of the deviation from the reference under stronger guidance (Appendix E).


4. FID/sFID evaluation settings (W4)

Table 1 uses 10k generated images for the main evaluation (Section 6), and Appendices I and J use 1k images for supplementary analyses and ablations. Table 6 reports the mean and standard deviation across 10 seeds, using 1k images for each evaluation. The numerical scale of FID and sFID can vary with the evaluation sample size. Within each experiment, we hold the prompts, sample count, and evaluation protocol fixed across methods. These controls support meaningful comparisons of the methods or components within each experimental setting.


5. Relation to Representative Guidance (W5)

Representative Guidance (RepG) [5] is relevant to SCG because both methods use reference targets to guide the sampling trajectory. RepG uses precomputed class-specific representatives from a self-supervised encoder and applies feature-matching gradients to predicted clean images. SCG uses a timestep-dependent reference power spectrum derived from the forward process and rescales the Fourier magnitudes of intermediate noisy states, preserving their phase. This correction requires no additional guidance encoder or backpropagation. We will expand Section 5 to include guidance based on reference targets and clarify the connection to RepG and the differences in target representation and correction mechanism.


Reference

[1] Kwon et al. TCFG: Tangential Damping Classifier-free Guidance. CVPR 2025.

[2] Chung et al. CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models. ICLR 2025.

[3] Jin et al. Stage-wise Dynamics of Classifier-Free Guidance in Diffusion Models. ICLR 2026.

[4] Karras et al. Alias-Free Generative Adversarial Networks. NeurIPS 2021.

[5] Dinh et al. Representative Guidance: Diffusion Sampling with Coherence. ICLR 2025.

AuthorsOpenReview ↗

Follow-up on our rebuttal

Thank you again for your review.

We added controlled ablations of the correction direction and reference spectral form. These experiments show that the correction direction and reference functional form have a meaningful effect on generation quality, providing evidence for the benefits of the proposed correction beyond distributional metrics. We also clarified the evaluation protocols and the use of an external reference for ImageNet generation.

We would appreciate your feedback on whether these additions address your main concerns.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response

We appreciate all reviewers’ careful evaluations and constructive suggestions. The reviewers highlighted the following strengths:

  • The derivation and analytic formulation of the spectral criterion (92Ps and nMah).
  • The diagnostic perspective based on power spectra (qEvT and nMah).
  • The training-free formulation and applicability across backbones (qEvT and nMah).
  • The extensive empirical evaluation (92Ps and nMah).

A central concern shared across the reviews is the practical value of the spectral criterion used in Spectral Correction Guidance (SCG). We address this concern through additional experiments and analyses examining the following aspects:

  • The effect of correcting toward versus away from the spectral reference (92Ps, qEvT, and nMah).
  • The role of the reference spectral form in generation quality (92Ps and qEvT).
  • The applicability of the latent spectral model across datasets (qEvT).
  • The relationship between spectral distance and generation quality across backbones (nMah).
  • The assessment of whether improvements in spectral consistency extend to individual sampling trajectories (qEvT and nMah).
  • The contribution of correction components beyond scalar guidance adjustment (qEvT).
  • The contribution of SCG under different guidance schedules (qEvT).

Detailed responses are provided in the individual threads below.

We also encourage reviewers to consult the related results in the other threads as several experiments address shared concerns. The corresponding changes will be incorporated into the revised manuscript and highlighted in blue.

We would be happy to discuss any remaining concerns. Please let us know if any points require further clarification.

AuthorsOpenReview ↗

Revised Manuscript

We have uploaded a revised manuscript incorporating the additional experiments and clarifications discussed in our rebuttal. Revisions are highlighted in blue.

We would appreciate your feedback on whether these additions address your main concerns.

AuthorsOpenReview ↗

Summary of the Discussion

We thank the reviewers and the Area Chair for their time and feedback. A central question across the reviews was whether the expected forward spectrum provides a useful reference for assessing spectral deviations in guided sampling trajectories and whether correcting these deviations improves generation quality.

In our initial rebuttal, we addressed this question through the following analyses.

  • Spectral criterion across datasets and backbones. We tested the clean-latent spectral model on ImageNet and found that the same functional form fits the mean spectra of all three VAEs (Table 10). We also extended the guidance-scale analysis to PixArt-α and SD3.5. On both backbones, the scale with the smallest spectral distance also yielded the lowest FID and highest IS among the settings tested (Table 11).

  • Effect of the reference spectrum and correction direction. To test whether correcting toward the reference contributes to generation quality, we held the CFG scale fixed and changed either the correction direction or the shape of the reference. On both SDXL and SD3.5, reversing the correction direction lowered all four preference scores and increased FID relative to SCG (Table 12). Each alternative reference shape produced the same pattern even though total power was preserved (Table 14).

  • Effect on individual trajectories and generated images. We measured the spectral discrepancy for each latent before averaging across trajectories. Even after correction ended, the mean discrepancy under SCG remained closer to the level observed in forward-noised real latents than under CFG (Table 15). Applying spectral correction during denoising also yielded larger gains in HPSv3 and ImageReward than applying it only to the final output (Table 7). In the human evaluation, participants preferred SCG over CFG for both visual quality and prompt alignment (Table 5).

In response to Reviewer qEvT’s follow-up, we plotted individual sample spectra alongside the shared reference (Figure 15). On SDXL, matching any one of the four suggested image statistics to its value in the corresponding SCG output did not reproduce the full preference gains of SCG (Table 13). The effective CFG weight corresponding to the projected correction varies across samples at the same denoising step (Figure 16). Adding the orthogonal component further improves HPSv3, ImageReward, and HPSv2 at both guidance scales (Table 16).

The revised manuscript incorporates these analyses and clarifies the derivation, the choice of correction and baseline hyperparameters, and the FID evaluation protocol. We also corrected the LF-CFG FID and sFID entries for SD3.5. We appreciate the reviewers’ questions, which helped us test the spectral criterion more directly and clarify the scope of our conclusions.

Best regards,
Authors of Submission 32501

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.