acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Aware Diffusion Guidance via Curvature-Adaptive Tubular Correction

Abstract

Gradient-guided diffusion samplers provide flexible priors for inverse problems and conditional generation, but strong guidance can move the sampling trajectory into regions where the learned score is poorly supported. Existing tangent-projection strategies limit first-order departure from an iso-density surface, yet discard potentially useful normal motion and overlook the second-order departure induced by tangent motion on a curved surface. We introduce curvature-adaptive tubular correction (CAT), a training-free plugin that regulates both effects within a shared, noise-dependent geometric budget. CAT decomposes the guidance gradient into normal and tangent components, charges normal displacement at first order and tangent displacement according to directional curvature, and obtains their jointly optimal magnitudes from a one-dimensional dual equation. Armijo backtracking calibrates the resulting finite step against the actual guidance objective, while matrix-free directional derivatives avoid constructing the full score Jacobian. We establish local guarantees for the tubular approximation, uniqueness of the correction, and sufficient objective decrease. Across seven inverse problems on FFHQ and ImageNet, CAT improves the evaluated pixel- and latent-space host samplers, with particularly consistent gains in perceptual metrics. It also improves black hole reconstruction on InverseBench. These results support curvature-aware tubular control as a reusable mechanism for stabilizing diffusion guidance.

Primary area:
generative models
Keywords:
Diffusion Models, Inverse Problems, Curvature-Aware Guidance

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rntEOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

Summary

  • This paper propose a looser constraint of the MCG manifold (M\mathcal{M}) [Improving Diffusion Models for Inverse Problems using Manifold Constraints] and constraint the updated latent to a safe region that is a tube around MCG manifold. This relaxation removes the linear operator limits of MCG, makes the algorithm appliable to a broader family of forward operators.
  • The authors provide thorough empirical study to validate the performance of proposed approch.
Critical strengths

Strength

  • The overall idea of relaxing MCG constraint to non-linear operators looks novel and interesting. The proposed approach is supported by theoretical analysis.
  • The empirical validation is thorough and solid. Both ImageNet and FFHQ dataset are included, both pixel space and pixel space are included. PSNR, LPIPS, SSIM and rFID are all included and 7 operators are included.
  • The Armijo frequency provides a better complexity-performance trade-off than vanilla increasing the parameter of DPS.
Critical weaknesses

Weakness

  • The proposed approach introduces quite some new parameters, which non trivially increase the complexity of the proposed approach. I am not sure whether all parameters are covered in Table 4 and Table 5. For example, γt\gamma_t seems to be determined by Eq.11, but how is β,c\beta,c in Eq.11 determined? Sec 3.2 introduces those parameters step by step, maybe adding a overall description and collection of all the parameters can help the readers understand the methods better.
  • The proposed approach can be hard to implement. According to Algorithm 1, the proposed approach written into sudo code is several time longer than DPS baseline (See algoritm 1 of [ICLR 22 Diffusion Posterior Sampling for General Noisy Inverse Problems]). This makes readers wonder whether the performance gain really comes from a robust neat improvement, or comes from additional complexity and nudging of parameters.
  • The ablation study in Table 2 is using non-linear deblurring as operator, which is kind of unusual. Can the authors provide results with more commonly used linear operators such as super-resolution? As baselines, DPS and PSLD seems to totally fail for this task (with PSNR 17dB). According to ablation studies, Armijo step seems to contribute most to performance, and the other components are not so important?
  • In complexity Table 6, what is the sampling step for vanilla DPS?
  • As the proposed approach follows the manifold assumption of MCG, it would be great if it could add MCG as baseline for linear operators. Besides, another possible assumption is the spherical manifold assumption, which is adopted by DSG [Guidance with Spherical Gaussian Constraint for Conditional Diffusion] and DMAP [Rethinking Diffusion Posterior Sampling: From Conditional Score Estimator to Maximizing a Posterior]. It would be better if the authors could discuss and compare to DSG and DMAP.
  • rFID using 100 images might not be credible enough. Please consider using 1000 or more image if rFID is a must.
Questions for authors

Questions

  • How is c,βc,\beta in Algorithm 1 determined? They are not in Table 3, 4.
  • See weakness

Justification

  • Overall I appreciate the strong and thorough validation on diverse tasks and dataset and recommand acceptance. However, many aspects of this paper remains unclear and should be improved. The proposed approach in Algorithm 1 is too complex compared with DPS baseline, with 5 more hyper-parameters. It becomes really hard to figure out whether the proposed approach's empirical advantage comes from the method itself or additional complexity. In ablation studies, the baseline method also fails to work (17 dB) and the contribution of component beyond Armijo step is unclear. I think this paper has too many new components and contributions, perhaps it would be better to reduce to 1-2 new part that is really effective.
2 comments
AuthorsOpenReview ↗

Part 1, Q1-Q2

Q1. Hyperparameter settings and overall method summary

Thank you for pointing out the need for a clearer account of the additional parameters. We added a “Method summary” paragraph at the end of Sec. 3.2 and expanded the reproducibility appendix to collect all the hyperparameter settings.

  1. Armijo parameters. We fix the sufficient-decrease coefficient to c=10−4c=10^{-4} and the backtracking shrinkage factor to β=0.5\beta=0.5 throughout the reported inverse-problem experiments, without task-specific tuning. Each search tests at most Jmax⁡=3J_{\max}=3 candidate scales. The correction scale γt\gamma_t is determined by the Armijo acceptance test, with the most recently accepted scale reused between scheduled searches. Thus, γt\gamma_t is computed during sampling rather than independently tuned.

  2. Overall parameter summary. After the definitions and derivations, the new “Method summary” paragraph in Sec. 3.2 brings together the two components of CAT and their additional hyperparameters. The tubular-budget coefficient ρ\rho is the main geometric hyperparameter requiring tuning and can be reused across multiple tasks sharing the same prior. The search period pp controls the frequency of Armijo evaluations and the efficiency–quality trade-off. The remaining Armijo settings, cc, β\beta, and Jmax⁡J_{\max}, are fixed as described above. This summary provides a single place to understand the roles and settings of the additional hyperparameters.

  3. Expanded reproducibility details. Appendix C now includes detailed settings for all hyperparameters used in our experiments.

Revision locations: Sec. 3.2, “Method summary,” page 5; Appendix C, pages 20–23. Revisions are highlighted in blue.

Q2. Implementation complexity and parameter tuning

Thank you for raising this concern about implementation complexity and parameter tuning. We reorganized the pseudocode to make the two components of CAT and their integration into a host sampler clearer, and summarized the additional hyperparameters in Sec. 3.2.

  1. Two explicit components. As described in the “Method summary,” CAT consists of a second-order curvature-aware tubular constraint that regulates geometric departure and Armijo backtracking that calibrates the correction scale. Algorithms 2 and 3 now describe these components separately. Algorithm 1 then presents their integration into the host sampling loop through two procedure calls, showing how CAT adjusts the measurement-guidance update after the host prior transition. This organization makes the overall procedure and the implementation of each component easier to follow.

  2. Limited additional tuning. The tubular-budget coefficient ρ\rho is the main additional hyperparameter requiring tuning and can be reused across multiple tasks sharing the same prior. We follow common Armijo backtracking settings with c=10−4c=10^{-4} and β=0.5\beta=0.5, and fix Jmax⁡=3J_{\max}=3. We use these values throughout the reported inverse-problem experiments without task-specific adjustment. The search period pp can simply be set to 1 to apply Armijo at every step, or increased to reduce computational cost. The scale γt\gamma_t is selected automatically by the acceptance test. The “Method summary” paragraph in Sec. 3.2 summarizes these parameter choices. Thus, the main tuning burden introduced by CAT is ρ\rho, which can be reused across tasks sharing the same prior, while the other parameters can be fixed with simple choices.

Revision locations: Sec. 3.2, “Method summary,” page 5; Appendix C, Algorithms 1–3, page 21, and Tables 3–5, pages 22–23. Revisions are highlighted in blue.

AuthorsOpenReview ↗

Part2, Q3-Q6

Q3. Ablations with linear operators and component contributions

Thank you for suggesting ablations with linear operators. We clarified the shared-initial-step protocol and expanded Table 2 with DPS Gaussian deblurring and HDR reconstruction. The results show task-dependent contributions from Armijo and geometric control, with their combination achieving the best results across the evaluated ablation settings.

  1. Why the host performs poorly in the ablations. Within each task, we fix the initial guidance step size across variants so that each variant adjusts the direction and step size through its own mechanism. The unmodified host collapses at this shared step size because it lacks adaptive direction adjustment and step-size control. This protocol isolates the effects of the added components. Table 1 separately reports comparisons using the respective guidance settings of the host and Host+CAT.

  2. Contributions of Armijo and geometric control. The expanded ablations show that Armijo provides substantial gains in deblurring, while second-order tubular control substantially improves HDR reconstruction even without backtracking. Their relative contributions vary across tasks, and combining them gives clearly stronger overall results. These ablations aim to isolate the contribution of each component and demonstrate the benefit of combining them across different tasks, with less emphasis on comparing their relative gains, which may depend on the task.

  3. Linear operators and super-resolution. Following your suggestion, we added DPS Gaussian-deblurring ablations, a common linear task, alongside the PSLD motion-deblurring ablation. We selected Gaussian deblurring because it more clearly distinguishes the effects of the components. The differences between variants on super-resolution are small, making it less informative for this component analysis. Appendix D.2, Table 7, includes Host+Armijo versus Host+CAT on super-resolution for both DPS and PSLD. We hope these results and clarifications address your concern.

Revision locations: Sec. 4.3, Table 2, page 9 (clarified protocol, additional ablations, and discussion of component contributions). Appendix D.2, Table 7, page 25 (Armijo–CAT comparisons including super-resolution). Revisions are highlighted in blue.

Q4. Sampling steps for vanilla DPS

Thank you for requesting this clarification. The vanilla DPS baseline uses N=1000N=1000 sampling steps. We specify the step counts in the table and added DPS runs with N=3000N=3000, 50005000, and 1000010000 for comparison.

Revision locations: Appendix D.4, Table 9, page 27 (originally Table 6 in the reviewed manuscript). Revisions are highlighted in blue.

Q5. Comparisons with MCG and DSG

Thank you for suggesting these additional comparisons. We added comparisons with MCG and DSG on linear inverse problems, where CAT achieves better perceptual quality in the evaluated settings.

  1. Additional baselines and tasks. We compare DPS+CAT with MCG on random-mask inpainting, and with DSG on random-mask inpainting, Gaussian deblurring, and 4× super-resolution. We chose the three operators for which its original paper recommends parameter settings, and use the respective papers' recommended settings for both baselines. All comparisons follow the FFHQ evaluation protocol with 100 images and report PSNR, SSIM, LPIPS, and DISTS.

  2. Perceptual quality. DPS+CAT achieves better perceptual quality than MCG and DSG in the evaluated comparisons, as reflected by lower LPIPS and DISTS.

Revision locations: Appendix D.1, pages 23–24. Revisions are highlighted in blue.

Q6. Reliability of rFID with 100 evaluation images

Thank you for raising this concern about the reliability of rFID with 100 images. We replaced FID with DISTS in the FFHQ and ImageNet inverse-problem evaluations to better assess reconstruction perceptual quality.

  1. Evaluation focus. We agree that FID estimated from 100 images cannot reliably establish distribution-level quality. Our conclusions primarily concern the perceptual fidelity of individual reconstructions, and FID was previously included only as an auxiliary metric. This change does not affect our main conclusions.

  2. Updated metrics. DISTS measures perceptual similarity to each reference image, complementing LPIPS without distribution-level estimation. We report DISTS and LPIPS alongside PSNR and SSIM to assess both perceptual fidelity and reconstruction accuracy.

  3. Results. The results show substantial perceptual improvements over competing methods across many evaluated settings. As a plugin, CAT also broadly improves both reconstruction accuracy and perceptual fidelity relative to the original host methods.

Revision locations: Sec. 4.1, pages 5-6 (evaluation metrics and main results); Table 1, page 6; Sec. 4.3, Table 2, page 9; Appendix D.4, Table 9, page 27. These tables all report DISTS. Revisions are highlighted in blue.

Reviewer_a2xwOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper addresses the problem that in gradient-guided diffusion samplers the measurement correction moves outside the region where the learned score is reliable. It places a noise-dependent tube around the iso-density surface and allocates the thickness budget between normal displacement (first order) and tangent displacement (second order, according to the directional curvature). The allocation is uniquely determined by a scalar dual equation, and Armijo backtracking calibrates the finite step against the actual guidance objective. No training is required, and it applies to both pixel-space and latent-space samplers.

Critical strengths
  1. Without training it attaches to several host samplers (DPS, PSLD, ReSample), and the perceptual metrics improve consistently across the seven tasks overall. Working across solver families is a meaningful property for a plugin.
  2. In Fig. 3, performance is maintained as the guidance step size ω\omega is increased, whereas the comparison methods collapse sharply. Making a previously sensitive parameter insensitive has practical value.
  3. That the allocation problem has a unique solution and is solved by scalar bisection, and that the curvature is evaluated in a single backward pass without the full Jacobian, are clean design choices.
Critical weaknesses
  1. Throughout the tables and figures in the main text, the computational cost is neither matched nor reported. According to Table 6, the representative setting in Table 1 (N=1000) takes 132 seconds per image, about 3.5× the 38 seconds of DPS. The authors show that the gain is not due to increased computation by pointing out that at N=200 the method takes 26 seconds, faster than DPS, while achieving higher PSNR but this comparison is performed for only one combination: FFHQ nonlinear deblurring with DPS as the host. The remaining six tasks in Table 1, the PSLD and ReSample hosts, and the sweeps in Figs. 3 and 4 contain no computational cost information at all. The comparison methods DAPS, DDNM, ReSample and others also have very different cost structures, and with no time column in Table 1 it is impossible to judge whether the quality differences reflect cost differences.
  2. Host+Armijo is not evaluated as a proper baseline. Table 2 shows that adding Armijo backtracking alone raises DPS nonlinear deblurring from 17.06 to 24.08 and PSLD motion deblurring from 17.96 to 25.74. Since Full CAT reaches 26.30 and 27.05 respectively, most of the total gain comes from the line search. Broken down further, First-order CAT which adds the normal/tangent decomposition and the first-order tube reaches 24.26 / 25.80, nearly the same as Host+Armijo, and the remaining gain comes from the curvature term. That is, the only part of this paper's geometric machinery that actually contributes is the second-order curvature term. Yet Host+Armijo appears only as a single row in a non-standard guidance setting where Host collapses, and is absent from both Table 1 and Fig. 3.
  3. The ρ\rho sweep is too narrow in range to support its conclusion. By Theorem 3, as ρ→∞\rho \to \infty we have λt=0\lambda_t = 0, so the proposed method coincides exactly with host+Armijo. However, Fig. 4 stops at ρ∈[0.1,2]\rho \in [0.1, 2] and nowhere in this interval does it reach that limit. The sweep should be extended to show whether it converges to host+Armijo at the upper end and where quality collapses at the lower end, before the claim of "robustness over a broad range" holds. Whether this behavior and the optimal value are the same across models and types of measurement noise is also not verified.
  4. There is no analysis whatsoever of K^t\hat K_t, which is central to the method. The curvature comes from the Jacobian of the learned score and governs the entire budget allocation, yet it is not reported how it varies with sampling step, or how it differs across models or types of measurement noise. The same applies to the fraction of steps where the constraint is actually active (λt>0\lambda_t > 0). Without these numbers there is no way to confirm that the geometric mechanism operated, and since Eq. (10) converges to First-order CAT when K^t→0\hat K_t \to 0, performance degradation would not be noticeable even if the curvature estimate were inaccurate.
Questions for authors
  1. Could you present comparisons at matched computational cost for the main tasks in Table 1? Please also state whether the computational cost was matched across methods in the sweeps of Figs. 3 and 4.
  2. Could you report host+Armijo in the setting of Table 1 and add the corresponding curve to Fig. 3?
  3. Does increasing ρ\rho substantially converge to host+Armijo performance? By Theorem 3 it should.
  4. Could you report the distribution of K^t\hat K_t and the fraction of active constraints by sampling step, model, and type of measurement noise?
5 comments
AuthorsOpenReview ↗

Part 1, Q1

Q1. Comparisons at matched computational cost

Thank you for raising this question about computational fairness. We clarify the matched-step protocol used for the plugin comparisons and have expanded the efficiency study with longer DPS runs showing that CAT retains better reconstruction quality even when DPS receives more runtime.

  1. Comparison protocol in Table 1. The compared solvers differ in their modeling, speed–accuracy priorities, and algorithmic constraints, making a single runtime or sampling-step budget difficult to impose uniformly across solver families. Within each host/CAT pair in Table 1, we keep the actual number of optimization steps the same. Since CAT is designed as a reusable plugin, this paired protocol provides a controlled comparison of its effect when integrated into the same host.

  2. Sweeps in Figs. 3 and 4. Figure 3 uses the same number of optimization steps for PSLD, PSLD+DiffState, and PSLD+CAT. Figure 4 likewise holds the step count fixed across CAT radius settings and uses the corresponding PSLD baseline as a reference. These sweeps match optimization steps to assess the effects of the plugins and their settings.

  3. Additional runtime comparison. We expanded the efficiency experiment on FFHQ nonlinear deblurring with longer DPS runs. DPS with N=5000N=5000 takes 164 seconds per image, compared with 132 seconds for DPS+CAT with N=1000N=1000. Despite receiving more runtime, DPS obtains PSNR/SSIM of 23.54/0.6861 and LPIPS/DISTS of 0.2845/0.1507, while DPS+CAT achieves 26.30/0.7623 and 0.2197/0.1157. Increasing DPS further to N=10000N=10000 takes 329 seconds and worsens all four metrics relative to N=5000N=5000. This behavior is consistent with semi-convergence in noisy inverse problems, where additional iterations can eventually degrade reconstruction quality [1]. In this experiment, extending host sampling to comparable or greater runtime does not recover the quality gains provided by CAT. These results indicate that its quality gains stem from the improved guidance mechanism beyond simply increasing host computation.

  4. Efficiency–quality trade-off. The efficiency study shows that fewer sampling steps or less frequent Armijo searches reduce CAT's runtime while retaining improvements over the host. For example, DPS+CAT with N=200N=200 takes 26 seconds and improves all four quality metrics over DPS with N=1000N=1000, which takes 38 seconds. Thus, CAT can also improve reconstruction quality at a lower runtime.

Revision locations: Appendix D.4, pages 26–27. Revisions are highlighted in blue. The comparison protocols discussed above correspond to Table 1, page 6, and Sec. 4.2, Figs. 3–4, pages 7–8.

References: [1] Elfving, T., Nikazad, T., and Hansen, P. C. (2010). Semi-convergence and relaxation parameters for a class of SIRT algorithms. Electronic Transactions on Numerical Analysis, 37, 321–336.

AuthorsOpenReview ↗

Part 2, Q2

Q2. Host+Armijo comparisons

Thank you for requesting these comparisons. We added Host+Armijo to the guidance-strength sweep, expanded the component ablations to additional tasks, and included a dedicated Armijo–CAT comparison across several host–task settings. These results show that the benefit of geometric control varies across tasks and that Full CAT provides stronger performance and robustness across the evaluated settings.

  1. Purpose of the fixed-step ablation. Within each ablation task, we keep the initial guidance step size identical across variants so that each variant adjusts the direction and step size through its own mechanism. The unmodified host collapse at this step size because it lacks the additional control used by the variants. This setting helps isolate the contributions of those mechanisms. Table 1 separately compares the host and Host+CAT with their respective guidance settings. We revised Sec. 4.3 to clarify the fixed-initial-step protocol, since a step that is too large for the unmodified host can still be appropriate for the controlled variants.

  2. Contributions of the two components. We expanded Table 2 to include DPS Gaussian deblurring and HDR reconstruction, alongside DPS nonlinear deblurring and PSLD motion deblurring. The relative gains from Armijo and geometric control depend on the task, while their combination gives clearly stronger overall results in these ablations. These ablations aim to isolate the contribution of each component and demonstrate the benefit of combining them across different tasks, with less emphasis on comparing their relative gains, which may depend on the task.

  3. Interpretation of the first-order comparison. CAT combines a second-order tubular constraint with Armijo backtracking. The tubular constraint jointly accounts for first-order normal displacement and curvature-induced second-order tangent departure. Full CAT outperforms First-order CAT in the ablations, supporting the benefit of jointly controlling first- and second-order departure. We would like to clarify that this comparison alone does not establish that second-order control is the only effective component.

  4. Added robustness curve. Figure 3 now includes PSLD+Armijo on FFHQ motion deblurring over guidance weights ω∈{0.1,0.5,1,2,5}\omega\in\{0.1,0.5,1,2,5\}, using 100 images, γPSLD=0.01\gamma_{\mathrm{PSLD}}=0.01, and measurement noise σy=0.05\sigma_y=0.05. As guidance increases in the strong-guidance regime, PSLD+Armijo degrades across the metrics, while PSLD+CAT remains comparatively stable. The corresponding visual comparison also shows sharper facial details with CAT. This sweep extends the comparison beyond the single guidance setting used in the component ablation.

  5. Broader comparison accompanying Table 1. We understand the reviewer's request as a suggestion to compare Host+Armijo and Host+CAT under standard guidance settings. We would first like to clarify that the guidance step used in the ablations is large for the unmodified host, but can still be appropriate for the controlled variants. We revised the wording in Sec. 4.3 to clarify this distinction, and the ablations also provide evidence for the comparison between Armijo and CAT. We additionally report comparisons on five DPS tasks and four PSLD tasks on FFHQ, covering linear and nonlinear operators, with 100 images per task and evaluation using PSNR, SSIM, LPIPS, and DISTS. These results show that geometric control provides substantial gains for several measurement operators, especially nonlinear ones, with its contribution varying across tasks and host samplers. Since Table 1 already includes many established and recent competitive methods, we present this additional Armijo comparison as an extended experiment in Appendix D.2. We hope that these clarifications and additional results address the reviewer's concern.

Revision locations: Sec. 4.2, Fig. 3, page 7 (Host+Armijo robustness curve and discussion). Sec. 4.3, Table 2, page 9 (clarified protocol and expanded ablations). Appendix D.2, Table 7, page 25 (additional Armijo–CAT comparisons). Appendix E, Fig. 11, page 31 (qualitative guidance sweep). Revisions are highlighted in blue.

AuthorsOpenReview ↗

Part 3, Q3-Q4

Q3. Convergence to Host+Armijo at large tubular radii

We expanded the radius sweep in both directions. As the radius increases, PSLD+CAT indeed approaches PSLD+Armijo performance, consistent with the expected limiting behavior. We also revised the robustness claims to better reflect the observed variation.

  1. Larger radii and convergence toward Armijo. We added ρ∈{5,10,15,20}\rho\in\{5,10,15,20\} to the PSLD+CAT sweep on FFHQ motion deblurring, keeping the host settings fixed and evaluating the same 100 images. Larger radii progressively relax geometric control, and the results approach those of PSLD+Armijo. At ρ=20\rho=20, the metrics are close to the Armijo baseline under the same guidance setting (ω=5\omega=5), as shown below.
MethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓
PSLD+CAT (ρ=20\rho=20)25.930.70940.33670.2020
PSLD+Armijo25.740.69470.34820.2072
  1. Smaller radii. We additionally tested ρ=0.05\rho=0.05 and 0.010.01, so the smallest coefficient is now one hundredth of the default value ρ=1\rho=1. The full sweep covers ρ∈{0.01,0.05,0.1,0.5,1,1.5,2,5,10,15,20}\rho\in\{0.01,0.05,0.1,0.5,1,1.5,2,5,10,15,20\}. The distortion metrics remain stable at small radii and vary moderately across the full sweep, while perceptual gains are stronger at smaller radii.

  2. Revised claim. We apologize for the overly broad wording of the original claim of “robustness over a broad range.” We revised the corresponding statements to describe the observed behavior more precisely, including the moderate variation across the full sweep and the stronger perceptual gains at smaller radii.

Revision locations: Sec. 4.2, Fig. 4, page 8. Revisions are highlighted in blue.

Q4. Curvature distributions and constraint activity

We added distributions of the estimated curvature and second-order contribution, together with constraint activity rates, at several sampling steps for the FFHQ and ImageNet models.

  1. Reported statistics. For each model and selected sampling step, we pool all samples from the seven evaluated tasks, with 100 images per task under the main DPS+CAT settings. We report the mean and the 10th, 30th, 50th, 70th, and 90th percentiles of the estimated curvature K^t\widehat K_t and the second-order share χt\chi_t, which measures the fraction of total geometric consumption due to curvature-induced tangent departure. We also report the fraction of active constraints, defined by λt>0\lambda_t>0. Since the evaluated tasks use Gaussian measurement noise, we report these statistics by sampling step and model. We hope that these results address the reviewer's request and help clarify how the geometric mechanism operates during sampling.

  2. The geometric mechanism remains active. Constraints are active at every reported stage for a nonzero fraction of samples, with particularly high activity early in sampling and at the final step. At steps 1 and 999, the active fractions are 100.0% and 99.9% for FFHQ, and 84.7% and 70.3% for ImageNet. Curvature-induced second-order departure also accounts for a substantial share of total geometric consumption, with particularly large contributions early in sampling and at the final step. The mean second-order shares at these steps are 99.8% and 94.1% for FFHQ, and 79.5% and 67.8% for ImageNet. These results confirm that the geometric mechanism operates during sampling and that curvature-induced departure makes a substantial contribution, especially at these stages.

  3. Consistency with the first-order ablation. The existing comparison between Full CAT and First-order CAT in Table 2 shows improvements across all four metrics for both evaluated hosts. Together with the diagnostics, this indicates that CAT does not effectively reduce to its first-order variant in these experiments and that the contribution of second-order curvature is consequential for reconstruction quality.

Revision locations: Sec. 4.3, page 9 (curvature and constraint activity discussion, alongside the existing ablation in Table 2); Appendix D.3, Table 8, page 26 (distributions and activity rates). Revisions are highlighted in blue.

Reviewer_y6tbOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents an approach for inference time steering of diffusion models. For this they introduce Curvature Adaptive Tubular correction (CAT). CAT decomposes the guidance to normal and tangent components and defines adaptive scaling and clipping approcimations for these vectiors. Armijo backtracking is used to calibrate the resulting step updates against the actual guidance objective. The authors also establish local gurantees for the tubular approximation. Results on some benchmrks show better performance over exisitng methods

Critical strengths
  1. The idea of decomposing the guidance to on manifold and tangential components is interesting and finding adaptive guidance scales is an interesting perspective.
  2. The authors have attempted to provide sufficeint ablations for their design choices.
Critical weaknesses
  1. All the evaluations are perfomed using 100 images from FFHQ and ImageNet. A reliable FID score requires more than 5000 images . This acts as a major drawback for the approach and the experiments.
  2. The paper is written is a very convoluted manner and is hard to follow
  3. The method imposes sever latency contraints and slowing down for better performance.
  4. How does the approach compare for non linear tasks and for text to image model guidance
  5. The baselines are outdates and main comapriosn are with DPS, Resample, DDRM which are from 2023. There are more recent and much better performing works in this field which are not reviewed or considered [1,2,3]
  6. Rissanen, Severi, Markus Heinonen, and Arno Solin. "Free hunch: Denoiser covariance estimation for diffusion models without extra costs." International Conference on Learning Representations. Vol. 2025. 2025.

[1] Shin, Seunghyeok, et al. "Geometry-Correct Diffusion Posterior Sampling with Denoiser-Pullback Curvature Guidance and Manifold-Aligned Damping." arXiv preprint arXiv:2605.27990 (2026). [2] Huang, William, et al. "Constrained diffusion with trust sampling." Advances in Neural Information Processing Systems 37 (2024): 93849-93873.

Questions for authors

I believe the lack of comaparisons with newer approaches, scale of datasets used for evaluations and the convoluted writing makes the paper lack on the standard for acceptance.

2 comments
AuthorsOpenReview ↗

Part 1, Q1-Q3

Q1. Reliability of FID with 100 evaluation images

Thank you for highlighting small-sample FID reliability. We replaced FID with DISTS in the FFHQ and ImageNet inverse-problem evaluations to better assess reconstruction perceptual quality.

  1. Evaluation focus. We agree that FID estimated from 100 images cannot reliably establish distribution-level quality. Our conclusions primarily concern the perceptual fidelity of individual reconstructions, and FID was previously included only as an auxiliary metric. This change does not affect our main conclusions.

  2. Updated metrics. DISTS measures perceptual similarity between each reconstruction and its reference image, complementing LPIPS without requiring distribution-level estimation. We report DISTS and LPIPS alongside PSNR and SSIM to assess both perceptual fidelity and reconstruction accuracy.

  3. Results. The results show substantial perceptual improvements over competing methods across many evaluated settings. As a plugin, CAT also broadly improves both reconstruction accuracy and perceptual fidelity relative to the original hosts.

Revision locations: Sec. 4.1, pages 5-6 (evaluation metrics and main results); Table 1, page 6; Sec. 4.3, Table 2, page 9; Appendix D.4, Table 9, page 27. These tables all report DISTS. Revisions are highlighted in blue.

Q2. Clarity and organization of the paper

Thank you for pointing out our presentation difficulties. We revised the method section, pseudocode, and overview figure to make CAT easier to understand.

  1. CAT in simple terms. After the host sampler takes its prior step, CAT adjusts the proposed measurement-guidance update. It splits this update into motion along and perpendicular to a local equal-density surface, then adjusts both components to keep their combined displacement within a small neighborhood of that surface, accounting for its curvature. Finally, periodic backtracking selects a step size that sufficiently reduces the measurement loss. This preserves useful motion toward the measurements while controlling departure from the local prior geometry.

  2. Clearer method presentation. Section 3.2 now opens with a concise “CAT in practice” paragraph explaining the idea and the four operational steps before the definitions and derivations. After the definitions and derivations, Section 3.2 concludes with a “Method summary” paragraph covering the components and key hyperparameters, helping readers connect the formulation to its implementation.

  3. Simplified pseudocode and illustration. We reorganized the pseudocode into the overall Host+CAT sampling loop and two supporting procedures for curvature control and backtracking. We also revised the overview figure to show how a single CAT correction adjusts the host update within the tube and then calibrates its step size.

Revision locations: Sec. 3.2, page 3 (“CAT in practice”), page 5 (“Method summary”); Appendix C, Algorithms 1–3, page 21; Fig. 1, page 1. Revisions are highlighted in blue.

Q3. Latency and performance trade-off

Thank you for raising the concern about runtime overhead. We expanded the efficiency study with longer DPS runs, showing that CAT retains better reconstruction quality even when the host receives more runtime.

  1. Comparison at comparable or greater host runtime. On FFHQ nonlinear deblurring, DPS with N=5000N=5000 takes 164 seconds per image, compared with 132 seconds for DPS+CAT with N=1000N=1000. DPS obtains PSNR/SSIM of 23.54/0.6861 and LPIPS/DISTS of 0.2845/0.1507, while DPS+CAT achieves 26.30/0.7623 and 0.2197/0.1157. Increasing DPS to N=10000N=10000 takes 329 seconds and worsens all four metrics relative to N=5000N=5000. This behavior is consistent with semi-convergence in noisy inverse problems, where additional iterations can eventually degrade reconstruction quality [1]. Here, extending host sampling to comparable or greater runtime does not recover the gains provided by CAT. Although CAT requires more runtime at the same number of optimization steps, these results indicate that its quality gains stem from the improved guidance mechanism beyond simply increasing host computation.

  2. Efficiency–quality trade-off. The efficiency study shows that fewer sampling steps or less frequent Armijo searches reduce CAT's runtime while retaining improvements over the host. For example, DPS+CAT with N=200N=200 takes 26 seconds and improves all four quality metrics over DPS with N=1000N=1000, which takes 38 seconds. Thus, CAT can also improve reconstruction quality at a lower runtime.

Revision locations: Appendix D.4, pages 26–27. Revisions are highlighted in blue.

References: [1] Elfving, T., Nikazad, T., and Hansen, P. C. (2010). Semi-convergence and relaxation parameters for a class of SIRT algorithms. Electronic Transactions on Numerical Analysis, 37, 321–336.

AuthorsOpenReview ↗

Part 2, Q4-Q5

Q4. Nonlinear inverse problems and text-to-image guidance

Thank you for this question about the applicability of CAT. Our evaluations already include HDR reconstruction, nonlinear deblurring, and black hole imaging, and we added text-to-image experiments to examine robustness under strong guidance.

  1. Nonlinear inverse problems. On HDR reconstruction and nonlinear deblurring, DPS+CAT achieves the best LPIPS and DISTS among the evaluated pixel-space methods on both FFHQ and ImageNet. Nonlinear deblurring also shows strong reconstruction accuracy. These results support its effectiveness beyond linear measurement operators. Appendix D.5 also reports nonlinear black hole imaging on InverseBench, where DPS+CAT achieves the highest mean PSNR and Blur PSNR among all compared methods under both observation-time settings.

  2. Extension to text-to-image guidance. We apply CAT to the difference between the conditional and unconditional model predictions, using the unconditional score to define the local geometry. Since this implementation provides a guidance direction without an explicit measurement-loss objective for line search, we disable Armijo backtracking and retain the second-order curvature constraint. We compare CFG+CAT with standard CFG and APG using Stable Diffusion 2.1 on 1,000 COCO 2017 validation captions, with identical sampling settings, prompts, and initial seeds at guidance scales of 10, 20, 50, and 100.

  3. Robustness results. As guidance increases from 10 to 100, the observed FID of CFG+CAT changes only from 64.67 to 66.26, compared with 66.63 to 104.88 for CFG and 65.26 to 74.59 for APG. CAT also maintains stable saturation and RMS contrast, consistent with the more natural colors and smoother tonal transitions in the qualitative examples. These results indicate greater robustness to increasing guidance strength.

  4. Role of the 1,000-image FID evaluation. We would like to clarify that this experiment is intended to examine how generation quality changes with guidance strength, rather than to establish the absolute best FID across all settings. We believe that the observed differences under strong guidance, evaluated with the same 1,000-image protocol and supported by saturation, contrast, and visual comparisons, provide useful evidence of improved robustness.

Revision locations: Sec. 4.1, Table 1, page 6 (nonlinear-task results, with updated DISTS entries); Sec. 4.5, second paragraph, page 10; Appendix D.6, Table 11, and Figs. 8–9, pages 28-29 (text-to-image protocol, robustness results, and visual comparisons). Revisions are highlighted in blue.

Q5. Comparisons with recent baselines

Thank you for suggesting these relevant works. Our main comparisons include several recent methods, and we added comparisons with the two most recent works among the three references we understand you to be recommending.

  1. Recent baselines in the main evaluation. Table 1 includes DAPS (2025), DCDP (2026), DiffRGD (2026), and DiffState (2025). We evaluate CAT with DPS, PSLD and ReSample as host samplers to assess the improvements obtained by integrating CAT into existing methods.

  2. Interpretation of the suggested references. We noticed that this comment refers to three works, while the reference list contains entries [1] and [2], and review item 6 separately lists Free Hunch. We have therefore understood Free Hunch as the third suggested reference. On this basis, we added comparisons with the two most recent works, CLAMP (2026) and Free Hunch (2025, using its FH+online variant).

  3. Additional comparisons and results. We evaluate these methods on FFHQ random-mask inpainting, Gaussian deblurring, and 4× super-resolution using the same 100-image protocol and four metrics. DPS+CAT achieves the lowest LPIPS and DISTS among all methods in this additional comparison across all three tasks, showing clear perceptual improvements over CLAMP and FH+online. The qualitative inpainting examples further illustrate how CAT preserves facial details and hair textures, consistent with its perceptual metrics.

Revision locations: Sec. 4.1, page 5, and Table 1, page 6 (main recent baselines). Appendix D.1, pages 23–24 (additional comparisons and qualitative results). Revisions are highlighted in blue.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.