acceptodds
Under review as a conference paper at ICLR 2027

Expose, Then Commit: Diffusion Posterior Sampling

Abstract

Diffusion posterior sampling solves inverse problems by combining a pretrained diffusion prior with measurement-consistency guidance. However, at high noise levels, clean-estimate errors can compromise full-band guidance, while stochastic refinement can modify details weakly constrained by the measurements. We argue that posterior sampling should coordinate measurement-frequency exposure during refinement with correction commitment afterward. Based on this principle, we propose a posterior continuation framework that constructs a family of intermediate posteriors whose noise-conditioned likelihoods emphasize low-frequency residuals and gradually return to full-band consistency. We instantiate this framework with a stabilized sampler combining a diffusion predictor, frequency-weighted Langevin refinement, and a Haar-domain commitment rule that fully adopts coarse corrections while progressively admitting detail corrections relative to the predictor. Across super-resolution, inpainting, and deblurring, our method achieves competitive-to-state-of-the-art restoration performance, including PSNR gains exceeding 5 dB on motion deblurring against the strongest evaluated baselines.

Primary area:
generative models
Keywords:
Diffusion Model, Inverse Problem

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_HgpDOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This manuscript addresses the challenge of diffusion posterior sampling (DPS) in ill-conditioned inverse problems (such as severe blur and heavy downsampling). The authors propose a "posterior continuation" framework based on two coordinated principles:

  1. Noise-conditioned frequency exposure: An intermediate posterior likelihood schedule that restricts measurement consistency to low-frequency bands at high diffusion noise levels and progressively expands to full-band consistency as noise decreases.
  2. Post-refinement Haar-domain detail commitment: A multi-resolution commitment mechanism operating after inner Langevin refinement, which fully accepts coarse-scale corrections while gating high-frequency detail corrections relative to the diffusion predictor x^0,i\hat{x}_{0,i} using an exposure-linked schedule.

The proposed framework is evaluated across four linear inverse problems (4×4\times super-resolution, motion deblurring, Gaussian deblurring, and random/box inpainting) on FFHQ and ImageNet at 256×256256 \times 256, with further extensions to real-world blur and 512×512512 \times 512 latent diffusion (Stable Diffusion v1.5). The method demonstrates significant improvements over competitive baselines, notably outperforming prior art by over 5 dB5\text{ dB} PSNR on motion deblurring.

Critical strengths
  • Clear Conceptual Intuition:
    The conceptual separation between measurement exposure (which frequencies should guide intermediate refinement) and detail commitment (how much refined high-frequency detail should actually be retained relative to the prior) is well-reasoned and addresses a genuine failure mode in existing diffusion inverse solvers.
  • Compelling Performance:
    The empirical results across super-resolution, inpainting, and deblurring are consistently strong. The improvements are particularly striking on severely ill-conditioned tasks, where the proposed sampler beats established baselines by large margins while visibly eliminating the severe ringing and spurious frequency artifacts common in prior methods.
  • Thorough Ablation Analysis:
    The authors do a good job isolating the contributions of each individual design choice. The ablations convincingly demonstrate that frequency continuation in the loss and staged commitment in the wavelet domain are complementary, and that simpler alternatives (like pure frequency re-weighting without spatial wavelet gating) fail to achieve the same stability.
Critical weaknesses
  1. Substantial Computational Overhead from Inner Langevin Iterations:
    While the diffusion backbone's evaluation cost is kept relatively standard (NFE=1000\text{NFE} = 1000, via N=200N = 200 outer steps with 5 PF-ODE steps per iteration), the inner Langevin refinement requires T=200T = 200 steps at every outer stage. This means the sampler evaluates the measurement gradient ∇Li\nabla \mathcal{L}_i—involving the forward operator, its adjoint, and 2D FFT operations—40,00040,000 times per image. This results in heavy wall-clock sampling times (130 s130\text{ s} per image on FFHQ and 200 s200\text{ s} on ImageNet). While the authors show that a low-budget configuration (N=9,T=90N=9, T=90) is feasible (Tables 8 and 9), the reconstruction quality drops noticeably compared to the full setting.

  2. Theoretical Gap Between Diagonalization Assumptions and Practical Degradations:
    The core error-attenuation guarantees in Appendix A.5 (Corollaries A.5 and A.8) strictly require that the forward operator A\mathcal{A} and the spectral projection mask PωP_\omega are simultaneously diagonalizable in the Fourier basis. While this holds for circular convolutions/blur, it naturally does not hold for operators like spatial inpainting (where spatial masking mixes Fourier frequencies) or non-periodic boundary conditions. While the authors briefly acknowledge in Section A.3.2 that masking mixes frequencies, the paper currently lacks formal bounds or characterizations of gradient mismatch for operators that do not share the Fourier eigenbasis. Clarifying the theoretical boundaries of the framework would strengthen the presentation.

  3. Heuristic Dual-Domain Representation (Fourier vs. Wavelet):
    The method relies on two different frequency tools: it filters the measurement loss in the Fourier domain (Eq. 4–6), but then switches to Haar wavelets to update image details after refinement (Eq. 11–13). While the authors test a few wavelet types in Table 3, the paper does not clearly explain why it mixes both Fourier and wavelets instead of using one consistent representation. For example, is a wavelet used for the update step because it works locally on image patches and avoids the wavy "ringing" artifacts that Fourier filtering can cause? Adding some plain intuition for this design choice would make the paper much clearer.

Questions for authors
  1. Inpainting and Extension to Non-Linear Operators:
    As noted in Section A.3.2, inpainting breaks the theoretical assumption of simultaneous Fourier diagonalizability because the spatial mask mixes frequencies. Despite this, your empirical results on random and box inpainting are quite strong. Since the method appears robust to violations of this assumption, did you try applying your framework to non-linear forward operators A(x)\mathcal{A}(x)? (non-linear blur, etc.)

  2. Fairness of Budget & Pareto Analysis:
    In Table 1, the baseline methods use different sampling parameters (e.g., DAPS uses T=100T=100 Langevin steps, whereas your method uses T=200T=200).

    • (a) Matched Budget: If DAPS is evaluated with the exact same inner refinement budget (T=200T=200 steps), how much of the performance gap remains?
    • (b) Pareto Frontier: To better highlight the efficiency–quality trade-off, would it be possible to provide a Pareto curve (e.g., PSNR vs. Runtime or PSNR vs. Total Sampling Steps) comparing your method against a strong baseline (such as DAPS or DPS) on at least one representative task and dataset (e.g., FFHQ motion deblurring)? This would make the computational comparison much clearer.
  3. Implementation in Latent Diffusion (Stable Diffusion v1.5):
    In Section A.3.3 and Table 5, you apply the method to Stable Diffusion v1.5 at 512×512512 \times 512. Could you briefly clarify how the Fourier residual filtering and Haar commitment are executed? Are they applied directly inside the 64×64×464 \times 64 \times 4 latent space, or is the intermediate estimate decoded to pixel space at each Langevin step?

6 comments
AuthorsOpenReview ↗

Official Comment Part 1

Thank you for recognizing the design and experimental results. All manuscript references below refer to the revised manuscript.

1. Computational overhead and reduced-budget evidence (Weakness 1)

The default uses 1000 diffusion evaluations and 40,000 inner updates. In pixel space, the predictor is fixed within refinement, so inner steps do not require denoiser calls, but operator/adjoint evaluations and Fourier transforms still incur cost. Haar commitment occurs once per outer stage. Latent refinement additionally differentiates through the decoder. We therefore report runtime rather than equating NFE with total cost.

Appendix A.3.4 and Table 8 report intermediate budgets on FFHQ motion deblurring. At (N,T)=(200,100)(N,T)=(200,100), our method achieves 36.86/0.945/0.050 in approximately 110 s/image; at (200,50)(200,50) it achieves 35.81/0.943/0.047 in 70 s/image, versus 36.69/0.940/0.054 in 130 s/image for the default. These are useful operating points beyond the aggressive low-budget settings in Tables 12–13, not a claim of universal Pareto dominance. The revised discussion of computational cost and the supplementary quality–efficiency comparison appear in Sec. 4.2; Appendices A.4.2–A.4.3 distinguish NFE from measurement/refinement costs and specify the separate Figure 5 protocol and timing convention.

2. Matched DAPS budget and quality–runtime comparison (Weakness 1; Question 2(a)–(b))

(a) Matched budget (Question 2(a)). At N=200,T=100N=200,T=100, our Table 8 result is 36.86 dB, compared with 29.66 dB for DAPS in Table 1: a 7.20 dB FFHQ motion-deblurring difference at equal update counts. The tables below additionally report DAPS at T=200T=200. All configurations use N=200N=200 and 1000 NFE; T=100/200T=100/200 corresponds to 20,000/40,000 inner updates. Equal counts do not imply equal runtime. Appendix A.4, “DAPS”, specifies the N=200,T=100N=200,T=100 baseline. The additional DAPS T=200T=200 results below are reported in this response; they are not tabulated in the current revised manuscript.

FFHQ:

MethodSR ×4 PSNR↑SSIM↑LPIPS↓Motion PSNR↑SSIM↑LPIPS↓Gaussian PSNR↑SSIM↑LPIPS↓Inpainting PSNR↑SSIM↑LPIPS↓Runtime (s/image)
DAPS (T=100)29.350.7820.19329.660.8470.15729.190.8170.16530.720.8000.159~80
DAPS (T=200)28.200.7040.25530.620.7790.16528.820.7310.21530.240.7740.175~100
Ours (T=200)31.880.8900.09036.690.9400.05429.990.8210.17532.860.9090.066~130

ImageNet:

MethodSR ×4 PSNR↑SSIM↑LPIPS↓Motion PSNR↑SSIM↑LPIPS↓Gaussian PSNR↑SSIM↑LPIPS↓Inpainting PSNR↑SSIM↑LPIPS↓Runtime (s/image)
DAPS (T=100)25.470.6370.29427.860.7660.19625.890.6580.26828.440.7750.135~160
DAPS (T=200)24.130.5260.34224.960.5590.26624.230.5280.35126.560.6750.217~185
Ours (T=200)27.670.7650.17034.700.9350.15526.020.6720.28328.600.8870.042~200

At T=200T=200, our method exceeds DAPS on all three metrics across the four tasks on both datasets. Increasing DAPS's inner budget does not consistently improve its results. We retain the T=100T=100 comparison, including its lower Gaussian-deblurring LPIPS, rather than selecting only the less favorable baseline configuration.

The DAPS-at-100 and default-method runtimes come from Table 1, whose caption specifies motion deblurring; they are not separate timing measurements for all four tasks. The supplementary curves below use a jointly re-evaluated benchmark with different hardware, image selection, and timing/batching conventions. Their absolute timings should not be combined with those of Tables 1 and 8.

(b) Quality–runtime comparison (Question 2(b)). Section 4.2 and Figure 5 compare our method with DAPS and DPS. Appendix A.4.2 specifies the shared 100-image FFHQ-256 subset, 61×6161\times61 motion-blur kernel with intensity 0.5, and observation noise σy=0.05\sigma_y=0.05. Appendix A.4.3 adds the execution budgets, NFE accounting, and timing/batching details for the H200 runs. This supplementary benchmark is separate from Tables 1 and 8.

AuthorsOpenReview ↗

Official Comment Part 2

For both our method and DAPS, the standard configuration uses five PF-ODE denoiser evaluations per outer stage, corresponding to 5N5N predictor evaluations; the inner budgets are T=200T=200 and T=100T=100, respectively. For the supplementary comparison in Figure 5, DAPS uses N∈{25,50,100,200,400}N\in\{25,50,100,200,400\}, and DPS uses K∈{100,250,500,1000}K\in\{100,250,500,1000\} respaced DDPM steps. Our method is evaluated at multiple computational budgets, and each plotted point retains its recorded total sampling NFE rather than being inferred from a shared outer-stage grid. In particular, the 36.91 dB result uses 1200 NFEs, comprising 200 additional network evaluations relative to the 1000-NFE configuration (a 20% increase); it is not an equal-NFE comparison with the 1000-NFE baselines. The 33.96 dB result retains its reported budget of 150 sampling NFEs. The sweep specifies seed 42 and one run per budget. Appendix A.4.3 separately discloses untimed warm-up calls retained in some exported DPS counts; the main-text NFE comparisons use sampling-only counts.

Runtime is batch-amortized sampling cost in GPU-seconds per image, obtained by summing synchronized sampling times across GPU workers and dividing by the total reconstructed-image count. It includes sampling-region work but excludes model loading, observation/initial-noise generation, warm-up, and final metric evaluation. The method-specific batching configurations and the distinction from isolated-image or parallel end-to-end latency are stated in Appendix A.4.3.

Within this benchmark, our method achieves 33.96 dB at 150 sampling NFEs, compared with the best evaluated DAPS result of 31.84 dB at 2000 NFEs and DPS result of 27.31 dB at 1000 NFEs. It reaches 36.70, 36.91, and 36.93 dB at 27.0, 56.0, and 117.0 s/image, respectively. These results support the quality–cost advantage over the evaluated configurations under the disclosed execution conditions, not a universal runtime ordering or a repeated-run significance claim.

3. Theory beyond simultaneous Fourier diagonalization (Weakness 2; Question 1 — inpainting)

The shared-basis assumption supports the frequency-wise corollaries, not execution of the sampler. We have added the following fixed-stage results to Appendix A.6.1 to clarify what remains valid without it. All vector and operator norms are Euclidean and spectral norms.

General linear operators. Fix A,y,βi>0A,y,\beta_i>0 and the scheduled weights. Let Pi=Pi⊤=Pi2P_i=P_i^\top=P_i^2 be the real residual-space projector induced by the conjugate-symmetric Fourier mask, including inverse transformation. Define

Qi=(1−λi)I+λiPi,Ei(x)=(Ax−y)⊤Qi(Ax−y)2βi2,0≤λi≤1.Q_i=(1-\lambda_i)I+\lambda_iP_i,\qquad E_i(x)=\frac{(Ax-y)^\top Q_i(Ax-y)}{2\beta_i^2},\qquad 0\leq\lambda_i\leq1.

For the same pair of clean estimates with difference δi\delta_i, differentiation gives

ei=∇Ei(x^0,i)−∇Ei(x0,i∗)=A⊤QiAδiβi2e_i = \nabla E_i(\hat{x}_{0,i}) - \nabla E_i(x_{0,i}^{\ast}) = \frac{A^{\top}Q_i A\delta_i}{\beta_i^2}

Since 0⪯Qi⪯I0\preceq Q_i\preceq I,

∥ei∥2≤∥A⊤QiA∥2∥δi∥2βi2≤∥A∥22∥δi∥2βi2,0⪯∇2Ei⪯A⊤Aβi2.\lVert e_i\rVert_2\leq\frac{\lVert A^\top Q_iA\rVert_2\lVert\delta_i\rVert_2}{\beta_i^2}\leq\frac{\lVert A\rVert_2^2\lVert\delta_i\rVert_2}{\beta_i^2},\qquad 0\preceq\nabla^2E_i\preceq\frac{A^\top A}{\beta_i^2}.

Equations (61)–(62) provide these identities and bounds. The matrix bound follows from (Av)⊤Qi(Av)≤∥Av∥22(Av)^\top Q_i(Av)\leq\lVert Av\rVert_2^2. It applies to non-periodic blur but bounds worst-case sensitivity, not pointwise improvement for every error direction. The score-error form additionally requires δi=σi2Δsi\delta_i=\sigma_i^2\Delta s_i.

Binary masks. For zero-filled inpainting, A=M=M⊤=M2A=M=M^\top=M^2. Because MM and QiQ_i are contractions,

∥ei∥2=∥MQiMδi∥2βi2≤∥Mδi∥2βi2=∥efull∥2.\lVert e_i\rVert_2=\frac{\lVert MQ_iM\delta_i\rVert_2}{\beta_i^2}\leq\frac{\lVert M\delta_i\rVert_2}{\beta_i^2}=\lVert e_{\mathrm{full}}\rVert_2.

Proposition A.13, Eqs. (63)–(64), proves non-increase of total mismatch for box or random masks without commutation, but not strict or per-frequency improvement. Also, (I−M)∇Ei(x)=0(I-M)\nabla E_i(x)=0: weighting does not create measurements of missing pixels.

For fixed A,Qi,βiA,Q_i,\beta_i and finite-second-moment errors, with Bi=A⊤QiAB_i=A^\top Q_iA and Ci=E[δiδi⊤]C_i=\mathrm{E}[\delta_i\delta_i^\top],

E∥ei∥22=tr⁡(BiCiBi)βi4.\mathrm{E}\lVert e_i\rVert_2^2=\frac{\operatorname{tr}(B_iC_iB_i)}{\beta_i^4}.

Lemma A.14 and Eqs. (65)–(66) establish this second-moment identity, including conditioning on a random mask. It follows by expanding the quadratic form. CiC_i includes bias; no isotropy is required. The mask inequality also holds in expectation, including over random masks. These results compare the same estimates, not different trajectories or final PSNR. The empirical inpainting evidence is reported in Table 1 for random masks and Appendix A.3.2, Table 5, and Figure 9 for box masks.

AuthorsOpenReview ↗

Official Comment Part 3

4. Fourier weighting versus Haar commitment (Weakness 3)

The revised Secs. 3.2–3.3 describe their respective roles, and Appendix A.6.3 adds the correction-magnitude and reconstruction-risk analysis below. Fourier weighting controls residual frequencies on the measurement grid, including the lower-resolution grid for super-resolution. Haar commitment controls image-space corrections in localized coarse/detail coordinates.

For an orthonormal Haar transform WW with compatible normalization and boundary handling, let

Δi=x0,iref−x^0,i\Delta _{i} = x _{0,i}^{\mathrm{ref}} - \hat{x} _{0,i}

Ki=W⊤diag⁡(Ic,giId)WK _{i} = W^{\top} \operatorname{diag}(I _{c}, g _{i} I _{d})W

gi=gi,d∈[0,1]g _{i} = g _{i,d} \in [0,1]

Then x0,ifuse=x^0,i+KiΔix_{0,i}^{\mathrm{fuse}}=\hat{x}_{0,i}+K_i\Delta_i, and orthonormality gives

∥KiΔi∥22=∥Δi,c∥22+gi2∥Δi,d∥22≤∥Δi∥22.\lVert K_i\Delta_i\rVert_2^2=\lVert\Delta_{i,c}\rVert_2^2+g_i^2\lVert\Delta_{i,d}\rVert_2^2\leq\lVert\Delta_i\rVert_2^2.

Proposition A.16, Eqs. (78)–(79), formalizes this non-expansiveness result. It limits correction magnitude, not ground-truth error. For fixed pre-fusion distributions with finite second moments, predictor detail error ai=(W(x^0,i−xgt))da_i=(W(\hat{x}_{0,i}-x_{\mathrm{gt}}))_d and correction bi=(WΔi)db_i=(W\Delta_i)_d satisfy

E∥ai+gbi∥22=E∥ai∥22+2g E⟨ai,bi⟩+g2E∥bi∥22.\mathrm{E}\lVert a_i+gb_i\rVert_2^2=\mathrm{E}\lVert a_i\rVert_2^2+2g\,\mathrm{E}\langle a_i,b_i\rangle+g^2\mathrm{E}\lVert b_i\rVert_2^2.

Proposition A.17, Eqs. (80)–(81), gives this risk identity and the corresponding fixed-stage oracle gate. The identity requires no independence assumption and explains the balance between suppressing uncertain changes and retaining useful corrections. It does not make Eq. (13) globally optimal.

Haar locality motivates the representation but does not guarantee artifact-free images or prove Haar uniquely optimal. Sec. 4.3, Table 2, and Figure 8 separate the two interventions. Table 3 compares the tested bases and a measurement-reweighting alternative; Appendix A.3.5 and Figure 12 provide qualitative comparisons of the wavelet bases. That alternative is not post-refinement Fourier gating, so we do not claim to exclude all Fourier-only designs.

5. Extension to blind bilinear inverse problems (Question 1 — nonlinear operators)

Did the authors apply the framework to nonlinear forward operators?

Yes: Appendix A.3.3 and Figure 10 report blind-deblurring examples on FFHQ and ImageNet under the BlindDPS setting (Chung et al., 2023a). Both the image and blur kernel are unknown during reconstruction. The ground-truth kernel is used to generate the observation, not supplied to the reconstruction algorithm. The revised text explicitly states the unknown-kernel protocol in Appendix A.3.3 and the Figure 10 caption.

The forward map is B(x0,k)=k∗x0B(x_0,k)=k*x_0. It is linear in either variable separately but nonlinear in the joint variable ϑ=(x0,k)\vartheta=(x_0,k). In particular,

B(x+u,k+v)=B(x,k)+B(u,k)+B(x,v)+B(u,v)B(x+u,k+v)=B(x,k)+B(u,k)+B(x,v)+B(u,v)

The cross term distinguishes the joint blind problem from known-kernel linear deblurring. The experiment uses the official BlindDPS motion-blur setting: 64×64 kernels, intensity 0.5, and Gaussian measurement noise σy=0.02\sigma_y=0.02 in normalized image coordinates. Exposure weights the residual computed with the estimated kernel; commitment controls the image correction after refinement. These operations and the joint nonlinearity are explained in Appendix A.3.3. Figure 10 provides qualitative comparisons with BlindDPS and ground truth.

Connection to the nonlinear analysis (Appendix A.6.4, Proposition A.18, Eqs. (82)–(84)). Appendix A.3.3 already identifies the applicability of this local result to the bilinear map; the following joint-variable calculation makes that specialization explicit. For fixed stage parameters, let PiP_i be the residual-space orthogonal frequency projector and define

Qi=(1−λi)I+λiPi⊤Pi,Eiblind(ϑ)=(B(ϑ)−y)⊤Qi(B(ϑ)−y)2βi2.Q_i=(1-\lambda_i)I+\lambda_iP_i^\top P_i,\qquad E_i^{\mathrm{blind}}(\vartheta)=\frac{(B(\vartheta)-y)^\top Q_i(B(\vartheta)-y)}{2\beta_i^2}.

The joint measurement gradient is

∇ϑEiblind(ϑ)=JB(ϑ)⊤Qi(B(ϑ)−y)βi2,JB(x0,k)[δx,δk]=k∗δx+δk∗x0.\nabla_{\vartheta}E_i^{\mathrm{blind}}(\vartheta)=\frac{J_B(\vartheta)^\top Q_i(B(\vartheta)-y)}{\beta_i^2},\qquad J_B(x_0,k)[\delta x,\delta k]=k*\delta x+\delta k*x_0.

Finite-dimensional convolution is bilinear, so its Jacobian is Lipschitz and is bounded on bounded neighborhoods. With a bounded residual on such a convex neighborhood, Proposition A.18 in Appendix A.6.4 applies to BB and yields

∥∇Eiblind(ϑ)−∇Eiblind(ϑ′)∥2≤∥Qi∥2βi2(MJ2+LJRy)∥ϑ−ϑ′∥2.\lVert\nabla E_i^{\mathrm{blind}}(\vartheta)-\nabla E_i^{\mathrm{blind}}(\vartheta')\rVert_2\leq\frac{\lVert Q_i\rVert_2}{\beta_i^2}(M_J^2+L_JR_y)\lVert\vartheta-\vartheta'\rVert_2.

Here, MJM_J bounds the Jacobian norm, LJL_J is its Lipschitz constant, and RyR_y bounds the residual. This is a local sensitivity statement, not a convergence or reconstruction-quality guarantee. Figure 10 supplies the empirical evidence for this blind extension. The fixed-linear Fourier-wise results are not transferred unchanged to the joint problem.

Reviewer_9pEjOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

The paper separates measurement guidance in diffusion posterior sampling into "which frequencies to expose" and "how much of the refinement result to adopt." The starting point is the observation that at high noise levels the high-frequency components of the clean estimate x^0\hat x_0 have low SNR and non-negligible score-induced error, whereas the low-frequency structure is more stable and preserves the global geometry. Since measurement corrections computed from an inaccurate x^0\hat x_0 can be misaligned with the local posterior geometry, the paper constructs a family of intermediate posteriors that emphasize low-frequency residuals according to the noise level and gradually return to full-band. Then, after frequency-weighted Langevin refinement, the coarse components are adopted entirely in the Haar domain while the detail components are admitted gradually through a gate. The proposed method is evaluated on four linear inverse problems.

Critical strengths
  1. The logical connections are clear. A corresponding theorem is placed alongside each claim, and the paper is written so that one can trace which design choice follows from which result. In particular, the flow from the problem statement that "frequency weighting alone leaves the detail components weakly constrained" to the introduction of Haar commitment is natural.
  2. The results are strong on motion deblurring, super-resolution, and random inpainting. FFHQ motion deblurring at 36.69 dB is more than 5 dB above the runner-up (31.54), and the margin is similar on ImageNet.
  3. The ablation in Table 2 separates the roles of the two components. Removing frequency continuation degrades drift and removing Haar fusion degrades detail quality, showing that the two mechanisms prevent different failures. The comparison against other wavelet bases and the frequency reweighting alternative in Table 3 also supports the design choices.
Critical weaknesses
  1. The levels of rigor in the problem formulation and in the method differ greatly. Sections 3.1-3.2 are developed densely with spectral characterization and theorems, yet the components that actually determine the algorithm are introduced without justification. There is no explanation of the functional forms of the cutoff ω(σ)\omega(\sigma) and the mixing weight λ(σ)\lambda(\sigma), of why Eq. (13) uses the power form unlock(i)=ds+(de−ds)ρiγ\mathrm{unlock}(i) = d_s + (d_e-d_s)\rho_i^\gamma, or of why the linear progress variable ρi=(N−i)/(N−1)\rho_i = (N-i)/(N-1) is used. The principle of "admitting detail corrections gradually" is convincing, but that principle does not require this particular functional form. It reads as if the theory justifies the design after the fact rather than having derived it.
  2. The ImageNet Gaussian deblurring result is not convincing. At PSNR 26.02, SSIM 0.672, and LPIPS 0.283, it is among the worst on all three metrics. In particular DDRM obtains PSNR 26.33, SSIM 0.713, and LPIPS 0.243 with NFE 20 and 5 seconds, while the proposed method falls behind on every metric despite using NFE 1000 and 200 seconds. DDNM (NFE 100) leads by 2 dB at 28.06. The paper only remarks that "several baselines are competitive on ImageNet Gaussian deblurring, but their advantages do not transfer to other degradations," without analyzing the cause. Since Gaussian blur is a typical low-pass operator, it would be expected to interact directly with the frequency exposure design, so an explanation is needed for why the method falls behind in precisely this setting despite using 50× the NFE.
  3. The notation impedes reading. In Section 3.3 the equation zd∣(y,zc)z_{\mathrm{d}} \mid (y, z_{\mathrm{c}}) appears before c (coarse) and d (detail) are defined in the following paragraph. In addition, the commitment gate wi,dw_{i,\mathrm{d}} and the cutoff ωi\omega_i are not visually distinguishable, which causes confusion when reading back and forth between Eq. (8) and Eq. (13). This is more pronounced because both symbols correspond to the two axes of this paper.
Questions for authors
  1. Could you analyze the cause of the low performance on ImageNet Gaussian deblurring?
  2. How were the schedules for ω(σ)\omega(\sigma) and λ(σ)\lambda(\sigma) determined? Were they chosen heuristically, or found by search? Does the spectral analysis in Section 3.2 yield a particular functional form?
  3. Is there a reason why unlock(i)\mathrm{unlock}(i) in Eq. (13) must take the form ρiγ\rho_i^\gamma? If the only requirement is monotone increase, how does performance change when it is replaced by a linear or logistic schedule?
  4. Parameter sensitivity is reported only on FFHQ motion deblurring. Does the same trend hold for other datasets or measurement operators?
7 comments
AuthorsOpenReview ↗

Official Comment Part 1

Thank you for your review. All section, equation, table, and figure references below refer to the revised manuscript.

1. Schedule design and parameter selection (Weakness 1; Question 2)

The theory motivates exposure and commitment; it does not uniquely prescribe their schedules. Corollary A.8 (Appendix A.5) characterizes a specified projector and mixing weight at a fixed stage. The numerical defaults were selected through empirical parameter search balancing quality and cost, then held fixed across datasets and degradations unless otherwise stated, without per-image or per-baseline tuning. Appendix A.4 (“Hyper-parameters in Alg. 1”) documents selecting the default once and keeping it fixed; the empirical-search origin is clarified here. Section 4.3 (“Unlocking schedules”) explicitly identifies the power-law form as an empirical design choice.

Appendix A.3.4 reports budget sensitivity in Tables 8–11 and continuation/fusion sensitivity in Tables 18–21. Table 22, referenced by Appendix A.4 .

Expanding the cutoff admits more residual frequencies at full weight; decreasing λi\lambda_i releases out-of-band attenuation. The initial weight outside the passband is 0.65, not zero, and λi=0\lambda_i=0 restores full-band consistency. This is a shared practical configuration, not a taskwise optimum. Response 4 tests alternative unlocking functions.

2. ImageNet Gaussian deblurring: operator-specific explanation (Weakness 2; Question 1)

The default ImageNet Gaussian-deblurring result remains less competitive than DDRM/DDNM. We now explicitly acknowledge this limitation in Section 4.2 (“ImageNet Gaussian deblurring”) and provide the operator-specific analysis in Appendix A.6.2. A low-pass operator does not imply maximal benefit from further low-pass weighting: reducing gradient error is different from recovering weakly observed information. The following fixed-stage analysis clarifies this distinction, without claiming to identify the unique cause of the empirical gap.

Operator dependence (Appendix A.6.2, Eqs. (75)–(76)). Under the idealized Fourier-diagonal Gaussian-convolution model [1],

ab(ν)=exp⁡(−2π2b2∥ν∥22),y~=abx~gt+ϵ~.a_b(\nu)=\exp(-2\pi^2b^2\lVert\nu\rVert_2^2),\qquad \widetilde y=a_b\widetilde x_{\mathrm{gt}}+\widetilde\epsilon.

Here, bb is the blur standard deviation and ν\nu is measured in cycles per spatial unit; finite kernels and boundary conditions determine the actual experimental response. For a fixed clean-estimate difference δi=x^0,i−x0,i∗\delta_i=\hat{x}_{0,i}-x_{0,i}^{*}, temperature βi>0\beta_i>0, and residual-space operator QiQ_i with Fourier weights qi(ν)=(1−λi)+λi1{ν∈Ωi}q_i(\nu)=(1-\lambda_i)+\lambda_i\mathbf{1}_{\{\nu\in\Omega_i\}}, the measurement-gradient mismatch satisfies

ei=A⊤QiAδiβi2,∥efull∥22−∥ei∥22=1βi4∑ν(1−qi(ν)2)∣ab(ν)∣4∣δ~i(ν)∣2≥0.e_i=\frac{A^\top Q_iA\delta_i}{\beta_i^2},\qquad \lVert e_{\mathrm{full}}\rVert_2^2-\lVert e_i\rVert_2^2=\frac{1}{\beta_i^4}\sum_\nu(1-q_i(\nu)^2)|a_b(\nu)|^4|\widetilde\delta_i(\nu)|^2\geq0.

This requires simultaneous Fourier diagonalization and compares the same pair of clean estimates; the score-error form additionally requires the relation in Corollary A.8. Gaussian attenuation contributes ∣ab(ν)∣4=exp⁡(−8π2b2∥ν∥22)|a_b(\nu)|^4=\exp(-8\pi^2b^2\lVert\nu\rVert_2^2), so strongly attenuated bands may already contribute little gradient mismatch. The available reduction depends on overlap with the prediction-error spectrum, not the low-pass label alone.

Measurement–predictor trade-off (Appendix A.6.2, Proposition A.15 and Eqs. (70)–(74)). At a fixed stage and predictor, the negative log-density in Eq. (9) is quadratic. Let hih_i be the predictor's Fourier coefficient and κi=βi2/σi2\kappa_i=\beta_i^2/\sigma_i^2. Completing the square gives the target mean and covariance eigenvalue:

μi=qiab∗y~+κihiqi∣ab∣2+κi,vi=βi2qi∣ab∣2+κi.\mu_i=\frac{q_i a_b^*\widetilde y+\kappa_i h_i}{q_i|a_b|^2+\kappa_i},\qquad v_i=\frac{\beta_i^2}{q_i|a_b|^2+\kappa_i}.

Consequently,

μi−x~gt=κi(hi−x~gt)+qiab∗ϵ~qi∣ab∣2+κi.\mu_i-\widetilde x_{\mathrm{gt}}=\frac{\kappa_i(h_i-\widetilde x_{\mathrm{gt}})+q_i a_b^*\widetilde\epsilon}{q_i|a_b|^2+\kappa_i}.

Reducing qiq_i increases the predictor-error coefficient and reduces the observation-noise coefficient; their correlation prevents a general monotonic risk conclusion. When qi∣ab∣2/κiq_i|a_b|^2/\kappa_i is small, the anchor dominates precision and vi≈σi2v_i\approx\sigma_i^2. For fixed positive κi\kappa_i and bounded residual, the mean displacement qiab∗(y~−abhi)/(qi∣ab∣2+κi)q_i a_b^*(\widetilde y-a_bh_i)/(q_i|a_b|^2+\kappa_i) vanishes as ∣ab∣→0|a_b|\to0. Thus, weighting cannot restore information suppressed by the operator, and stochastic detail variation can remain.

Haar commitment controls that variation but can also withhold useful corrections. Appendix A.6.3, Proposition A.17 and Eq. (80) formalize this trade-off. For a fixed pre-fusion distribution of detail error aia_i and correction bib_i with finite second moments, its risk is

Ri(g)=E∥ai∥22+2g E⟨ai,bi⟩+g2E∥bi∥22.R_i(g)=\mathrm{E}\lVert a_i\rVert_2^2+2g\,\mathrm{E}\langle a_i,b_i\rangle+g^2\mathrm{E}\lVert b_i\rVert_2^2.

AuthorsOpenReview ↗

Official Comment Part 2

The cross term and correction energy determine the trade-off; neither stronger suppression nor one shared schedule is universally optimal. This is not a Fourier interpretation of Haar subbands or a guarantee for the full trajectory.

Evidence and cost. Table 2 in Section 4.3 shows that the full method improves all three metrics over either removal on FFHQ super-resolution, motion deblurring, and random inpainting. For FFHQ Gaussian deblurring it improves PSNR/SSIM, whereas removing exposure gives lower LPIPS. These findings are compatible with operator- and metric-dependent benefits, not a universal ranking of tasks. Different FFHQ/ImageNet predictors can also interact differently with the shared schedule.

The default uses 1000 diffusion evaluations and 40,000 inner updates, as specified by Table 1 and the N=T=200N=T=200 configuration in Table 22. More iterations do not remove target mismatch, predictor error, or discretization bias, and more accurate target sampling need not improve restoration metrics. Appendix A.6.2, Eq. (77) distinguishes sampling a target from minimizing reconstruction error. Section 4.2 and Appendix A.6.2 explicitly distinguish this mechanism analysis from a verified causal diagnosis of the ImageNet performance gap. The runtime column of Table 1 is measured on motion deblurring, as its revised caption specifies; it is not a task-specific Gaussian-deblurring timing comparison.

3. Notation and definition order (Weakness 3)

We have moved the Haar-coordinate definitions before the conditional-distribution statement in Section 3.3 (text preceding and including Eq. (11)):

z=Wx0=(zc,zd),z^i=Wx^0,i,z=Wx_0=(z_c,z_d),\qquad \hat z_i=W\hat x_{0,i},

where zcz_c is the coarse block and zdz_d collects detail coefficients. We have retained ωi\omega_i for the cutoff and renamed the commitment gate wi,dw_{i,d} as gi,dg_{i,d}. The revised notation appears in Eqs. (12)–(13), Algorithm 1, and Table 22, without changing the gate definition or algorithm. This response also uses the revised notation.

4. Normalized progress and alternative unlocking schedules (Question 3)

We have clarified the normalization and specified all three unlocking functions in Appendix A.4.1. The comparisons on both datasets are included in Table 4 and the accompanying “Unlocking schedules” paragraph in Section 4.3, and are reproduced below as Tables R1–R2. Equation (13) separates endpoints from adoption timing:

ui=ds+(de−ds)ρiγ,gi,d=ui(1−λi),ρi=N−iN−1.u_i=d_s+(d_e-d_s)\rho_i^\gamma,\qquad g_{i,d}=u_i(1-\lambda_i),\qquad \rho_i=\frac{N-i}{N-1}.

For N>1N>1, ρi\rho_i runs from zero to one; it normalizes outer-stage progress, not diffusion noise. We retain ds=0.05,de=1d_s=0.05,d_e=1. Linear unlocking sets γ=1\gamma=1 instead of the default 10; at ρi=0.5\rho_i=0.5, the respective unlocking values are 0.525 and 0.0509.

For logistic unlocking, we replace ρiγ\rho_i^\gamma with

hk,c(ρ)=s(k(ρ−c))−s(−kc)s(k(1−c))−s(−kc),s(t)=11+exp⁡(−t),k=10,c=0.5.h_{k,c}(\rho)=\frac{s(k(\rho-c))-s(-kc)}{s(k(1-c))-s(-kc)},\qquad s(t)=\frac{1}{1+\exp(-t)},\qquad k=10,\quad c=0.5.

All three schedules retain the same endpoints and exposure-linked factor. “Linear” describes unlocking, not necessarily the complete gate. Bold marks the best reported value, including ties.

Table R1. Unlocking schedules on FFHQ motion deblurring (revised manuscript: Table 4, FFHQ block).

SchedulePSNR ↑SSIM ↑LPIPS ↓
Linear (γ=1\gamma=1)37.030.9490.046
Normalized logistic37.040.9530.041
Power law (γ=10\gamma=10, submitted default)36.690.9400.054

Table R2. Unlocking schedules on ImageNet motion deblurring (revised manuscript: Table 4, ImageNet block).

SchedulePSNR ↑SSIM ↑LPIPS ↓
Linear (γ=1\gamma=1)32.600.9080.071
Normalized logistic32.640.9080.070
Power law (γ=10\gamma=10, submitted default)34.700.9350.155

On FFHQ, linear/logistic unlocking improve all metrics, with 0.34/0.35 dB PSNR gains. Logistic's 0.01 dB advantage over linear does not establish statistical superiority. On ImageNet, the default retains higher PSNR/SSIM, whereas linear/logistic reduce LPIPS from 0.155 to 0.071/0.070, with PSNR decreases of 2.10/2.06 dB.

Thus, the specific power law is not indispensable, but adoption timing matters. No schedule dominates across datasets and metrics. These comparisons change both curve shape and timing, so they do not isolate their individual effects.

5. Parameter sensitivity across datasets and operators (Question 4)

We extended the original FFHQ motion-deblurring analysis to ImageNet motion deblurring and 4× super-resolution on both datasets. All results have been incorporated into Appendix A.3.4, “Parameter Sensitivity”: Tables 9–11 report sampling budgets, Tables 15–17 report measurement noise, and Tables 19–21 report continuation/fusion parameters. Tables R3–R11 below reproduce these results.

AuthorsOpenReview ↗

Official Comment Part 3

(a) Sampling budgets

Table R3. Sampling-budget sensitivity on ImageNet motion deblurring (revised manuscript: Table 9).

(N,T)(N,T)PSNR ↑SSIM ↑LPIPS ↓Runtime (s/image)
(50, 200)33.680.9180.179~55
(100, 200)34.230.9290.162~100
(300, 200)35.020.9450.142~300
(500, 200)34.970.9420.145~430
(200, 50)33.760.9320.159~110
(200, 100)34.800.9390.152~170
(200, 300)34.940.9430.144~245
(200, 500)33.910.9120.176~330
(300, 100)34.010.9330.157~270
(400, 50)33.950.9360.153~310
(500, 25)34.310.9410.148~370
(300, 300)35.060.9440.140~345
(400, 400)34.730.9330.156~490
(500, 500)34.340.9220.166~655
(200, 200) — default34.700.9350.155~200

Table R4. Sampling-budget sensitivity on FFHQ 4× super-resolution (revised manuscript: Table 10).

(N,T)(N,T)PSNR ↑SSIM ↑LPIPS ↓Runtime (s/image)
(50, 200)31.440.8730.092~35
(100, 200)31.650.8760.091~65
(300, 200)31.850.8800.090~195
(500, 200)31.830.8800.090~280
(200, 50)31.660.8780.094~70
(200, 100)31.740.8790.092~110
(200, 300)31.770.8780.090~160
(200, 500)31.730.8780.089~215
(300, 100)31.800.8800.092~175
(400, 50)31.730.8790.089~200
(500, 25)31.600.8780.099~240
(300, 300)31.910.8990.090~225
(400, 400)31.820.8800.089~320
(500, 500)31.850.8800.089~425
(200, 200) — default31.880.8900.090~130

Table R5. Sampling-budget sensitivity on ImageNet 4× super-resolution (revised manuscript: Table 11).

(N,T)(N,T)PSNR ↑SSIM ↑LPIPS ↓Runtime (s/image)
(50, 200)27.330.7550.176~55
(100, 200)27.440.7580.173~100
(300, 200)27.390.7580.174~300
(500, 200)27.410.7580.172~430
(200, 50)27.400.7580.178~110
(200, 100)27.460.7600.174~170
(200, 300)27.470.7600.172~245
(200, 500)27.720.7660.169~330
(300, 100)27.300.7580.175~270
(400, 50)27.370.7590.176~310
(500, 25)27.210.7540.181~370
(300, 300)27.400.7580.171~345
(400, 400)27.340.7580.171~490
(500, 500)27.420.7590.180~655
(200, 200) — default27.670.7650.170~200

Budget effects are non-monotonic. On ImageNet motion deblurring, (500,500)(500,500) gives 34.34 dB, below the default's 34.70 dB, while (200,100)(200,100) improves all metrics to 34.80/0.939/0.152 at approximately 170 rather than 200 s/image. The best PSNR is 35.06 dB at (300,300)(300,300), a 0.36 dB gain at higher reported cost.

Super-resolution PSNR spans 31.44–31.91 dB on FFHQ and 27.21–27.72 dB on ImageNet; the best values exceed the defaults by only 0.03/0.05 dB. This shows task-dependent saturation over the tested grids, not statistical equivalence or universal parameter optimality.

(b) Measurement noise

These experiments vary observation conditions, not a sampler hyperparameter or noise-model misspecification.

Table R6. Measurement-noise sensitivity on ImageNet motion deblurring (revised manuscript: Table 15).

σy\sigma_yPSNR ↑SSIM ↑LPIPS ↓
0.0036.300.9600.130
0.0335.400.9480.143
0.05 — default34.700.9350.155
0.0733.200.9200.185

Table R7. Measurement-noise sensitivity on FFHQ 4× super-resolution (revised manuscript: Table 16).

σy\sigma_yPSNR ↑SSIM ↑LPIPS ↓
0.0032.700.9120.078
0.0332.250.9000.084
0.05 — default31.880.8900.090
0.0731.150.8740.106

Table R8. Measurement-noise sensitivity on ImageNet 4× super-resolution (revised manuscript: Table 17).

σy\sigma_yPSNR ↑SSIM ↑LPIPS ↓
0.0028.300.7850.155
0.0327.950.7740.163
0.05 — default27.670.7650.170
0.0727.050.7500.189

Increasing noise consistently reduces PSNR/SSIM and increases LPIPS. From σy=0\sigma_y=0 to 0.07, PSNR drops by 3.10, 1.55, and 1.25 dB, respectively. The trend transfers across these settings, although its magnitude differs.

Reviewer_kvGdOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors tackle posterior sampling in diffusion models. The idea is to emphasize the low-frequency corrections at high noise levels before resorting to a full-band correction. A second Haar step controls detail corrections for further improvements. Experiments are presented on a set of linear inverse problems.

Critical strengths

The idea of frequency-based weighting is novel and intuitive and the method is training-free so can be applied to existing pre-trained diffusion models.

Critical weaknesses
  1. Hard to follow: The paper was very hard for me to parse. Some examples include (but are not restricted to) – Phrases like “measurement-frequency exposure during refinement with correction commitment” in the abstract make the paper hard to parse, as these are non-standard trade-offs in this literature.

– Terms like “measurement exposure” are never defined in the introduction.

– Section 3.1: The theorems and corollaries are just mentioned in passing, and no intuitive explanation is given.

– The paragraph between lines 195-205 is useful for an intuitive understanding, which raises the question of what purpose the preceding theory serves.

– Similarly, Eqs. 8,9,10 could be explained more intuitively, but these do most of the heavy lifting in terms of understanding the method

  1. Dubious evaluations: There seem to be some inconsistencies in the evaluation:

– The entries for FGPS are odd. For instance, for motion deblurring, this method performs worse on FFHQ than ImageNet (PSNR: 21.87 vs 21.70) while ImageNet is the harder dataset to run on. For all other baselines, the trend is reversed, as expected. Similarly, the results in the ablations are a bit odd. Removing either the Haar or the frequency continuation in Table 2 leads to large degradation in performance. However, their combination is extremely strong, which suggests that the two are strongly coupled. Can the authors elaborate on why this is the case?

– Comparisons with state-of-the-art baselines like DiffPIR (which is cited)[1] and NDTM [2] are missing. Can the authors show comparisons on one inverse problem, let's say 4x super-resolution, for the rebuttal with these baselines?

– Does the model generalize to non-linear inverse problems, such as non-linear deblurring?

– Large-scale generalization: How does the method generalize to large-scale text-to-image models? Can the authors demonstrate the applicability to another task like Style Guidance?

References:

[1] Denoising Diffusion Models for Plug-and-Play Image Restoration, Zhu et al.

[2] Variational Control for Guidance in Diffusion Models, Pandey et al.

Questions for authors

See weaknesses

8 comments
AuthorsOpenReview ↗

Official Comment Part 1

Thank you for recognizing the novelty and intuition of our approach. All manuscript references below refer to the revised manuscript. Results and implementation details supplied only in this response are identified explicitly.

1. Terminology in the abstract (Weakness 1, opening concern)

Exposure weights measurement residuals during refinement; commitment controls how much of the resulting correction is retained. We have rewritten the Abstract and clarified both terms at first use in the Introduction. The revised abstract now explains how residuals are weighted during refinement and how the resulting corrections are retained. In plain terms:

At high diffusion noise levels, we down-weight high-frequency measurement residuals during refinement. We then fully adopt coarse corrections while gradually admitting detail corrections relative to the diffusion prediction.

2. Definition of measurement exposure (Weakness 1, bullet 1)

We now define both controls directly in the Introduction and explain their different roles in the Introduction and Figure 1:

"exposure determines how measurement frequencies contribute to refinement (a), whereas commitment determines how corrections enter the next sampling state (b)."

The revised Introduction explains why down-weighting residuals does not restrict all detail changes introduced by stochastic refinement, followed by the implementation overview. The mathematical definitions appear in Section 3.2, Eq. (8) and Section 3.3, Eq. (12).

3. Intuition accompanying the theory (Weakness 1, bullet 2)

We have revised Section 3.1 to explain the practical implication of each theoretical result alongside its citation. The opening paragraph links clean-estimate errors to unreliable measurement corrections; the subsequent discussion explains how passband restriction attenuates error contributions, why inexact gradients can disrupt descent, and why band limitation does not increase the curvature bound. Section 3.2 connects Corollary A.8 to residual weighting, while Section 3.3 connects weakly constrained detail directions to post-refinement commitment.

4. Purpose of the theoretical analysis (Weakness 1, bullet 3)

The overview explains the procedure; the theory characterizes the mechanisms motivating it. Under the stated linear-Gaussian and Tweedie-estimator assumptions, Theorem A.4 gives

Δg=σ2σy2A⊤A Δs,\Delta g=\frac{\sigma^2}{\sigma_y^2}A^\top A\,\Delta s,

relating score-estimation error to the difference between measurement-energy gradients at estimated and oracle clean images. Under simultaneous Fourier diagonalization, Corollaries A.5/A.8 characterize frequency-dependent attenuation, including the factor 1−λi1-\lambda_i outside the passband. Theorem A.6 explains how gradient errors can disrupt descent, and Corollary A.7 bounds curvature without requiring a strict reduction. Lemma A.11/Corollary A.12 describe weakly constrained detail directions in an idealized low-pass setting, motivating a separate commitment step.

These results support the design under explicit assumptions; they neither prove convergence of the complete sampler nor uniquely determine the schedule functions. The revised manuscript explains these connections in Sections 3.1–3.3, with the supporting results in Appendix A.5: Theorem A.4, Corollary A.5, Theorem A.6, Corollary A.7, Corollary A.8, and Lemma A.11/Corollary A.12. The fixed-stage scope is made explicit at the start of Appendix A.6, and Appendix A.6.3 explains why the analysis does not select a universally optimal gate or schedule.

5. Interpretation of Equations (8)–(10) (Weakness 1, bullet 4)

Let r=A(x)−yr=A(x)-y, with Fourier coefficients rL,rHr_L,r_H inside and outside the passband. Equation (8) is

Li(x;y)=∥rL∥22+(1−λi)∥rH∥22.L_i(x;y)=\lVert r_L\rVert_2^2+(1-\lambda_i)\lVert r_H\rVert_2^2.

It retains unit weight inside the passband and down-weights the complement, returning to full-band consistency as λi→0\lambda_i\to0. Equation (9) combines this objective with a predictor-centered anchor. Its negative log-density, up to an additive constant, is:

Ui(x)=Li(x;y)2βi2+∥x−x^0,i∥222σi2.U_i(x)=\frac{L_i(x;y)}{2\beta_i^2}+\frac{\lVert x-\hat{x}_{0,i}\rVert_2^2}{2\sigma_i^2}.

Equation (10) then performs Langevin refinement: the drift combines measurement correction and attraction toward the predictor, while Gaussian noise enables stochastic exploration. Haar commitment subsequently controls the resulting correction. These explanations have been incorporated directly beside Eqs. (8)–(10) in Section 3.2: passband weights after Eq. (8), the measurement–anchor interpretation after Eq. (9), and the roles of drift and injected noise after Eq. (10). The equivalent energy UiU_i is also written explicitly in Appendix A.6, Eq. (60).

AuthorsOpenReview ↗

Official Comment Part 2

6. FGPS results on FFHQ and ImageNet (Weakness 2, bullet 1 — FGPS results)

FGPS was evaluated on the same test images and with the same metric implementation as our method, not copied from its original paper. Section 4.1 specifies aligned operators, noise levels, resolution, metric implementation, and 100 validation images per dataset. Table 1 reports 21.70 dB on FFHQ and 21.87 dB on ImageNet for motion deblurring. Appendix A.4, “FGPS” specifies the FGPS blur settings. The published comparisons and counterexample below provide additional clarification in this response.

The opposite dataset ordering also occurs in published evaluations. Each pair below is ImageNet vs. FFHQ, taken within the cited paper:

FGPS — MCG: Table 1 of [1] reports 12.01 vs. 11.64 dB for Gaussian deblurring and 11.80 vs. 11.54 dB for motion deblurring.

DPS — Score-SDE/ILVR: Appendix E, Tables 6–7 of [2] report 15.97 vs. 7.12 dB for Gaussian deblurring, 7.21 vs. 6.58 dB for motion deblurring, and 18.62 vs. 13.52 dB for random inpainting.

DAPS — PSLD: Table 1 of [3] reports 25.42 vs. 24.35 dB for 4× super-resolution, 25.86 vs. 23.27 dB for Gaussian deblurring, and 31.30 vs. 30.31 dB for random inpainting. Appendix F.3 specifies SD1.5 for PSLD on ImageNet, so this last example is not a fixed-backbone comparison isolating dataset identity.

There is no necessary PSNR ordering from dataset diversity alone. A direct counterexample uses X=(C+aV,C−aV)X=(C+aV,C-aV) with independent V∼Uniform⁡[−1,1]V\sim\operatorname{Uniform}[-1,1] and

P1: C∼Uniform⁡[7/16,9/16], a=1/4;P2: C∼Uniform⁡[1/4,3/4], a=1/8.P_1:\ C\sim\operatorname{Uniform}[7/16,9/16],\ a=1/4;\qquad P_2:\ C\sim\operatorname{Uniform}[1/4,3/4],\ a=1/8.

Both distributions lie in [0,1]2[0,1]^2. If the interval for CC has width ww, the transformation has Jacobian magnitude 2a2a, giving h(X)=log⁡(4aw)h(X)=\log(4aw) and h(P2)−h(P1)=log⁡2>0h(P_2)-h(P_1)=\log2>0 in these fixed coordinates. With the same observation Y=(X1+X2)/2+ϵY=(X_1+X_2)/2+\epsilon and reconstruction f(Y)=(Y,Y)f(Y)=(Y,Y),

MSE=ϵ2+a2V2\mathrm{MSE} = \epsilon^2 + a^2V^2

MSE(P2)<MSE(P1)\mathrm{MSE}(P_{2}) \lt \mathrm{MSE}(P_{1})

under a coupling using the same VV and independent Gaussian noise realization. With peak value one, this implies higher expected per-image PSNR for the higher-entropy distribution, without inferring mean PSNR from mean MSE. The expectations are finite; with no noise the gap is 20log⁡102≈6.0220\log_{10}2\approx6.02 dB, and ff is the posterior mean under both distributions.

This counterexample does not model FFHQ/ImageNet or validate particular numerical entries. It shows why the 0.17 dB reversal alone is insufficient to establish an evaluation inconsistency. The reported entries are supported by the common evaluation protocol above.

7. Interaction between exposure and commitment (Weakness 2, bullet 1 — component ablations)

Frequency weighting changes the measurement-gradient term in Eq. (10), not the injected Langevin noise. Haar commitment instead controls the entire predictor-relative correction, including stochastic detail changes:

Δi=x0,iref−x^0,i\Delta_i = x_{0,i}^{\mathrm{ref}} - \hat{x}_{0,i}

x0,ifuse=x^0,i+W⊤diag⁡(Ic,gi,dId)WΔix_{0,i}^{\mathrm{fuse}} = \hat{x}_{0,i} + W^{\top}\operatorname{diag}(I_c, g_{i,d}I_d)W\Delta_i

Thus, weighting the objective does not replace controlling the correction, and post-refinement gating does not replace reliable measurement guidance. The predictor itself is not filtered.

The ablations isolate these interventions: “w/o. Freq” changes only Eq. (8), keeping the full-method Haar gates fixed; “w/o. Haar” retains exposure but passes the refined estimate forward without fusion. We have specified these controls explicitly in Section 4.3. The quantitative and visual results appear in Table 2 and Figure 8, respectively. The distinction between modifying the drift and controlling the entire correction is also explained in Section 3.2 and Section 3.3. We use the revised gate notation gi,dg_{i,d} consistently.

8. Comparisons with DiffPIR and NDTM (Weakness 2, bullet 2)

We use the same test subsets, image identities, degradation settings, and evaluation pipeline as our method, including the same 100 FFHQ validation images.

Metrics. Inputs in [−1,1][-1,1] are mapped by clip⁡((x+1)/2,0,1)\operatorname{clip}((x+1)/2,0,1); [0,1][0,1] outputs are not rescaled again. Metrics are evaluated per image on full RGB images without luminance conversion or border shaving. We use piq.psnr (data_range=1.0, convert_to_greyscale=False, −10log⁡10(MSE+10−8)-10\log_{10}(\mathrm{MSE}+10^{-8})); piq.ssim (data_range=1.0, 11×11 Gaussian window, standard deviation 1.5, k1=0.01,k2=0.03k_1=0.01,k_2=0.03); and piq.LPIPS(replace_pooling=True, reduction='none') with VGG16 features, learned weights, average-pooling replacement, and internal ImageNet normalization.

Our (N,T)=(200,200)(N,T)=(200,200) and (9,90)(9,90) configurations use 40,000 and 810 inner updates, respectively. Runtime is reported separately.

AuthorsOpenReview ↗

Official Comment Part 3

FFHQ:

MethodPSNR ↑SSIM ↑LPIPS ↓Runtime (s/image)
DiffPIR26.710.7690.259~4
NDTM24.420.5000.353~35
Ours (N=200,T=200N=200,T=200)31.880.8900.090~130
Ours (N=9,T=90N=9,T=90)30.610.8550.104~2

ImageNet:

MethodPSNR ↑SSIM ↑LPIPS ↓Runtime (s/image)
DiffPIR23.150.6020.371~15
NDTM22.510.5000.320~70
Ours (N=200,T=200N=200,T=200)27.670.7650.170~200
Ours (N=9,T=90N=9,T=90)26.860.7410.188~4

Both configurations improve all three quality metrics over DiffPIR and NDTM on both datasets. The reduced-budget results show that these gains are not confined to the full sampling budget. Runtime values are approximate measurements reported for the corresponding configurations.

9. Extension to blind bilinear inverse problems (Weakness 2, bullet 3)

Does the method generalize to nonlinear inverse problems, such as nonlinear deblurring?

Appendix A.3.3, “Blind deblurring with unknown kernels”, and Figure 10 report blind deblurring on FFHQ and ImageNet following the BlindDPS setting (Chung et al., 2023a). The blur kernel is unknown during reconstruction and is estimated together with the image. The ground-truth kernel is not supplied to the reconstruction algorithm.

The observation model is

y=B(x0,k)+ϵ,B(x0,k)=k∗x0.y=B(x_0,k)+\epsilon,\qquad B(x_0,k)=k*x_0.

Although convolution is linear in x0x_0 for a fixed kernel, BB is bilinear, and therefore nonlinear, in the jointly unknown image and kernel. This experiment extends beyond the known-linear-operator benchmarks by additionally requiring inference of the degradation parameters. The degradation settings are specified in Appendix A.3.3: 64×64 kernels generated with intensity 0.5 and additive Gaussian noise with standard deviation σy=0.02\sigma_y=0.02 in the implementation's normalized image coordinates. These settings are specific to this blind experiment, rather than the main known-operator benchmark. Frequency exposure acts on the measurement residual formed with the estimated kernel, while Haar commitment controls the predictor-relative image correction after refinement.

Figure 10 presents the degraded input, BlindDPS reconstruction, our reconstruction, and ground truth. These examples provide qualitative evidence for the evaluated blind bilinear problem. The unknown-kernel protocol, the roles of exposure and commitment, and the joint nonlinearity are now described in Appendix A.3.3 and the revised Figure 10 caption. The fixed-linear Fourier-wise guarantees are not asserted for the joint blind problem. Appendix A.6.4, Proposition A.18, provides the local Jacobian-based framework under its stated assumptions.

10. Pretrained text-to-image models (Weakness 2, bullet 4 — backbone generalization)

Appendix A.3.3, “Stable-Diffusion-based restoration at 512×512”, Table 6, and Figure 11 evaluate frozen SD1.5 on FFHQ at 512×512 without additional training; these results are also highlighted in Section 4.2.

For completeness, we supply the SD-specific settings here: CFG 7.5, prompt “A natural looking human face.”, 50 outer steps, five Euler PF-ODE steps per outer stage, and a polynomial noise schedule with exponent 7 from 10 to 0.1. PF-ODE integration uses exponent 7 and minimum noise 0.02. These detailed settings are provided in this response and are distinct from the main pixel-space defaults in Table 22.

Refinement updates the latent through decoder gradients of the pixel-space measurement residual. The decoded-latent objective and its gradient are now stated in Appendix A.6.4, Eqs. (85)–(86). Commitment is applied after refinement, before re-noising.

Against SILO with the same backbone and resolution, Table 6 improves all three metrics for Gaussian deblurring and super-resolution. Box inpainting improves SSIM/LPIPS slightly, with PSNR 0.11 dB lower. Figure 11 provides the corresponding visual evidence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.