acceptodds
Under review as a conference paper at ICLR 2027

LARC: A General Late-stage Adaptive Reconstruction Corrector for Diffusion Inverse Problems

Abstract

Diffusion models provide powerful generative priors for solving imaging inverse problems, enabling high-quality reconstructions from incomplete or corrupted measurements. However, further improving their reconstruction quality can be computationally expensive, particularly when refinement involves repeated inner sampling or optimization across noise levels. Motivated by the coarse-to-fine behavior of reverse diffusion sampling, we introduce the **L**ate-stage **A**daptive **R**econstruction **C**orrector (**LARC**), a general corrector that refines image details based on the structure established during early sampling. LARC can be integrated into a broad range of diffusion inverse solvers, preserving the host solver's early sampling trajectory and refining its later stages. At each selected noise level, LARC performs a two-stage correction. The first is measurement-guided initialization, which solves a regularized minimization problem to obtain an initial estimate guided by the observation. The second is preconditioned score refinement, which further refines the initial estimate through score updates with direction-dependent step sizes. Together, these stages combine measurement guidance with the diffusion prior to refine image details through a small number of local updates. Experiments on FFHQ and ImageNet across inpainting, super-resolution, and deblurring demonstrate that LARC outperforms competing baselines in reconstruction accuracy and consistently improves the tested host solvers with limited additional computation.

Primary area:
generative models
Keywords:
Diffusion Models, Inverse Problem, Adaptive Correcter

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_B58YOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Summary

This paper proposes LARC, a late-stage reconstruction corrector for diffusion-based inverse problems. It preserves the early sampling trajectory of an existing solver and replaces subsequent updates with a two-stage procedure: measurement-guided initialization, which balances data fidelity with proximity to the denoised prediction and current sampling state, followed by preconditioned score refinement at a fixed noise level. The preconditioner adjusts updates according to the measurement operator’s geometry. Experiments on FFHQ and ImageNet cover inpainting, super-resolution, and deblurring, demonstrating improvements in reconstruction accuracy across four host solvers. Component and scheduling ablations further support the framework’s practical effectiveness. However, its technical novelty and the specific limitation it addresses remain insufficiently established relative to existing PnP and score-based correction methods. Moreover, the compatibility between the residual statistics after initialization and the original noise level used for subsequent score updates requires further clarification.

Critical strengths

Strengths

  1. Consistent improvements in reconstruction accuracy. The experiments demonstrate PSNR improvements across multiple inverse problems on FFHQ and ImageNet.

  2. Flexible integration with existing solvers. LARC reuses pretrained diffusion models without additional training and is evaluated with four different solvers.

  3. Informative component and schedule ablations. The paper examines initialization, preconditioning, inner iteration counts, and correction placement. In particular, the comparison against isotropic updates with matched average scaling and the placement experiments under fixed NFE budgets provide useful evidence for the effectiveness of the proposed design choices.

Critical weaknesses

Weaknesses

  1. Limited technical novelty. The main components are closely related to existing methods. Adapting updates across directions according to the measurement operator’s spectral structure is already reflected in DDRM [1], while SNIPS [2] employs curvature-based preconditioned score updates. Score correction at a fixed noise level is related to existing predictor–corrector approaches [3], and the measurement-guided initialization resembles the data-consistency subproblems in PnP methods such as DiffPIR [4]. Although the specific updates differ and I acknowledge the reported performance improvements, the contribution appears primarily to be a combination of these established ideas applied during late-stage sampling. The paper does not sufficiently clarify the substantive technical novelty of this combination.

  2. The distinct contribution remains unclear. Despite the performance improvements, the paper does not sufficiently explain which limitations of existing solvers the proposed design choices address or what independent benefits they provide. The ablations support the effectiveness of initialization and preconditioning, but the necessity of the additional current-state constraint in Eq. (10) lacks targeted justification. Consequently, it remains unclear what generalizable methodological insight the paper provides beyond an effective combination of existing techniques.

  3. The consistency between initialization and subsequent score updates requires clarification. The initialization in Eq. (10) may alter both the magnitude and statistical structure of the reconstruction residual, which is not guaranteed to follow the Gaussian perturbation model associated with noise level tt. Nevertheless, the second stage continues to use the score network conditioned on the original noise level tt. Previous studies have shown that a mismatch between optimization-induced residuals and the input distribution of pretrained denoisers can degrade denoising performance or introduce artifacts [5, 6]. The authors should therefore explain why retaining the original noise level is appropriate and whether this potential mismatch introduces additional reconstruction error. The theoretical analysis assumes an exact score and does not address this practical concern, while targeted discussion or empirical validation is also missing.

References

[1] Kawar et al. Denoising Diffusion Restoration Models. 2022.

[2] Kawar et al. SNIPS: Solving Noisy Inverse Problems Stochastically. 2021.

[3] Song et al. Score-Based Generative Modeling through Stochastic Differential Equations. 2021.

[4] Zhu et al. Denoising Diffusion Models for Plug-and-Play Image Restoration. 2023.

[5] Shrestha and Fu. Taming Score-Based Denoisers in ADMM: A Convergent Plug-and-Play Framework. 2026.

[6] Du et al. Plug-and-Play Diffusion Meets ADMM: Dual-Variable Coupling for Robust Medical Image Reconstruction. 2026.

Questions for authors

Questions

  1. The initialization in Eq. (10) may alter the magnitude and statistical structure of the reconstruction residual. Could the authors explain why it remains appropriate to use the score network conditioned on the original noise level tt? Could this potential distribution mismatch compromise the reliability of the updates or introduce additional reconstruction error?

  2. What specific limitation of existing diffusion-based inverse solvers does LARC aim to address? Compared with existing data-consistency corrections, preconditioned updates, and score refinement methods, what is LARC’s central distinguishing contribution? Could the authors clarify how its specific design choices address this limitation?

7 comments
AuthorsOpenReview ↗

Response to Reviewer B58Y (1/3): summary; what is new and relation to prior solvers (W1)

We thank the reviewer for the expert reading and for [5, 6], which the revision cites and discusses in the related work. New results are on FFHQ (100 test images) for inpainting (r = 0.3) / SR 4× / motion deblurring.

Summary.

  • W1. LARC is a corrector designed for the late stage of an existing solver. Its data step contains the host's current state utu_t, and removing this term costs 2.52–9.77 dB on SR 4× and motion deblurring. In the same position, an ADMM-type and a SNIPS-style corrector fall below the host's own continuation in 3 of 9 and 6 of 12 settings; LARC improves all 12 (Table R6).
  • W2 and W3. Part 2 gives the limitation, the design rule and the measured role of the current-state term (Table R7). Part 3 treats the noise level of the inner score.

W1. What is new, and relation to DDRM, SNIPS, PC sampling and DiffPIR. The components of LARC are known. What is new is a correction step designed for the late stage of an existing solver.

1. The design. LARC takes over from a host at a late activation step and replaces only the remaining updates. Each corrected step has two parts, and one quadratic energy FtF_t (Eq. 10) defines both.

  • (a) A data step that contains the host's current state. The start is the minimizer of FtF_t, which is anchored to the denoised estimate u^0\hat u_0 and to the host's current state utu_t. Anchoring to utu_t keeps the correction on the host's trajectory. Removing the term with all else fixed costs 2.52–9.77 dB on SR 4× and motion deblurring (Table R7, part 2).
  • (b) K prior-score steps at the same noise level, scaled by GtG_t. By Proposition 1, Gt=(τt∇2Ft)−1G_t=(\tau_t\nabla^2F_t)^{-1}, so each step moves a direction only by the share the measurement leaves undetermined: zero where a noiseless measurement fixes the content, largest where it gives no information.

2. Other correctors in the same position. On each of the four hosts, after the same prefix and at LARC's total NFE, the remaining updates are replaced by FAST-DIPS, a SNIPS-style corrector (annealed Langevin with spectral preconditioning, on a geometric noise schedule down to 0.01) or LARC. Each competitor starts from its authors' recommended setting and is tuned per host and task over a range that contains it, on 10 development images (FFHQ 60000–60009, disjoint from the test set): for FAST-DIPS, ADMM penalty ∈ {175, 200, 225} and tolerance ∈ {0.04, 0.05, 0.06} (recommended: 200 and 0.05); for the SNIPS-style corrector, a step-size factor c ∈ {0.01, 0.018, 0.033, 0.06, 0.1, 0.2, 0.5, 1} and 1, 3, 5 or 10 updates per noise level, at most 5 on the DiffPIR host (official setting: c = 0.033, 5 updates). LARC keeps its Table 1 parameters. The code and all settings will be released.

Table R6. PSNR change (dB, computed before rounding) relative to the host continuing from the same state at the same NFE: FAST-DIPS / SNIPS-style / LARC. Brackets: that continuation (Table R1, reply to Reviewer 1jHd: A for DiffPIR; A′ for DPS and DAPS; the stronger of A and A′ for FAST-DIPS). "–": on the FAST-DIPS host, a FAST-DIPS late stage is the host's own continuation, i.e., the reference. Bold: best.

Host prefixInpaintingSR 4×Motion
DiffPIR [34.08 / 29.26 / 29.20]−0.34 / −1.50 / +0.19+0.65 / +0.60 / +0.71−0.61 / +0.05 / +1.15
DPS [33.16 / 28.92 / 28.09]−0.03 / −0.49 / +0.82+1.08 / +0.51 / +1.41+0.35 / −0.70 / +2.63
DAPS [28.77 / 27.86 / 27.76]+4.05 / +3.50 / +4.01+2.17 / −5.15 / +2.54+0.58 / +1.43 / +2.89
FAST-DIPS [34.04 / 30.50 / 28.82]– / −1.10 / +0.70– / −0.34 / +0.25– / +0.58 / +1.95

Only LARC improves every host on every task. Among the correctors it has the best SSIM in all 12 settings and the best PSNR in 11; on DAPS inpainting FAST-DIPS is 0.04 dB higher. The FAST-DIPS corrector falls below the host's own continuation in 3 of 9 settings, and the SNIPS-style corrector in 6 of 12.

3. Relation to the solvers named by the reviewer.

  • DDRM [1] uses the spectrum of AA to combine measurement and prediction per direction, with one network call per level. LARC's start contains a blend of this type; LARC then uses the spectrum to decide where the prior may still act.
  • SNIPS [2] runs annealed Langevin sampling of the posterior from pure noise through all levels. LARC takes K prior-score steps at one late level, and the data enter only through FtF_t.
  • PC sampling [3] uses a Langevin corrector that keeps the sample distributed as ptp_t. LARC's steps start from a measurement-guided point and return a clean estimate (Eq. 16).
  • DiffPIR [4] takes one data step per level, anchored to u^0\hat u_0 only. LARC uses its transition (17)–(18); on the DiffPIR host at the same 100 NFE, LARC gains +0.19 / +0.71 / +1.15 dB.

None of [1–4] puts the host's current state in the data step or applies K prior-score evaluations at one late level through that data step.

AuthorsOpenReview ↗

Response to Reviewer B58Y (2/3): limitation, design rule and the current-state term (W2/Q2)

Notation. For an eigenvector qiq_i of A⊤AA^\top A with eigenvalue νi\nu_i: wi=τtνi/σ2w_i=\tau_t\nu_i/\sigma^2, rt=τt/ℓtr_t=\tau_t/\ell_t, πi=wi/(1+rt+wi)\pi_i=w_i/(1+r_t+w_i) (the share of direction ii fixed by the measurement), bi=qi⊤A⊤y/νib_i=q_i^\top A^\top y/\nu_i (νi>0\nu_i>0) and ht=(u^0+rtut)/(1+rt)h_t=(\hat u_0+r_tu_t)/(1+r_t).

W2/Q2. Limitation addressed, design rule and the current-state term.

  • Limitation. Late in sampling the measurement has little left to give: once data consistency holds, the remaining error lies in directions the measurement does not determine (all of it, for noiseless measurements), and only the prior can reduce it. Existing solvers spend one network evaluation per level on this (DDRM, ΠGDM, DiffPIR, DPS) or run costly inner loops at every level (DAPS, SITCOM: 10–25× DiffPIR's runtime in Table 1). Giving the host the same or a larger late budget on its own schedule is worse than LARC in all 12 settings (Table R1, reply to Reviewer 1jHd).
  • Design rule (the transferable insight). At each late noise level, spend several network evaluations on the prior, and let one energy around the host state decide both where to start and how far each direction may move. Per eigendirection, the minimizer of Eq. (10) is πibi+(1−πi)qi⊤ht\pi_i b_i+(1-\pi_i)q_i^\top h_t and GtG_t has eigenvalue (1−πi)/(1+rt)(1-\pi_i)/(1+r_t): each direction is taken from the data by the share πi\pi_i and refined by the prior in proportion to the remaining share 1−πi1-\pi_i. The rule uses only the host's state and noise level, so the same rule applies to all four hosts.
  • Independent benefits (Table 3; PSNR gain over using neither, inpainting / SR / motion). Start alone: +1.79 / +0.08 / +2.58 dB. Metric alone: +1.32 / +0.22 / −2.58 dB. Both: +4.24 / +0.40 / +2.96 dB. Start and metric work together: without the Eq. (10) start, GtG_t hurts motion deblurring (24.81 vs. 27.39 dB with isotropic steps); with it, GtG_t is best (30.35 vs. 29.97 dB).

The current-state term of Eq. (10). u^0\hat u_0 is a posterior mean, while utu_t is the state the host has committed to. The term has two effects. (i) It moves the start in weakly observed directions from u^0\hat u_0 toward utu_t by the share rt/(1+rt)r_t/(1+r_t), which keeps the correction local to the host trajectory and gives the first score evaluation a noisier input (see W3, part 3). (ii) It rescales the trust in the data and the step: LARC equals the variant without the term whose anchor is hth_t and whose τt\tau_t and ηt\eta_t are divided by 1+rt1+r_t. Table R7 measures both. On the noiseless tasks, rows 2 and 3 coincide and give effect (i) alone; on motion deblurring they also change the trust in the noisy data.

Table R7. ΔPSNR (dB) relative to LARC; inpainting / SR 4× / motion. Row 1 is relative to the Table 1 run; rows 2 and 3 are paired with a LARC run from their own batch (same measurements, seeds and host prefix). Row 2: the term is removed and the step ηt\eta_t is divided by 1+rt1+r_t (τt\tau_t unchanged). Row 3: only the start is computed without the term; the refinement steps are unchanged.

VariantLARC(DiffPIR)LARC(DAPS)
Term removed, all else fixed−0.06 / −2.52 / −4.90−0.42 / −2.69 / −9.77
Term removed, step compensated−0.03 / −0.03 / −1.59−0.08 / −2.79 / −1.90
Term removed from the start only−0.03 / −0.03 / −0.36−0.08 / −2.79 / +0.12
  • Step scale. Compensating the step recovers the SR loss with DiffPIR (−2.52 → −0.03 dB) and most of the motion losses (−4.90 → −1.59 dB, −9.77 → −1.90 dB).
  • Anchor, effect (i). On the noiseless tasks it accounts for the whole 2.79 dB on SR with DAPS and for at most 0.08 dB in the other three settings.
  • Motion deblurring. With the step compensated, removing the term still costs 1.59 dB (DiffPIR) and 1.90 dB (DAPS), where the term also sets the trust in the noisy data; removing it from the start alone changes PSNR by −0.36 dB and +0.12 dB.
AuthorsOpenReview ↗

Response to Reviewer B58Y (3/3): noise level of the inner score evaluations (W3/Q1)

W3/Q1. Noise level of the inner score evaluations.

  • Why level t. Stage 2 is descent on Φt=−log⁡pt\Phi_t=-\log p_t, the prior smoothed at the host's level, so tt plays the role of a regularization strength, as the denoiser level does in PnP. For the exact MMSE denoiser, Tweedie's identity Dt(u)=u+λt∇log⁡pt(u)D_t(u)=u+\lambda_t\nabla\log p_t(u) holds at every uu, not only at level-tt samples, and DtD_t is the proximal map of an implicit penalty whose Moreau envelope is λtΦt\lambda_t\Phi_t up to a constant (Gribonval, 2011; Gribonval & Nikolova, 2020). Level tt is also the level of u^0\hat u_0 in Eq. (10) and of the transition (17)–(18) that follows.
  • How far the inner inputs are from level-t samples. In directions the measurement fixes, Stage 1 replaces the host's perturbation by data; there the weight (1−πi)/(1+rt)(1-\pi_i)/(1+r_t) of GtG_t vanishes in the noiseless case, so those score components are not applied. In unobserved directions the first input keeps the share rt/(1+rt)r_t/(1+r_t) of the host's level-tt noise (1/3, 0.55 and 2/3 for inpainting, SR and motion with DiffPIR; 1−πi1-\pi_i times this share in partly observed directions), and the following steps reduce it. The current-state term is what keeps this share: without it, the first input carries none of the host's noise in these directions (effect (i) in part 2).
  • The mismatch studied in [5, 6]. [5, 6] study this mismatch for PnP iterates fed to a diffusion denoiser. It is most severe with ADMM, whose dual variable accumulates structured residuals: in [5] it "further distorts" the noise geometry, and [6] finds the shift "milder" without such a memory variable (its spectral homogenization adds +0.15 dB to PnP-HQS). [5] corrects the mismatch with the AC-DC denoiser, [6] with spectral homogenization. LARC has no dual variable and takes at most 8 steps per level. Eqs. (17)–(18) then move the corrected estimate to the next level with DiffPIR's transition: the new perturbation combines fresh noise (a fraction ζ of its variance; all of it when ζ = 1, e.g. LARC(DiffPIR) on inpainting) with ϵ^t\hat\epsilon_t, the rescaled difference between the current state and the corrected estimate. The next level restarts from the minimizer of its own FtF_t.
  • Net effect. From the same state and with no more NFE than the host's own continuation, LARC ends above it in all 12 settings (Table R1), by +0.19 / +0.71 / +1.15 dB for DiffPIR: whatever mismatch remains does not outweigh the correction.

The revision adds the comparison with [1–6] to §2.2, the per-direction reading of Eq. (11) and the role of the current-state term to §3.1, the noise-level discussion to §3.2, and Tables R6–R7 to the appendix.

Reviewer_T775OpenReview ↗
Rating
2
Confidence
3
Informativeness
1(−1)
Presentation
3
Soundness
2
Read review
Paper summary

The present paper proposes LARC, a late-stage corrector for diffusion solvers of linear inverse problems. It aims to improve the final reconstruction using the structure already established by an existing solver. LARC retains the early sampling trajectory and replaces the later updates with a two-stage correction. First, it minimizes a quadratic objective combining measurement fidelity with proximity to the denoised prediction and current sampling state. The estimate is then refined through repeated score updates at a fixed noise level, using an operator-dependent preconditioner to limit changes in well-observed directions. The refined estimate is blended with the original prediction before continuing sampling. The method is evaluated with DPS, DiffPIR, DAPS, and FAST-DIPS on FFHQ and ImageNet for inpainting, super-resolution, and deblurring.

Critical strengths
  • The experiments show PSNR improvements for all four host solvers across the seven tasks and both datasets. This breadth is relevant because the method is intended to improve different existing solvers. The results support the usefulness of LARC within the evaluated linear inverse problems, although the gains are more consistent for PSNR than for LPIPS.

  • The controlled ablations support the contributions of preconditioning and late correction. With initialization retained, the proposed preconditioner improves FFHQ inpainting PSNR from 31.82 to 34.27 dB over isotropic refinement with matched average scaling. For FAST-DIPS, late correction also performs better than early or middle correction when the correction window, inner iteration count, and total number of network evaluations are kept fixed. These comparisons provide useful evidence for the contribution of the main design choices to the reported improvements.

Critical weaknesses
  • The main comparison is difficult to fully assess because the baseline configurations and tuning procedures are not described in enough detail. The authors state that official implementations are used and provide detailed LARC configurations, but the complete standalone baseline schedules and parameter settings remain unclear. The activation point is tuned on ten development images, while the selection procedure for other task-specific parameters and the tuning effort for each baseline are not clearly explained. The comparisons using shared early trajectories and the controlled schedule ablations provide useful evidence, but their relationship to the configurations used in the main comparison is not fully clear. These details matter because LARC changes the later sampling procedure and the allocation of network evaluations. Without them, it is difficult to assess the reported advantage under comparable sampling budgets and tuning effort.

  • The claim of limited additional computation needs qualification because the runtime cost varies across hosts and tasks. For FFHQ Gaussian deblurring, DiffPIR runtime increases from 1.447 to 5.282 seconds per image with LARC, while PSNR increases from 27.27 to 28.07 dB. This shows an accuracy gain accompanied by a substantial runtime increase for a fast solver. Other reported settings have similar or shorter runtimes with LARC, so the computational benefit depends on the host and task. The paper should explain this variation more clearly and relate the reconstruction improvements to the corresponding computational cost. Since efficiency is part of the claimed contribution, this variation limits how broadly the practical advantage can be stated.

Questions for authors
  • Provide the complete sampling schedules, numbers of network evaluations, and parameter settings for the standalone baselines in the main comparison. Explain how these settings correspond to those used in the comparisons with shared early trajectories, including any differences in the total sampling budget. These details would help assess whether the main reported gains are supported under comparable computation.

  • Explain how the task-specific LARC parameters other than the activation point were selected. In particular, specify how the number of inner updates, refinement step size, and blending weight were chosen, and what selection criterion was used. Were the same ten development images used for these choices? Were these images separate from the evaluation sets, and did the standalone baselines receive comparable tuning effort?

  • Explain the runtime increase for DiffPIR on FFHQ Gaussian deblurring and how this setting should be interpreted in relation to the claim of limited additional computation. What do the existing budget comparisons establish about the reconstruction improvement relative to the additional cost for fast host solvers?

5 comments
AuthorsOpenReview ↗

Response to Reviewer T775 (1/2): summary; settings and shared-trajectory comparisons (Q1)

We thank the reviewer for the careful reading. New results are on FFHQ (100 test images) for inpainting (r = 0.3) / SR 4× / motion deblurring unless noted.

Summary.

  • Budgets (Q1). Table R3 lists the NFE of both branches of the shared-trajectory comparisons at the Table 1 configurations. The host never receives fewer NFE than LARC.
  • Tuning (Q2). LARC's activation step, K, γ and ρ were selected on 10 development images disjoint from the test set. LARC improves on the strongest version of every host (its Table 1 setting, the host re-tuned at LARC's NFE, or its same-state continuation) in all 12 settings, by +0.19 to +2.07 dB (Table R4).
  • Runtime (Q3). The 5.282 s entry was likely caused by differences in the hardware environment and interference from other jobs on the GPU. Re-measured on an idle GPU, LARC(DiffPIR) takes 1.440 s against 1.430 s for DiffPIR at the same 100 NFE (Table R5).

Q1. Complete settings and correspondence with the shared-trajectory comparisons.

Standalone baselines (Table 1). Each method runs with its official code and the authors' settings for the task; where a paper lists no setting for a degradation level, we use its setting for the same task type. Sources: DPS (App. D.1), DiffPIR (App. B.1, Table 3; its deblurring setting for motion), DAPS (implementation details; we set τ = 0.03 for motion), FAST-DIPS (recommended settings), SITCOM (App. K), DDiff (App. H) and ΠGDM (App. B.1, §3.3). NFE (inpainting / SR / motion): 1000 for DPS and for DAPS (200 annealing levels × 5 ODE steps), 100 for DiffPIR, 75 / 75 / 50 for FAST-DIPS.

Host settings inside LARC. The host prefix uses the host block of ffhq.json in the supplementary code (the larc block of that file matches Appendix Tables 4–7). For inpainting / SR / motion: DPS guidance scale 4 / 1 / 4; DAPS τ = 0.01 / 0.01 / 0.03; DiffPIR λ = 1 / 7 / 7; FAST-DIPS ADMM penalty 50 / 1600 / 200 with 3 ADMM iterations, 2 inner steps and tolerance 0.05. The revision adds one table with the full schedule, NFE and hyperparameters of every baseline and of every host prefix.

Shared-trajectory comparisons. Both branches start from the same host state at LARC's activation step. The host branch keeps the late part of its own schedule (DPS: every timestep from t = 217 to 0; DAPS: its remaining annealing levels; DiffPIR and FAST-DIPS: their schedules re-spaced to LARC's budget, with the host settings above), so it never receives fewer NFE than LARC. At the exact Table 1 configurations:

Table R3. NFE (inpainting / SR / motion).

HostShared prefixAfter branching: host vs. LARCTotal: host vs. LARC
DPS350 / 200 / 350218 vs. 100 / 50 / 30568 / 418 / 568 vs. 450 / 250 / 380
DAPS150 / 150 / 200850 / 850 / 800 vs. 150 / 60 / 901000 vs. 300 / 210 / 290
DiffPIR55 / 55 / 6445 / 45 / 36 (both)100 (both)
FAST-DIPS175 / 175 / 10045 / 45 / 30 (both)220 / 220 / 130 (both)

From these shared states, LARC beats the host's own continuation in all 12 settings (column A of Table R1, reply to Reviewer 1jHd). The margins over the strongest host version are in Table R4. DiffPIR's same-state branch differs from standalone DiffPIR in its schedule, which places 45 / 45 / 36 of the 100 evaluations on the last interval, and in λ for inpainting; this is why it reaches 34.08 / 29.26 / 29.20 dB against 33.13 / 29.16 / 29.08 dB for standalone DiffPIR.

Figure 4 shows the same kind of comparison. Its FAST-DIPS panels use the Table 1 configuration. Its DPS, DAPS and DiffPIR panels use separately chosen activation steps and budgets: the LARC branches end at 540 / 325 / 440, 450 / 910 / 840 and 95 / 100 / 100 NFE (as printed in the figure), against 450 / 250 / 380, 300 / 210 / 290 and 100 in Table 1. The host branches end at the budgets of Table R3. The revision lists the Figure 4 configurations and adds the version at the Table 1 configurations (column A of Table R1).

AuthorsOpenReview ↗

Response to Reviewer T775 (2/2): parameter selection (Q2); runtime and cost (Q3)

Q2. Parameter selection, development images and tuning effort.

  • LARC. Besides the activation step, K, γ and ρ were selected per task by mean PSNR on 10 development images per dataset (FFHQ: 60000–60009), disjoint from the test images, with a coarse-to-fine grid search over K ∈ {1, 2, 3, 4, 5, 6, 8, 10}, γ ∈ {0.1, 0.2, …, 8.0} and ρ ∈ {0.1, 0.2, …, 1.0}. The selected configuration was then frozen and evaluated on the 100 test images (the subsets released with DiffPIR).
  • Baselines. Table 1 uses the authors' recommended settings. In addition, on the same development images and with the same criterion, we re-tuned at LARC's NFE every host whose budget we changed: DPS (guidance scale, 6 values), DAPS (τ, 4 values) and FAST-DIPS (12 ADMM settings), and again for the late stage alone under A′ (6, 4 and 9 settings). DiffPIR keeps its paper's parameters at its paper's 100 NFE; its strongest version in our runs is the same-state branch, which moves evaluations to the late levels and uses the host setting of LARC(DiffPIR) (λ = 1 instead of 7 for inpainting).

Table R4. PSNR (dB) of the strongest host version in each setting → LARC (gain). T1: Table 1. A: same state, at least LARC's NFE. A′: same state, LARC's NFE, late-stage parameters tuned. B: host alone at LARC's NFE, tuned (details in our reply to Reviewer 1jHd). Every strongest version uses at least LARC's NFE.

HostInpaintingSR 4×Motion
DPS33.16 (A′, B) → 33.98 (+0.82)29.22 (B) → 30.33 (+1.11)28.66 (A) → 30.73 (+2.07)
DAPS31.11 (T1, A) → 32.78 (+1.67)29.28 (T1) → 30.40 (+1.12)28.77 (A) → 30.66 (+1.89)
DiffPIR34.08 (A) → 34.27 (+0.19)29.26 (A) → 29.97 (+0.71)29.20 (A) → 30.35 (+1.15)
FAST-DIPS34.04 (A′) → 34.73 (+0.70)30.50 (A) → 30.75 (+0.25)28.82 (A′) → 30.77 (+1.95)

LARC is better in all 12 settings. For scale, the 100-image mean PSNR varies by at most 0.02 dB across three seeds (Table R5). Table 1 compares LARC with each host at its recommended setting; Table R4 gives the gain that remains over the strongest host version. The revision adds Table R4 to §4.2 and states the gains in the abstract, §4.2 and the conclusion relative to it.

Q3. Runtime and cost.

  • The 5.282 s entry was likely caused by differences in the hardware environment and interference from other jobs running on the GPU. Both methods use exactly 100 NFE, and LARC adds only closed-form FFT solves. Re-measured on one GPU without concurrent jobs (batch size 1, CUDA-synchronized, sampling time only):

Table R5. FFHQ Gaussian deblurring, 100 NFE; mean over three seeds, with the range over seeds in brackets.

PSNR (dB)SSIMTime (s/image)
DiffPIR27.29 [27.28–27.29]0.7671.430 [1.428–1.432]
LARC (DiffPIR)28.08 [28.07–28.09]0.7951.440 [1.436–1.442]

The other 13 LARC(DiffPIR)/DiffPIR runtime ratios in Table 1 lie between 0.97 and 1.02. The revised Table 1 corrects this entry and states the timing protocol.

  • Cost across hosts. LARC(DPS) and LARC(DAPS) take 0.21–0.57× and 0.05–0.39× the runtime of DPS and DAPS at their default budgets. The two fast hosts:

    • DiffPIR: the gain costs no extra NFE and less than 1% extra time (Table R5).
    • FAST-DIPS: it has the smallest budget (75 / 75 / 50 NFE), and LARC(FAST-DIPS) takes 1.04–2.71× its runtime in 12 of the 14 Table 1 settings (0.78–0.80× on Gaussian deblurring). At equal budget LARC is still better: at LARC's NFE, FAST-DIPS reaches 34.04 / 30.50 / 28.82 dB from the same state (the stronger of A and A′; 32.22 / 29.95 / 27.87 dB alone), against 34.73 / 30.75 / 30.77 dB for LARC, a gain of +0.70 / +0.25 / +1.95 dB. Of the Table 1 gain (3.42 / 0.59 / 3.24 dB), 2.73 / 0.34 / 1.29 dB is what FAST-DIPS itself gains from the larger budget.

    The revision states the computation claim per host.

Thank you for the clarifications and additional experiments. The response addresses my concerns about the evaluation protocol and DiffPIR timing. After considering the discussion, I share Reviewer B58Y's concern about limited algorithmic novelty. The authors explain the role of the current sampling state and the common objective for initialization and preconditioning. However, I still view the contribution as an incremental combination of existing correction and score refinement ideas applied during late-stage sampling. The reported improvements support its practical value, but the algorithmic advance is not sufficiently convincing to me. I will maintain my original recommendation.

Reviewer_1jHdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LARC, a training-free corrector for diffusion-based inverse problems. The method combines measurement-guided initialization with score refinement to improve reconstruction quality. Experiments on FFHQ and ImageNet across inpainting, super-resolution, and deblurring tasks show improvements over several diffusion inverse solvers, including DPS, DiffPIR, DAPS, and FAST-DIPS.

Critical strengths

LARC can be integrated into multiple existing diffusion inverse solvers without additional training, making it practically useful. The method achieves consistent improvements across multiple image restoration tasks, supported by extensive comparisons and ablation studies.

Critical weaknesses
  1. The approach combines measurement-consistency optimization with preconditioned score updates, which are similar to existing plug-and-play and proximal optimization methods. The paper does not sufficiently point out what distinguishes LARC from these established techniques.
  2. LARC introduces additional score evaluations and uses task-dependent sampling configurations. It remains unclear whether the improvements come from the proposed correction mechanism or from reallocating computational resources toward later sampling stages.
Questions for authors
  1. What is the novelty of LARC compared with existing PnP and proximal optimization approaches?
  2. Please provide comparisons of the original solvers with optimized sampling schedules under identical NFE and runtime budgets.
  3. How sensitive is LARC to the manually selected activation timestep and the number of inner refinement iterations?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 1jHd (1/2): summary; novelty relative to PnP (W1/Q1)

We thank the reviewer for the constructive comments. New results are on FFHQ (100 test images) for inpainting (r = 0.3) / SR 4× / motion deblurring unless noted.

Summary.

  • Novelty (W1/Q1). LARC is a plug-in corrector: it takes over from an existing solver, including a PnP solver, at an activation step and replaces only the remaining updates. Its update differs from a PnP iteration in two ways that follow from this role.
  • Correction, not reallocation (W2/Q2). Given the same NFE, with its schedule re-spaced and its parameters re-tuned, or continued from the same state where LARC takes over, every host stays below LARC in all 12 settings. The margin over the strongest host version is +0.19 to +2.07 dB (Tables R1 and R4).
  • Sensitivity (Q3). Every tested activation step improves on standalone DiffPIR. On the noiseless tasks K = 8–10 is best and every K ≥ 2 matches or exceeds it (Table R2).

W1/Q1. Novelty relative to PnP and proximal methods. Inside each corrected step LARC uses a PnP-type splitting (Appendix A.1.4). It differs from PnP and proximal methods in its role and in its update.

1. Role: a plug-in corrector for an existing solver, not a standalone solver. PnP and proximal methods, including DiffPIR and recent ADMM-based score PnP (Shrestha & Fu, 2026; Du et al., 2026), define the whole reconstruction. LARC keeps the host's trajectory up to an activation step selected on development images and replaces only the remaining updates. It needs only the host's current state and noise level, so the same corrector runs on four hosts with different mechanisms: gradient guidance (DPS), PnP (DiffPIR), Langevin annealing (DAPS) and ADMM (FAST-DIPS). The evidence:

  • It improves every host from the host's own state. From the same state and with no more NFE than the host's own continuation, LARC is better in all 12 settings (Table R1). This includes DiffPIR, itself a PnP method: 34.08 / 29.26 / 29.20 → 34.27 / 29.97 / 30.35 dB at the same 100 NFE.
  • Other correctors placed at the same position do not improve every host. After the same prefix and at the same NFE, an ADMM-type corrector (FAST-DIPS, on the three other hosts) and a SNIPS-style corrector (on all four) fall below the host's own continuation in 3 of 9 and 6 of 12 settings. LARC improves all 12 and has the best PSNR in 11 of them (Table R6, reply to Reviewer B58Y).
  • The position matters. The same correction placed early or mid-trajectory loses 1.47 / 0.09 / 3.71 dB, and LARC without a host loses 2.58 / 0.81 / 0.40 dB (FAST-DIPS host, Table 8).

2. Update: two differences from a PnP iteration.

  • (a) The data step corrects the host's state instead of starting a new reconstruction. Eq. (10) keeps the start close to both the host's denoised estimate u^0\hat u_0 and its current state utu_t, the state the host would continue from. This fits a late-stage corrector: anchoring to utu_t keeps the correction on the host's trajectory. DiffPIR's data step, like other PnP data steps, is anchored to the denoised estimate only. Table R7 (reply to Reviewer B58Y) measures what the term does: it rescales the refinement step and the trust in the data, and it moves the start toward utu_t. Computing the start without it lowers PSNR by 2.79 dB on SR 4× with DAPS and by 0.36 dB on motion deblurring with DiffPIR, and changes it by at most 0.12 dB in the other four settings.
  • (b) The prior step changes only what the measurement leaves open. A PnP prior step re-denoises the whole image. LARC takes a few score steps at the host's noise level and scales each one, direction by direction (along the eigenvectors of A⊤AA^\top A), by the share of that direction the measurement does not determine: zero where a noiseless measurement fixes the content, largest where it gives no information. With the same start and the same average step size, this scaling gives 34.27 dB against 31.82 dB for isotropic steps on inpainting, and +0.32 / +0.38 dB on SR / motion (Table 3).

We will add this discussion to §2.2.

AuthorsOpenReview ↗

Response to Reviewer 1jHd (2/2): equal-NFE comparisons (W2/Q2); sensitivity (Q3)

W2/Q2. Correction or reallocation of computation; identical NFE and runtime. We compare LARC with three versions of each host:

  • (A) same state, at least LARC's NFE: the host continues its own schedule from the state where LARC takes over (1.3–1.7× LARC's NFE for DPS, 3.3–4.8× for DAPS, equal for DiffPIR and FAST-DIPS). This is the comparison of Fig. 4, run here at the Table 1 configuration;
  • (A′) same state, exactly LARC's NFE (DPS, DAPS, FAST-DIPS): the host's remaining schedule is fitted to LARC's budget and its late-stage parameters are tuned on 10 development images (FFHQ 60000–60009, disjoint from the test set);
  • (B) host alone, exactly LARC's NFE: the full schedule is re-spaced to LARC's budget and the parameters are tuned on the same images (DPS: guidance scale; DAPS: τ; FAST-DIPS: ADMM penalty and inner steps). DiffPIR runs at its paper's 100 NFE with its paper's parameters.

A and A′ test reallocation directly: from the same state, the host receives the same or a larger late-stage budget and still falls short. For DiffPIR, A is itself a re-scheduled DiffPIR (more evaluations on the late levels, λ = 1 for inpainting) and is 0.95 / 0.10 / 0.12 dB above DiffPIR at its paper setting (B).

Table R1. PSNR (dB) of the host under A / A′ / B → LARC ("–": DiffPIR's A already runs at exactly LARC's NFE, with the host setting used inside LARC(DiffPIR)). LARC's NFE, also the host's under A′ and B: DPS 450 / 250 / 380, DAPS 300 / 210 / 290, DiffPIR 100, FAST-DIPS 220 / 220 / 130.

HostInpaintingSR 4×Motion
DPS32.89 / 33.16 / 33.16 → 33.9827.13 / 28.92 / 29.22 → 30.3328.66 / 28.09 / 27.36 → 30.73
DiffPIR34.08 / – / 33.13 → 34.2729.26 / – / 29.16 → 29.9729.20 / – / 29.08 → 30.35
DAPS31.11 / 28.77 / 29.68 → 32.7829.11 / 27.86 / 28.57 → 30.4028.77 / 27.76 / 28.39 → 30.66
FAST-DIPS33.18 / 34.04 / 32.22 → 34.7330.50 / 29.89 / 29.95 → 30.7528.01 / 28.82 / 27.87 → 30.77
  • LARC is best in all 12 settings under every protocol. Over the strongest host version, Table 1 settings included, the margin is +0.19 to +2.07 dB (Table R4, reply to Reviewer T775). Under B it also has the better SSIM and NRMSE in all 12 settings.
  • DPS and DAPS at their full 1000 NFE (2.2–4.8× LARC's budget) also stay below LARC (Table 1).
  • Runtime. At equal NFE, LARC adds only closed-form linear solves: with DiffPIR, 1.440 vs. 1.430 s per image at 100 NFE (FFHQ Gaussian deblurring, re-measured; reply to Reviewer T775). Against the default hosts of Table 1, LARC takes 0.21–0.57× the runtime of DPS and 0.05–0.39× that of DAPS. FAST-DIPS has the smallest default budget (75 / 75 / 50 NFE), so LARC(FAST-DIPS) takes 1.04–2.71× its runtime in 12 of 14 settings; given LARC's NFE, FAST-DIPS stays below LARC in every version (Table R1).
  • The revision adds Table R1 to §4.2 and reports each LARC variant against its strongest host version.

Q3. Sensitivity to the activation step and K. LARC(DiffPIR); DiffPIR alone gives 33.13 / 29.16 / 29.08 dB; bold marks the Table 1 setting.

Table R2. PSNR (dB). Activation step j: LARC is active from step j to the last step, so the budget changes with j (NFE 132 / 116 / 100 / 84 for j = 52 / 54 / 56 / 58; 108 / 104 / 100 / 96 for j = 61 / 63 / 65 / 67). K: the sweep of Appendix Fig. 6, with the outer schedule fixed (NFE = 60 + 5K; motion: 76 + 12K).

Activation step jK
Inpainting52: 33.62, 54: 34.08, 56: 34.27, 58: 33.981: 33.00, 2: 33.55, 3: 33.86, 5: 34.15, 8: 34.27, 10: 34.25, 15: 34.06, 20: 33.75
SR 4×52: 30.15, 54: 30.06, 56: 29.97, 58: 29.561: 28.55, 2: 29.54, 3: 29.17, 5: 29.34, 8: 29.97, 10: 29.97, 15: 29.49, 20: 29.67
Motion61: 30.28, 63: 30.33, 65: 30.35, 67: 30.351: 25.68, 2: 30.35, 3: 29.35, 5: 29.00, 8: 28.67, 10: 27.89, 15: 26.58, 20: 24.88
  • Activation step. Every tested step improves on DiffPIR, including the steps that use fewer NFE than DiffPIR's 100. Against the strongest DiffPIR version of Table R1 (34.08 / 29.26 / 29.20 dB), every step is better on SR and motion; on inpainting the selected step is better and j = 54 matches it. Moving two steps away from the selected step changes PSNR by at most 0.02 dB on motion, 0.29 dB on inpainting and 0.41 dB on SR. The steps were selected on the development images, not on the test set: on SR the selected j = 56 (29.97 dB) is not the best test value (30.15 dB at j = 52, with 132 NFE).
  • K. K = 1 is too few on all three tasks. On the noiseless tasks K = 8–10 is best and every K ≥ 2 matches or exceeds standalone DiffPIR. On noisy motion deblurring, K and the step size jointly set the correction strength: at the selected step size, K = 2–3 improves on DiffPIR and larger K over-corrects (Appendix Fig. 6), so K is selected together with γ on the development images.

Dear Reviewer,

As the discussion period is approaching its end, we would like to kindly follow up on our rebuttal and invite you to take a look at our responses to your concerns.

We would greatly appreciate any feedback you might have, and we would be happy to provide further clarification if needed.

Thank you very much for your time and consideration!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of our responses and the revised PDF

Dear Reviewers and Area Chair,

Thank you for your reviews and the discussion. As the discussion period closes, we summarize where each concern is addressed. All changes are marked in blue in the revised PDF, whose latest revision adds Tables R9–R11; the main text remains within 10 pages.

Reviewer T775 confirmed that our response addresses the concerns about the evaluation protocol and the DiffPIR timing, and Reviewer B58Y noted that the matched-budget comparisons and Table R7 strengthen the empirical evidence and clarify the roles of the components. The remaining points concern the algorithmic contribution (T775, B58Y), the reliability of the score network at the corrected iterates, and the roles of anchoring, step scaling and measurement weighting (B58Y). Our reply to Reviewer T775 on the algorithmic contribution and our two latest replies to Reviewer B58Y, with the new Tables R9–R11, address them.

Concern (raised by)What the revision addsWhere
Matched budgets (1jHd, T775)With the same or a larger NFE for the host solver, LARC exceeds the strongest version of each of the four host solvers by 0.19–2.07 dB in all 12 FFHQ settings§4.2, Table R1; App. A.2.4, Table R3
Run time (T775)Re-measured with the same 100 NFE: 1.440 s for LARC (DiffPIR) vs. 1.430 s for DiffPIR; corrected entry in Table 1Table 1, §4.2; App. A.4.4, Table R5
Settings and tuning (T775)Tuning on 10 development images disjoint from the test images; settings of the baselines and of the host solvers§4.1; App. A.2.2–A.2.3, Table R8
Sensitivity (1jHd)Activation step and number of inner updates KKApp. A.4.3, Table R2
Difference from PnP methods, DDRM, SNIPS, predictor–corrector sampling and DiffPIR (1jHd, T775, B58Y)At the same position with the same NFE, LARC is the only corrector that improves every host solver on every task§1, §2.2; App. A.4.1, Table R6
Roles of anchoring, step scaling and measurement weighting (B58Y)Factor-by-factor ablation; the joint gain of the initialization and GtG_t exceeds the sum of their separate gains on all three tasks§3.1; App. A.1.5 and A.4.2, Tables R7 and R11
Reliability of the score network at the corrected iterates (B58Y)Residual diagnostic; across three controlled adjustments of the noise conditioning, including two that use the ground truth, the largest PSNR gain is 0.08 dB§3.2; App. A.1.6, Tables R9–R10

In short, LARC is a plug-in corrector whose single local energy anchors the correction to the state of the host solver and lets the prior refine only what the measurements leave undetermined. It improves four different host solvers at matched budgets, and its components are examined by controlled ablations (Table 3 and Tables R2, R7, R9–R11).

We would be grateful if you could take this summary into account in the final discussion, and we remain happy to clarify any point before the discussion closes.

Best regards, The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.