{"id":"9ac14353-bf24-429d-a6b2-5167910c84c0","arxiv_id":"2412.00664","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Guided Decoupled Posterior Sampling (GDPS) adds a gradient step on the measurement mismatch ||y - A(x_t)||^2 during the reverse process, improving reconstruction accuracy over DAPS, SITCOM, Resample, and DPS.","lead":"This paper adds a simple data-consistency gradient step to the reverse process of decoupled posterior sampling methods for diffusion-based image reconstruction. The modification consistently improves reconstruction accuracy across linear and nonlinear inverse problems on FFHQ and ImageNet.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The guidance gradient is evaluated on the noisy x_t rather than on a clean estimate, so at high σ_t it is noise-dominated and the claimed data-consistency mechanism is unverified; a noise-level ablation is needed.","rationale":"The empirical support is broad and consistent: GDPS improves over DAPS on essentially every reported task and dataset, and the ablation over N and n in Table 13 is reassuring. I therefore do not think the paper should be rejected. However, the load-bearing mechanistic claim is currently supported only by end-to-end metrics with per-task-tuned γ and no error bars. The concern is specifically that Eq. (6) applies the data-consistency loss to the noisy sample x_t rather than to the clean estimate, so the gradient at high noise is not a posterior-consistent direction. The proposed noise-level ablation would settle whether the claimed mechanism is real or whether the gains come from a low-noise projection artifact. I also note an internal numerical inconsistency worth a careful pass: in Table 13, DAPS on SR with N=200, n=5 has SSIM 0.978, while the same configuration in Table 1 has SSIM 0.816, despite matching PSNR and LPIPS. This does not change the verdict but reinforces the need for a full numerical re-check. The reader's CONDITIONAL verdict already accounts for the absence of statistical quantification and tuned hyperparameters, so my read leaves it unchanged.","tokens_in":16098,"tokens_out":9434,"duration_ms":93979,"concrete_test":"Run the FFHQ super-resolution task (Table 1) with GDPS in four configurations: full guidance, guidance only at σ(t_k) below the median schedule value (low-noise half), guidance only above the median (high-noise half), and no guidance (DAPS), keeping the same γ, N, n, and Langevin steps. Report means and 95% confidence intervals over the 100 images. If the high-noise-only run is statistically indistinguishable from DAPS while the low-noise-only run retains the Table 1 gains, the central mechanism is not operative where the paper claims it is; if the high-noise-only run also improves, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central assumption is that Eq. (6) / Algorithm 1 line 4 is a data-consistency correction that steers the reverse trajectory toward the posterior. This is unverified in precisely the regime where decoupled sampling is supposed to help. The measurement y constrains x0, not the intermediate noisy sample x_t. For linear A, writing x_t = x0 + σ_t ε gives ∇_{x_t}‖y−A(x_t)‖² = −2Aᵀ(y−A x_t) = −2Aᵀn + 2σ_t AᵀA ε. At large σ_t the noise-dependent term dominates, so the step mainly damps the component of ε in range(AᵀA); it does not directly reduce ‖y−A(x0)‖². The update also pushes x_t off the marginal distribution on which the score model was trained, biasing subsequent score evaluations. Since γ is hand-tuned per task and per variant over a 100× range (Appendix C, Tables 11–12), the consistent end-to-end gains could come from a low-noise projection/denoising effect rather than from the claimed 'smooth transition' toward the target distribution. The paper provides no analysis of how γ should scale with σ_t, no noise-level ablation, and no check that the guidance direction correlates with the posterior-consistent direction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Guided Decoupled Posterior Sampling (GDPS), a modification of decoupled posterior sampling methods such as DAPS. The only algorithmic change is an extra gradient-descent step in the reverse process, x_t^guided = x_t - gamma * grad_{x_t} ||y - A(x_t)||^2 (Eq. (6), Algorithm 1 line 4), intended to inject measurement information before the Langevin optimization stage. The guidance step is also adapted to latent diffusion models (G-LatentDAPS) and to SITCOM's Tweedie-based update (G-SITCOM). The paper reports PSNR, SSIM, and LPIPS on 100-image FFHQ and ImageNet validation sets across five linear and three nonlinear inverse problems, plus four challenging settings, and claims consistent state-of-the-art improvements over DAPS, LatentDAPS, SITCOM, Resample, PSLD, and DPS.","tokens_in":1585,"tokens_out":3470,"duration_ms":67340,"significance":"If the empirical gains are robust, GDPS is a simple, nearly cost-free plug-in enhancement to a popular family of diffusion-based inverse problem solvers, and the latent and Tweedie extensions broaden its applicability. The paper's strengths are its breadth of experiments: two datasets, linear and nonlinear operators, challenging regimes, and two additional base methods, with running time comparable to DAPS (Appendix B). The central limitation is that the proposed mechanism is not analyzed: the gradient is applied to the noisy intermediate sample, gamma is hand-tuned per task over a 100x range, and no variance or significance information is given for the reported improvements. The consistent gains are suggestive but the explanatory claim of a 'smoother transition' remains unquantified.","major_comments":[{"comment":"The data-consistency gradient is evaluated at the noisy sample x_t rather than at a clean estimate of x0. For a linear forward operator, writing x_t = x0 + sigma_t * epsilon gives grad_{x_t} ||y - A(x_t)||^2 = -2A^T(y - A x0) + 2 sigma_t A^T A epsilon. At large sigma_t the noise-dependent term dominates, so the step is not obviously reducing ||y - A(x0)||^2; it may only damp components of the noise in range(A^T A). The paper provides no analysis of how this update interacts with the probability-flow ODE, no scaling law for gamma in sigma_t, and no noise-level ablation. This is load-bearing because the abstract and Section 3.2 claim that the constraint provides a 'smoother transition' toward the target distribution. I would like to see either an analysis of the update in the high-noise regime or an experiment that reports the contribution of the guidance step as a function of noise level, e.g., by applying the update only above a threshold sigma_t or by re-scaling gamma with sigma_t.","section":"Section 3.2, Eq. (6) and Algorithm 1 line 4"},{"comment":"The step size gamma is selected separately for every task and every variant, with values ranging from 0.1 to 10 (a 100x range). No sensitivity analysis is reported, and no principled selection rule is given. Since gamma is the only new hyperparameter of the method, the consistent end-to-end gains could in principle come from favorable per-task tuning rather than from the proposed mechanism. I request a sensitivity study for Table 1 with respect to gamma (at least for two or three tasks), or a data-driven schedule for gamma that does not require per-task hand-tuning.","section":"Appendix C, Tables 11 and 12"},{"comment":"All metrics are reported as single-point estimates over 100 images, with no error bars, standard deviations, or paired significance tests. Many of the reported advantages are small, for example Table 4 phase retrieval PSNR 22.72 vs 22.17 and SSIM 0.580 vs 0.577, and Table 8 HDR PSNR 24.82 for both LatentDAPS and G-LatentDAPS. To support the claim that GDPS 'consistently' outperforms the baselines, the authors should report variance across random seeds or per-image paired statistics (e.g., Wilcoxon signed-rank tests) for at least the main tables.","section":"Tables 1-6 and 8-9"},{"comment":"The claimed 'smoother transition' and 'more effective convergence toward the target distribution' are not quantified anywhere in the paper. There is no trajectory analysis, no plot of the data-consistency loss over the reverse process, and no comparison of the guided and unguided reverse trajectories. This is a presentation issue for the main claimed mechanism, and it should be either substantiated with a quantitative diagnostic or softened to a qualitative observation.","section":"Abstract and Section 3.2"}],"minor_comments":[{"comment":"The update line has a missing closing parenthesis: 's_theta(x_{t_k}, sigma(t_k) Delta t' should read 's_theta(x_{t_k}, sigma(t_k)) Delta t'.","section":"Algorithm 1, line 3"},{"comment":"Several captions (Figures 3-6, 7-11) contain the typo 'taks' instead of 'tasks'.","section":"Figure captions"},{"comment":"For G-SITCOM the guidance is applied to the clean Tweedie estimate (denoted by a hat), whereas Eq. (6) applies it to the noisy x_t. The paper should state explicitly that the guidance acts on different objects in the pixel/latent/Tweedie variants, since the claimed mechanism is not identical across extensions.","section":"Section 3.3 and Algorithm 3"},{"comment":"The displayed definition of F(x,y) is garbled in the text and should be reformatted for readability.","section":"Equation (11)"},{"comment":"Reference [29] gives incomplete page information ('pages 681-') and should be completed.","section":"Reference [29]"}],"recommendation":"major_revision","confidential_remarks":"The paper would benefit from releasing code and from a clear statement of how gamma was selected for each cell, as the per-task tuning is central to reproducibility. The stress-test concern about the high-noise behavior of Eq. (6) is real and should be addressed with an experiment rather than with a purely analytic argument, since the method is ultimately empirical."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper proposes GDPS, a one-line modification to DAPS that adds a data-consistency gradient step to the reverse process. On the tested benchmarks, it consistently improves over DAPS and other baselines, usually by 0.2–1 dB PSNR. The contribution is a legitimate plug-in enhancement: it is simple, applicable to latent diffusion and Tweedie-based variants, and the empirical coverage is broad (FFHQ, ImageNet, linear and nonlinear tasks, plus challenging settings). This is the correct first thing to know.\n\nThe main soft spot is the unexamined mechanism. As the stress-test note observes, for linear A the guidance gradient is -2Aᵀn + 2σ_t AᵀA ε. At high noise levels the second term dominates, so the step mostly damps the noise component in the range of AᵀA. That could be a denoising or projection effect rather than a posterior-steering one. The paper calls the transition 'smoother' but doesn't quantify it, and there is no analysis of how γ should scale with σ_t. The per-task tuning of γ over a wide range (Tables 11–12) is also a concern, because it invites the question of whether the gains are robust or a product of fitting.\n\nThat said, the empirical claim is consistent across many cells, which is more than many papers offer. The lack of error bars is a real but fixable issue; even a single-seed study would benefit from a sensitivity analysis on γ (e.g., a coarse grid) to show the improvement is not a knife-edge effect. I'd also ask for a noise-level ablation, like running the guidance only at low σ_t, to test whether the mechanism is indeed about data consistency or just denoising.\n\nOverall, the central claim holds up in the reported experiments. The paper is a useful incremental advance for the decoupled-sampling subfield. It deserves a serious referee, though the revision should address variance, gamma sensitivity, and the high-noise behavior. I'd bring it to a reading group focused on diffusion inverse problems, and I'd cite it if I worked on decoupled sampling.","headline":"A simple, plausibly useful plug-in guidance step for decoupled posterior sampling, with consistent but modest empirical gains and a real need for variance reporting and analysis of the guidance mechanism.","tokens_in":16906,"tokens_out":3080,"would_cite":true,"duration_ms":28203,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding a data-consistency gradient step to the reverse process of decoupled posterior sampling improves reconstruction accuracy on linear and nonlinear inverse problems.","keywords":["inverse problems","diffusion models","posterior sampling","decoupled posterior sampling","data consistency","image restoration","phase retrieval","latent diffusion models"],"falsifier":"Take a fixed inverse problem and compare GDPS with an ablation that applies the gradient step only after the first half of the reverse steps, keeping $\\gamma$ fixed. If the ablated version matches GDPS in PSNR and LPIPS, then the guidance on nearly pure-noise samples carries no useful information, contradicting the paper's picture of a smooth, measurement-guided transition. A complementary check is to run both methods at a measurement noise level well above $\\sigma = 0.05$, such as $\\sigma = 0.2$, and see whether the GDPS advantage shrinks or vanishes.","tokens_in":15882,"feed_emoji":"🖼️","tokens_out":6108,"duration_ms":51788,"temperature":0.7,"pith_summary":"Decoupled posterior sampling methods for inverse problems separate denoising, optimization, and renoising, and they suffer because the denoising step ignores the measurement. This paper proposes Guided Decoupled Posterior Sampling (GDPS), which inserts a single gradient-descent step on the squared data-consistency loss, $\\|y - A(x_t)\\|^2$, after each reverse-process ODE step. The guided state is then fed to the Langevin optimization stage, so the reconstruction starts from a measurement-aware point. The paper reports that this modification improves PSNR, SSIM, and LPIPS over DAPS and other baselines on FFHQ and ImageNet across linear and nonlinear tasks, and that the same step improves LatentDAPS and SITCOM when applied in latent space or through Tweedie's formula.","feed_headline":"One guidance step lifts diffusion inverse-problem accuracy","feed_subtitle":"GDPS inserts a data-consistency gradient into the reverse process, improving DAPS and SITCOM on FFHQ and ImageNet.","key_machinery":"The carrying object is the guided reverse update $x_t^{\\text{guided}} = x_t - \\gamma \\nabla_{x_t}\\|y - A(x_t)\\|^2$, a single gradient-descent step on the squared data-consistency loss applied after each probability-flow ODE step. In latent diffusion models the step is taken in latent space through the decoder, $z_t^{\\text{guided}} = z_t - \\gamma \\nabla_{z_t}\\|y - A(D(z_t))\\|^2$; in Tweedie-based sampling it is applied to the estimated clean image, $\\hat{x}_0^{\\text{guided}} = \\hat{x}_0 - \\gamma \\nabla_{\\hat{x}_0}\\|y - A(\\hat{x}_0)\\|^2$. This gradient term is what carries measurement information into the reverse process, while the scalar $\\gamma$ controls how strongly the state is pulled toward consistency.","core_discovery":"The central claim is that the reverse process of decoupled posterior sampling should not be run blind to the measurement. The paper's update, $x_t^{\\text{guided}} = x_t - \\gamma \\nabla_{x_t}\\|y - A(x_t)\\|^2$, is applied after each inner ODE step in Algorithm 1, and this one-line change is enough to make the optimization stage start from a state already pulled toward data consistency. On the tested tasks, including super-resolution, inpainting, deblurring, phase retrieval, nonlinear deblurring, and high dynamic range, with measurement noise $\\sigma = 0.05$, GDPS reports the best PSNR, SSIM, and LPIPS among the compared methods on both FFHQ 256x256 and ImageNet 256x256, and the improvement persists on harder versions of those tasks. The paper also claims the guidance transfers: G-LatentDAPS and G-SITCOM outperform their base methods.","pith_inferences":["One editorial extension is that $\\gamma$, currently hand-set per task, could be made noise-adaptive: a schedule $\\gamma(\\sigma_t)$ that shrinks as the sample becomes cleaner would likely remove the per-task tuning shown in the appendix.","Another extension is to read the guidance step as a lightweight measurement-conditional perturbation of the probability-flow ODE; that reading suggests testing it with other ODE solvers, such as higher-order or few-step samplers, to see whether the benefit survives when the reverse trajectory is shorter.","Because the guidance is computed on noisy samples, its behavior at very high noise levels is the most likely failure mode; an ablation that applies the gradient only after the sample has been substantially denoised would isolate where the gain actually comes from."],"forward_implications":["If the central claim is right, any decoupled posterior sampling method can be improved by adding the data-consistency gradient to its reverse process, without changing its optimization or forward processes.","The improvement should appear on both linear and nonlinear forward operators, because the gradient is computed directly from $A$, not from a linearized adjoint.","The gains should hold under harder measurement regimes, including 16x super-resolution, heavy masking, and phase retrieval with low oversampling.","The guidance step adds only a few seconds per image relative to DAPS, so the accuracy gain does not require a large computational cost."],"supporting_citations":[{"why":"Defines the decoupled posterior sampling loop that GDPS modifies and supplies the main DAPS baseline, including the latent variant.","marker":"[22]"},{"why":"Provides the Tweedie-based SITCOM method that GDPS extends into G-SITCOM, and serves as a comparison baseline.","marker":"[24]"},{"why":"Supplies the pre-trained score model used for the pixel-space experiments and is one of the baselines GDPS must beat.","marker":"[14]"},{"why":"Provides the probability-flow ODE solver whose per-step output GDPS corrects with the guidance term.","marker":"[25]"},{"why":"Serves as a baseline in the linear and nonlinear experiments and provides the latent diffusion model settings used in the latent extensions.","marker":"[35]"}],"fun_headline_variants":["A single guidance step fixes blind reverse diffusion","GDPS guides reverse diffusion with measurement data","Data consistency in reverse diffusion step boosts inverse accuracy","Guided reverse step prevents diffusion error accumulation","One-line change in decoupled posterior sampling yields SOTA results"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benefit rests on the assumption that the gradient of $\\|y - A(x_t)\\|^2$ with respect to the noisy sample $x_t$ points toward a measurement-consistent reconstruction. At early reverse steps, where $x_t$ is close to pure noise, that gradient may be dominated by noise rather than by signal, and the paper does not analyze how the guidance interacts with the probability-flow ODE or how $\\gamma$ should scale with the noise level.","fun_headline_variants_meta":{"raw":{"variants":["A single guidance step fixes blind reverse diffusion","GDPS guides reverse diffusion with measurement data","Data consistency in reverse diffusion step boosts inverse accuracy","Guided reverse step prevents diffusion error accumulation","One-line change in decoupled posterior sampling yields SOTA results"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000611,"raw_usage":{"total_tokens":2817,"prompt_tokens":896,"completion_tokens":1921,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":1850}},"tokens_in":512,"tokens_out":1921,"duration_ms":12290,"temperature":1.0,"reasoning_tokens":1850,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:07:41.595277+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed inverse problem and compare GDPS with an ablation that applies the gradient step only after the first half of the reverse steps, keeping $\\gamma$ fixed. If the ablated version matches GDPS in PSNR and LPIPS, then the guidance on nearly pure-noise samples carries no useful information, contradicting the paper's picture of a smooth, measurement-guided transition. A complementary check is to run both methods at a measurement noise level well above $\\sigma = 0.05$, such as $\\sigma = 0.2$, and see whether the GDPS advantage shrinks or vanishes.","supporting_citations":[],"review_version":1}