{"id":"0dfa13f3-6d9e-418c-bc41-90ea12af1b76","arxiv_id":"2412.03941","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Measurements Optimization, which alternates SGLD steps on the measurement objective with denoiser projection, achieves SOTA or near-SOTA image restoration at 50-100 diffusion NFEs across eight linear and nonlinear tasks.","lead":"This paper introduces Measurements Optimization (MO), a plug-in module for diffusion-based inverse problem solvers that runs several stochastic-gradient Langevin updates on the measurement fit and then denoises through a pretrained diffusion model at every sampling step. If the reported results hold, it would cut the number of diffusion evaluations by roughly an order of magnitude while preserving or improving image quality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The noising/denoising projection in Algorithm 1 is the theoretical hinge of MO, and the paper's support for it (Theorem 1) does not cover real diffusion models; a direct measurement-residual check is needed.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing issue that I would stress-test: the projection step in Algorithm 1 is where the algorithm's correctness claim rests, and Theorem 1 does not provide a rigorous guarantee for real diffusion models. The empirical tables are suggestive, but they do not isolate this step, and the lack of error bars or code means a failure here would undermine the central claim rather than merely weaken it. I do not find a more decisive internal inconsistency: the ODE update in Algorithm 2 is consistent with EDM when s(t)=1, and the reported hyperparameters, while numerous, are not themselves contradictory. Therefore, my read does not move the verdict; it remains CONDITIONAL, pending the concrete residual-based check and, ideally, code and error-bar reporting.","tokens_in":17019,"tokens_out":7263,"duration_ms":69777,"concrete_test":"Run DPS-MO on FFHQ box inpainting with the Table 1 configuration (50 NFEs) and, for every diffusion step before the ODE update, compute ρ_t = ||y − A(Dθ(xSGLD_t + σ(t)ε))|| / ||y − A(xSGLD_t)||, the ratio of measurement residuals after versus before the noising/denoising projection. If ρ_t > 1 for most steps with σ(t) > 1, the projection is actively discarding measurement information and the §4 mechanism is not what drives the reported gains; if ρ_t ≤ 1 for those steps, the projection preserves the measurement constraints and the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of Measurements Optimization is the alternating loop in Algorithm 1: SGLD drives x toward measurement consistency, then lines 9–10 add σ(t)-noise and denoise with the pretrained diffusion model Dθ, supposedly projecting the SGLD solution back onto the natural-image manifold while retaining the measurement information gained. The only theoretical support offered is Theorem 1, adapted from EDM §B.3, which characterizes the optimal denoiser for an empirical distribution with a finite training set and an ideal denoiser. That result is not a statement about a real diffusion model at finite σ, and even in the idealized setting the optimal denoiser returns a weighted average of training images—not an element of the training set—so the claim that the projection 'pulls back onto the training data manifold' is not actually established. If, at the large σ values used early in the schedule, the denoiser removes the measurement information that SGLD has injected, then the 50–100 NFE success would not be explained by the proposed mechanism. The paper reports no measurement residuals after the denoising step and no ablation that isolates this projection, so the concern is unresolved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Measurements Optimization (MO), a plug-and-play module for diffusion-based inverse problem solving. At each diffusion step, MO runs several SGLD steps on the measurement loss and then adds noise to the resulting image and denoises it with a pretrained diffusion model, aiming to project the solution back onto the image manifold. The module is integrated into DPS and Red-diff, yielding DPS-MO and Red-diff-MO. The authors report state-of-the-art or near-state-of-the-art performance on five linear and three nonlinear tasks on FFHQ-256 and ImageNet-256, typically with 50-100 NFEs (with ImageNet phase retrieval requiring 1000 NFEs). They also provide ablations over sampling schedules, optimizer choice, SGLD step count, and memory/time comparisons.","tokens_in":17236,"tokens_out":6547,"duration_ms":55244,"significance":"If the empirical claims hold, the paper offers a practically valuable 10-40x speedup over existing diffusion-based inverse problem solvers, with a simple module that is orthogonal to the underlying sampler and can be inserted into existing methods. The paper is thorough in task coverage, clearly specifies the algorithm, and provides extensive qualitative results. The significance is moderated, however, by the weak theoretical support for the core manifold-projection mechanism and by evaluation choices that do not quantify uncertainty and involve substantial per-task hyperparameter tuning. The claimed SOTA gains would be more convincing with variance-aware reporting and a direct check of measurement consistency after the denoising step.","major_comments":[{"comment":"The claim that the noising/denoising step \"projects the optimization solution back onto the training data manifold\" is not supported by Theorem 1. The theorem, adapted from EDM §B.3, characterizes the optimal denoiser for an empirical distribution as a weighted average of training images; a weighted average is not an element of the training set, so the theorem does not establish a manifold projection in the sense used in the paper. Moreover, the result assumes an ideal denoiser and a finite training set, and the paper provides no evidence that a real diffusion denoiser at finite σ preserves the measurement information introduced by the SGLD steps. I recommend adding a direct check of the measurement residual after the denoising step, e.g., ||y − A(Dθ(xSGLD_t + σ(t)ε, σ(t)))||, and an ablation that isolates the effect of the projection (e.g., skipping it or replacing it with a simple proximal step).","section":"§5.1–5.2, Tables 1–4"},{"comment":"All quantitative comparisons are reported as point estimates (means over 100 images) with no variance or confidence intervals, and for nonlinear tasks the metrics are the best over four independent samples. The headline difference, e.g., DPS-MO 28.71 dB vs. DAPS 27.12 dB on FFHQ HDR, could be within run-to-run variability, especially under best-of-four selection. Please report means ± standard deviations (or confidence intervals) for at least the key DPS-MO and best-baseline comparisons, and clarify whether the best-of-four selection is applied to baselines as well.","section":"§5.1–5.2, Tables 1–4"},{"comment":"Many hyperparameters are tuned per task and dataset: NFE, σmax, σmin, and the number of SGLD steps N_SGLD are chosen to \"achieve the best performance\" via grid search, and σmax is set to 1 (rather than the default 80) for ImageNet SR and random inpainting. This makes the 50–100 NFE claim conditional on substantial per-task calibration and weakens the \"plug-and-play\" generality. Please state the selection procedure (e.g., a validation split) and report sensitivity of the main results to these hyperparameters, or at least discuss the cost of tuning relative to the baselines.","section":"§5.2, Table 8, Appendix 8.3"}],"minor_comments":[{"comment":"The SGLD update in Algorithm 1 uses a single learning rate η, but Appendix 8.4 defines a time-dependent schedule η_i with p=2, r=0.01 and a loss scaled by 1/(2τ²) with τ=0.01. The algorithm and main text should be aligned with the implementation for reproducibility.","section":"Appendix 8.4"},{"comment":"The section title contains a typo: \"Performace\" should be \"Performance.\"","section":"§5.3"},{"comment":"The caption says \"100 and 100 NFEs\" but should presumably read \"1000 and 1000 NFEs\"; the same issue appears in Table 12. Please check all captions for this typo.","section":"Table 10 caption"},{"comment":"The axis labels contain \"uni00A0\" artifacts (non-breaking space characters) that should be removed before camera-ready.","section":"Figures 3 and 5"},{"comment":"The claim that SGLD \"guarantees asymptotic convergence to global minima in non-convex settings\" with reference [7] is stronger than what Gelfand and Mitter (1991) prove, which concerns annealed Langevin dynamics under specific conditions. Please soften the claim or cite a more specific SGLD convergence result.","section":"§4, references"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the proposed method is potentially useful, but the empirical evidence needs strengthening: the core theoretical justification is overstated, the comparisons lack uncertainty quantification, and the hyperparameter tuning is extensive. These issues are addressable with additional experiments and careful reporting, so I recommend major revision rather than rejection. If the authors can provide a measurement-residual check and variance-aware results, the paper would be much more convincing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, the thing to know about 2412.03941: it's an engineering paper in the best sense. The authors take known ingredients—SGLD and a pretrained diffusion denoiser—and combine them into a per-step inner loop they call Measurements Optimization. Re-running SGLD from the previous denoised estimate at each diffusion step, then adding noise and denoising, consistently beats DPS and DAPS on eight linear and nonlinear tasks at 50–100 NFEs. The reported 10–40x NFE reduction is credible, and the central claim holds up as an empirical statement.\n\nWhat's new is the specific recipe, not the parts. SGLD and noising/denoising projections are standard, but the combination with re-initialization from the previous step, and the demonstration that it works as a plug-in for both DPS and Red-diff, is a real contribution. The paper is well-ablated: it tests optimizer choice, sampling schedule, SGLD step count, and initialization, and the supplementary shows that re-solving SGLD at each step matters (Table 9). It also reports memory and wall-clock time, which is more than many papers in this area bother to do. The closing limitation about non-differentiable or expensive forward operators is honest.\n\nThe soft spots are in the theory and the evaluation protocol. The theoretical justification for the projection step, Theorem 1, is adapted from EDM and assumes an ideal denoiser on a finite training set. That does not cover the real, finite-sigma diffusion models used here. The stress-test concern is legitimate: no measurement-residual check is reported after the denoising step, so the claim that the projection retains the measurement information while returning to the manifold is not directly verified. This is the weakest link, but it is a weakness of explanation, not a demonstrated failure. The empirical results stand on their own.\n\nThe evaluation has shortcuts that a referee should push on: best-of-four selection for nonlinear tasks, means over 100 images with no error bars, per-task hyperparameters tuned on the test set, and baselines copied from DAPS rather than rerun. These are common in this literature but they do limit how strongly the SOTA claims can be stated. And there is no code or data link, which makes independent verification harder.\n\nWho should read it: anyone building diffusion-based inverse solvers. It is a useful, practical result and deserves a serious referee. My recommendation: send it to review, but ask for code release, error bars or a fixed evaluation protocol, and a direct check of measurement consistency after the denoising projection.","headline":"A fast, well-ablated plug-in for diffusion inverse solvers; the empirical story is convincing enough to referee, but the projection theory is thin and the evaluation shortcuts need tightening.","tokens_in":17800,"tokens_out":3639,"would_cite":true,"duration_ms":32247,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that repeatedly fitting the measurement with Stochastic Gradient Langevin Dynamics between diffusion denoising steps turns 1000–4000-step inverse problem solvers into 50–100-step solvers without sacrificing quality.","keywords":["diffusion models","inverse problems","measurements optimization","stochastic gradient Langevin dynamics","plug-and-play module","image restoration","posterior sampling","phase retrieval"],"falsifier":"Run MO on any task with a known forward operator and compare the measurement residual $\\|y - A(\\hat{x}_0)\\|_2$ immediately before and after the denoising query at several noise levels; if denoising consistently raises the residual above the level SGLD already reached, the manifold projection is discarding measurement information and the alternating scheme cannot converge to a measurement-consistent solution.","tokens_in":16774,"feed_emoji":"⚡","tokens_out":14028,"duration_ms":104467,"temperature":0.7,"pith_summary":"Diffusion models can restore images from corrupted measurements, but current solvers need 1000 to 4000 diffusion evaluations per image because they feed in only a sliver of measurement information at each step. This paper introduces Measurements Optimization (MO), a plug-in module that injects far more measurement information per step: it first runs Stochastic Gradient Langevin Dynamics to fit an image to the measurement, then adds noise and denoises with a pretrained diffusion model to pull the fit back onto the natural-image manifold. Plugging MO into existing solvers such as DPS and Red-diff cuts the required evaluations to 50–100 for most tasks, a 10 to 40-fold speedup, while matching or beating the previous state of the art in quality.","feed_headline":"Diffusion inverse problems solved in 50–100 steps","feed_subtitle":"A plug-in module fits measurements harder per step, so 1000-step solvers match or beat their past quality in a tenth of the time.","key_machinery":"The load-bearing object is the Measurements Optimization (MO) module, an inner-outer loop wrapped around any diffusion sampler. At each diffusion time step $t$, an inner loop of $N$ stochastic gradient ascent steps of the form $x \\leftarrow x + \\eta \\nabla_x \\|y - A(x)\\|_2^2 + \\sqrt{2\\eta}\\,\\epsilon$ with $\\epsilon \\sim \\mathcal{N}(0, I)$ fits an image to the measurement (this is SGLD); then the outer loop noises the result to level $\\sigma(t)$ and feeds it to the mean-predicted denoiser $D_\\theta$, which returns the estimate $\\hat{x}_0$ that seeds the next step. Theorem 1 (adapted from Section B.3 of EDM) justifies the denoising query: an ideal denoiser trained on a finite dataset returns a likelihood-weighted mixture of training images, so the query projects the SGLD solution onto the training-data manifold, and the Gaussian noise added by SGLD lets the diffusion model overwrite unrecoverable content. The paper uses the EDM schedule $s(t)=1$, $\\sigma(t)=t$ with a $\\rho=7$ discrete time grid, and sets the SGLD learning rate to $5 \\times 10^{-5}$ for all tasks except ImageNet phase retrieval.","core_discovery":"The central claim is that the high NFE cost of diffusion inverse problem solving is not an inherent property of the diffusion prior but a consequence of injecting too little measurement information per step. Existing methods take a single gradient step on the fidelity loss $\\|y - A(\\hat{x}_0)\\|_2^2$ per diffusion evaluation; MO instead runs an inner SGLD loop that repeatedly minimizes $\\|y - A(x)\\|_2^2$, then asks the pretrained denoiser $D_\\theta(x; \\sigma(t))$ to pull the result back onto the data manifold via a noise-add-and-denoise query. The paper asserts that this two-step iteration, run at every diffusion time step, achieves state-of-the-art or near-state-of-the-art results on five linear and three nonlinear tasks on FFHQ 256 and ImageNet 256, with at most 100 function evaluations per image in all cases except ImageNet phase retrieval (1000). The headline numbers are DPS-MO reaching 28.71 dB PSNR on FFHQ HDR at 100 NFEs versus DAPS's 27.12 dB at 4000 NFEs, and Red-diff-MO raising Red-diff's phase-retrieval PSNR from 15.60 to 27.34 dB at a tenth of the NFEs.","pith_inferences":["Because MO only needs gradient access to the forward operator, the same inner SGLD loop should extend to other differentiable measurement models such as magnetic resonance or computed tomography reconstructions without retraining; the paper demonstrates the generality only through its eight tasks.","The fresh Gaussian noise injected by SGLD at every inner step suggests MO is sampling from an approximate posterior rather than merely minimizing a loss, so the module could be repurposed to generate diverse measurement-consistent samples beyond the four-run phase-retrieval and inpainting illustrations.","A direct stress test would be to run MO under larger measurement noise or with fewer SGLD steps; the paper fixes Gaussian noise at $\\sigma_n = 0.05$ and tunes the step count per task, so the claimed 10–40-fold speedup under realistic noise levels remains untested."],"forward_implications":["DPS-MO reaches state-of-the-art or near-state-of-the-art quality on eight linear and nonlinear tasks using no more than 100 NFEs, except for ImageNet phase retrieval which uses 1000 NFEs (Tables 1–4).","MO is a plug-and-play module: integrating it into DPS yields DPS-MO and into Red-diff yields Red-diff-MO, and both improve on their base methods; Red-diff-MO raises phase-retrieval PSNR from 15.60 to 27.34 dB at 100 NFEs versus 1000 (Table 5).","The module reduces wall-clock time and memory overhead relative to DPS: DPS-MO produces an inpainted image in 5.3 seconds versus over 60 seconds for DPS and Red-diff, with memory comparable to Red-diff (Table 7).","The choice of inner optimizer matters: SGLD clearly outperforms Adam for inpainting, and the authors attribute this to the Gaussian noise SGLD adds, which the diffusion model can then use to redraw masked or corrupted content (Table 6).","Refreshing the SGLD solution at every diffusion step with an initialization from the previous denoised output is essential; reusing a single SGLD solution across all steps degrades LPIPS/PSNR substantially (Table 9)."],"supporting_citations":[{"why":"The sampling-based solver that MO plugs into; supplies the baseline posterior-sampling framework and the comparison numbers for DPS.","marker":"[5]"},{"why":"Provides the pretrained diffusion model design, the EDM sampling schedule used in all experiments, and the proof of Theorem 1 on denoiser manifold projection.","marker":"[11]"},{"why":"The training-based solver that MO plugs into; supplies the second integration target and the baseline whose phase-retrieval and HDR results MO improves.","marker":"[15]"},{"why":"Introduces Stochastic Gradient Langevin Dynamics, the inner optimizer that fits measurements aggressively and injects Gaussian noise.","marker":"[29]"},{"why":"The strongest nonlinear-task baseline; supplies the task configuration, evaluation protocol, and baseline numbers that the MO comparisons are measured against.","marker":"[30]"},{"why":"Non-asymptotic analysis of SGLD in non-convex learning, cited as the reason SGLD is well-suited to highly non-convex forward operators.","marker":"[18]"}],"fun_headline_variants":["Measurement optimization cuts diffusion steps to 100","Plug-in module solves inverse problems in 100 steps or less","100-step diffusion inverse solvers match 4000-step quality","Measurement optimization accelerates diffusion inverse solving"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme rests on the assumption that a real pretrained diffusion denoiser, when queried by adding noise and denoising, reliably returns an estimate that stays on the natural-image manifold while still carrying the measurement information the SGLD steps just encoded; if that projection corrupts or discards the measurement fit, the alternating loop cannot converge to a correct solution.","fun_headline_variants_meta":{"raw":{"variants":["Measurement optimization cuts diffusion steps to 100","Plug-in module solves inverse problems in 100 steps or less","100-step diffusion inverse solvers match 4000-step quality","Measurement optimization accelerates diffusion inverse solving"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00033,"raw_usage":{"total_tokens":1916,"prompt_tokens":1100,"completion_tokens":816,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":716,"completion_tokens_details":{"reasoning_tokens":755}},"tokens_in":716,"tokens_out":816,"duration_ms":7144,"temperature":1.0,"reasoning_tokens":755,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:55:07.414428+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MO on any task with a known forward operator and compare the measurement residual $\\|y - A(\\hat{x}_0)\\|_2$ immediately before and after the denoising query at several noise levels; if denoising consistently raises the residual above the level SGLD already reached, the manifold projection is discarding measurement information and the alternating scheme cannot converge to a measurement-consistent solution.","supporting_citations":[{"cited_title":"Elucidating the design space of diffusion-based generative models","cited_arxiv_id":null,"evidence_quote":"Provides the pretrained diffusion model design, the EDM sampling schedule used in all experiments, and the proof of Theorem 1 on denoiser manifold projection."},{"cited_title":"Bayesian learning via stochas- tic gradient langevin dynamics","cited_arxiv_id":null,"evidence_quote":"Introduces Stochastic Gradient Langevin Dynamics, the inner optimizer that fits measurements aggressively and injects Gaussian noise."},{"cited_title":"Non-convex learning via stochastic gradient langevin dy- namics: a nonasymptotic analysis","cited_arxiv_id":null,"evidence_quote":"Non-asymptotic analysis of SGLD in non-convex learning, cited as the reason SGLD is well-suited to highly non-convex forward operators."}],"review_version":1}