{"id":"ec17539a-847c-43e4-bbfc-cb9b2531ae2b","arxiv_id":"2510.02208","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"MACS replaces the variance term in aDDIM with a measurement residual, yielding better FID/KID than fast ODE baselines in two-step consistency-model inverse problems.","lead":"This paper proposes MACS, a test-time change to how consistency models sample during image reconstruction: it injects extra noise scaled by how far the current estimate is from the measured data. It improves two-step perceptual reconstruction (FID/KID) across super-resolution, inpainting, and deblurring on Fashion-MNIST and LSUN Bedroom, without retraining the generative backbone.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation shows the measurement-residual term is the active ingredient; FID/KID gains may simply reflect added stochasticity, not measurement-awareness.","rationale":"The reader's weakest_assumption centers on the theoretical derivation (inequality vs. equality) for replacing the variance term with the measurement residual. That is a valid concern, but it is not the most load-bearing for the central claim, which is empirical: MACS gives consistent perceptual improvements. For that claim to substantiate the proposed mechanism, one must show the residual term itself is responsible. The update rule combines DDIM with an extra stochasticity term; any positive R (including a constant) will alter the noise level. The paper provides no control experiment distinguishing between 'measurement-aware adaptive noise' and 'simply more noise'. Since increased stochasticity is a well-known way to improve FID/KID, the empirical gains might have nothing to do with measurement awareness. This is a concrete, experimentally decidable issue. The theoretical derivation, while flawed, would be secondary: even a perfect derivation would not establish that the residual is causally responsible for the gains. Therefore the missing ablation is the single most load-bearing concern. The reader's conditional verdict already calls for revisions, so I maintain the same verdict rather than escalating; the concern reinforces the need for an ablation but does not by itself invalidate the method's usefulness.","tokens_in":14336,"tokens_out":7538,"duration_ms":60540,"concrete_test":"Run the Fashion-MNIST inpainting experiment (Table II) under three settings: (1) MACS as proposed; (2) R = γ·c where c is the mean residual over the validation set (a constant per task); (3) R = γ·||ε̂||² (depends only on the CM estimate, not the measurement y). Keep all other hyperparameters fixed and report FID/KID with multiple seeds to assess variance. If settings (2) or (3) achieve FID/KID within a small margin (e.g., 5%) of MACS, the measurement-residual term is not the active ingredient and the measurement-awareness claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MACS regulates stochasticity through the measurement residual R = γ||y−A(x̂)||², and that this measurement-awareness drives the reported perceptual improvements. In the update (Eq. 36), the added noise has squared norm ρ||ε̂||² + (1−√ρ)²R. For any γ>0, this increases stochasticity relative to DDIM (γ=0). The paper compares against DDIM-like multistep sampling and ODE solvers, but never against a control where R is replaced by a y-independent scalar. Since extra stochasticity is known to improve FID/KID at the cost of distortion (the perception–distortion tradeoff the authors themselves cite), the observed gains could arise purely from increased noise, with the specific measurement dependence being irrelevant. The theoretical derivation does not rescue this: Eqs. (37)–(41) only establish an upper bound E||y−A(x̂)||² ≤ ||A||₂²||x−x̂||² + mσ_y², and for rank-deficient operators (inpainting, super-resolution) the residual can be much smaller than the conditional variance, so the residual is not a faithful proxy. Even if the bound were tight, it would not prove that the residual—rather than the added stochasticity—is the active mechanism. Without an ablation that replaces R with a constant or a y-independent proxy, the empirical claim that the method is measurement-aware is unsupported.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MACS, a test-time sampling strategy for consistency models in inverse problems. The core idea is to replace the variance-compensation term of the aDDIM sampler with a measurement residual γ||y−A(x̂)||² inside the noise-injection coefficient, so that the sampler's stochasticity is adjusted according to how well the current estimate matches the observations. The method is evaluated on Fashion-MNIST and LSUN Bedroom for super-resolution, inpainting, linear deblurring, and nonlinear deblurring, using fixed pretrained CM backbones and conditioning networks. The reported results show consistent FID/KID improvements at 2 NFE relative to ODE baselines (Euler, Heun, DPM-Solver) and standard multistep consistency sampling, with comparable or slightly lower PSNR/SSIM.","tokens_in":14710,"tokens_out":4172,"duration_ms":37823,"significance":"If the mechanism works as claimed, MACS would be a simple, retraining-free improvement to CM-based inverse solvers, and the empirical gains on two datasets and four tasks are suggestive. The paper's strengths are its clear experimental setup, comparison at matched NFE, and use of standard metrics. However, the theoretical justification in Eqs. (37)–(41) is only an upper bound, not a derivation of equivalence with aDDIM, and the experiments do not include an ablation that isolates measurement-awareness from added stochasticity. These gaps leave the central causal claim—that the measurement residual, rather than increased noise, is the active ingredient—unsupported. The contribution is thus best viewed at present as a promising heuristic rather than a validated mechanism.","major_comments":[{"comment":"The derivation establishes only E||y−A(x̂)||² ≤ ||A||₂²||x−x̂||² + mσ_y², and Eq. (41) absorbs ||A||₂² and mσ_y² into a tunable γ. This is an inequality, not an equivalence; it does not show that the residual is proportional to the conditional variance that aDDIM compensates. For rank-deficient operators (inpainting, super-resolution), the residual can be far smaller than Var[x|x_t]. The sentence following Eq. (41) that the residual is 'fully consistent' with aDDIM is therefore not justified. The theoretical link is load-bearing for the 'measurement-aware' claim.","section":"§IV, Eqs. (37)–(41)"},{"comment":"No ablation separates the effect of the residual's dependence on y from the effect of increased stochasticity. For any γ>0, the update (Eq. 36) adds more noise than DDIM/ODE baselines. Since the authors themselves cite the perception–distortion tradeoff, and extra stochasticity alone is known to improve FID/KID, the observed gains could arise simply from larger noise. A control with R replaced by a constant or by a y-independent variance proxy is needed to attribute the improvements to measurement-awareness. Without it, the central mechanism remains untested.","section":"§V-C, Eq. (36), Tables I–III"},{"comment":"The guidance scale γ is chosen per task (0.15–0.8) with no sensitivity analysis, and the reported metrics appear to be point estimates without error bars or multiple seeds. For example, on LSUN SR the FID difference between MACS (40.13) and Multistep (40.85) is small; it is unclear if this exceeds run-to-run variability. Adding confidence intervals or a γ-sweep would strengthen the claim of consistent, robust gains.","section":"§V-B, Tables I–III"},{"comment":"The theoretical derivation in Eqs. (37)–(41) assumes a linear forward operator A. The nonlinear deblurring experiment in Table III is presented as an extension of the same principle, but the spectral norm bound and the expectation expansion do not apply to nonlinear maps. The nonlinear result should be described as heuristic/empirical, or a separate argument must be provided; as written, the paper extends the theoretical justification beyond its domain of validity.","section":"§V-A, Table III"}],"minor_comments":[{"comment":"The Multistep row is difficult to read due to missing spacing (e.g., '226.03 0.77140.90 5.67' appears to combine the NFE '2' with PSNR '26.03' and SSIM '0.771'). Please reformat for clarity.","section":"Table I"},{"comment":"The notation x_{var,t} is introduced as a scalar η||x−x̂||² but then appears multiplicatively with the vector ε̂. Please clarify whether it is a scalar coefficient or a vector, and how the product is computed.","section":"§II-D, Eq. (25)"},{"comment":"The claim that increasing NFE beyond 2 does not improve results is supported only by the residue curves in Fig. 4, not by metric tables at NFE=3 or 4. A small table or a sentence reporting metrics at higher NFE would make this observation reproducible.","section":"§V-C"}],"recommendation":"major_revision","confidential_remarks":"The empirical comparison is competently executed, but the central mechanism is not convincingly demonstrated. The missing ablation (measurement-dependent residual vs. constant noise) is the key risk: the method may work, but for a different reason than the one claimed. I would encourage the editor to request the ablation and a tightening of the theoretical claims before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is Eq. (36): MACS swaps the teacher-norm variance term in aDDIM for a measurement residual γ||y−A(x̂)||². That specific update is not in the cited papers, and it is not a re-coordinatization of an existing equation. The experiments are decent: two datasets, four inverse tasks, a frozen CM backbone, and consistent FID/KID gains over the tested baselines at two NFEs. That is a real, useful empirical contribution for anyone doing few-step CM-based restoration without retraining.\n\nThe soft spot is the one the stress-test flags, and it lands. There is no ablation where the residual term is replaced by a y-independent scalar of the same magnitude. Without that control, the reported gains could come entirely from added stochasticity—the perception–distortion tradeoff the authors themselves cite. So the central claim that the method is measurement-aware is not actually demonstrated by these tables. The theory does not rescue it: Eqs. (37)–(41) prove an upper bound, not equality, and then absorb ||A||² and mσ_y² into a free γ. That makes the connection to aDDIM unfalsifiable from the derivation. The same hand-wave is then extended to a nonlinear operator even though the derivation assumes linear A.\n\nOther issues are smaller but worth fixing: no error bars on any table, γ is chosen per task with no sensitivity analysis, and CM4IR—arguably the strongest existing zero-shot few-step CM inverse solver—is missing from the comparisons. Any of those alone would be minor; together they weaken the paper but do not sink the empirical finding.\n\nI want to be clear: the empirical gains in Tables I–III look real and consistent, and the method is simple enough to be a useful plug-in for other CM inverse pipelines. The mismatch is between the modest evidence and the strong mechanism claim. That is fixable with one well-designed ablation, sensitivity curves, and a rewritten theory section that states the heuristic status honestly.\n\nThis paper deserves a serious referee. It is competent, reproducible in spirit, and the update rule is novel enough to matter to the subfield. I would send it to review, with the expectation of moderate revision rather than acceptance as-is.","headline":"A useful, empirically solid sampler variant, but the paper never isolates whether the measurement-residual term—not just extra stochasticity—is what improves the metrics.","tokens_in":15188,"tokens_out":1245,"would_cite":true,"duration_ms":13330,"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":"This paper claims that a measurement-residual term can replace the variance-compensation term in a consistency-model sampler, yielding a test-time-only guidance mechanism that improves inverse-problem reconstructions in two steps without re","keywords":["inverse problems","consistency models","few-step sampling","measurement consistency","image restoration","test-time adaptation","super-resolution","deblurring"],"falsifier":"Run a calibration experiment on a fixed linear operator A: set gamma from the derived bound (||A||_2^2 plus the measurement-noise contribution) and compare reconstruction quality against the paper's tuned gamma. If the tuned gamma deviates by more than a small factor from that value, or if the optimal gamma varies strongly across images or noise levels, the substitution is not fully consistent with the aDDIM variance-compensation mechanism.","tokens_in":14240,"feed_emoji":"🎯","tokens_out":4549,"duration_ms":51495,"temperature":0.7,"pith_summary":"The paper tries to establish that the measurement residual—how far a current estimate is from matching the observed data through the degradation operator—can stand in for the variance-compensation term in a consistency-model sampler. It introduces MACS, which scales the injected noise at each sampling step by this residual: larger residuals allow more exploration, smaller residuals refine. This is a purely test-time change that leaves the consistency model and any conditioning network untouched, and the authors show consistent perceptual and pixel-level gains across super-resolution, inpainting, and deblurring on two datasets using only two sampling steps. The derivation treats the residual as an upper bound on the variance term and absorbs the constant factors into a single tunable scalar gamma.","feed_headline":"Measurement-aware sampling beats fast baselines in 2 steps","feed_subtitle":"Residual-driven noise injection enforces data fidelity in consistency models without retraining, across super-resolution, inpainting, and de","key_machinery":"The MACS update step is the central mechanism: given the model's clean prediction x_hat and residual direction eps_hat = x_t - x_hat, the next latent is x_s = x_hat + sqrt(rho + ((1 - sqrt(rho))/||eps_hat||)^2 * gamma*||y - A(x_hat)||^2) * eps_hat, where rho is the noise-level ratio and gamma is a hyperparameter. The residual term replaces the variance-compensation term from the adjusted DDIM sampler, modulating the injection scale as a function of measurement fidelity while keeping the update direction aligned with the model's own estimate. This single algebraic change is what allows measurement awareness to be injected into the sampling trajectory without retraining.","core_discovery":"The central claim is that the stochasticity of a consistency-based sampler can be regulated by the measurement residual instead of an unobserved ground-truth difference. Concretely, MACS replaces the variance term eta||x - x_hat||^2 in the adjusted DDIM update with gamma||y - A(x_hat)||^2, where A is the degradation operator. The paper argues that for linear operators this residual is bounded by a scaled version of the variance term plus a measurement-noise floor, and that with a suitable gamma the two are functionally interchangeable. This substitution produces an update that injects more noise when the estimate is far from the measurements and less when it is close, steering the few-step t","pith_inferences":["The single scalar gamma bundles the operator norm, measurement noise variance, and the unknown conditional variance; a per-operator or per-step calibration of gamma based on ||A||_2 and sigma_y could turn the upper bound into a tighter equality and may improve robustness.","Because the residual term uses the forward operator directly, MACS could be adapted to non-Gaussian measurement noise by replacing the L2 residual with a log-likelihood term, a natural extension the paper does not explore.","The method's success on nonlinear deblurring suggests the residual behaves more like a general data-fidelity scheduler than a strict variance estimator; a theory that does not rely on linearity would be a valuable next step.","A concrete stress test would be to hold gamma fixed while varying the conditioning of A or the noise level; if the optimal gamma shifts widely, the claimed interchangeability with aDDIM's variance term is only approximate."],"forward_implications":["If the claim holds, any consistency-model inverse solver can adopt MACS as a plug-in sampling schedule with no retraining, since it only modifies the update step.","Two-step reconstruction becomes practical for real-time inverse imaging, cutting the cost by an order of magnitude relative to diffusion solvers that need hundreds of function evaluations.","The residual-guided noise injection tends to shift the distortion–perception trade-off toward better perceptual quality while keeping pixel-level metrics competitive, as the authors observe.","The same residual-guidance principle could extend to other few-step samplers that share the DDIM-style update structure, not just consistency models."],"fun_headline_variants":["Measurement-aware sampler cuts steps for inverse problems","Few-step inverse imaging with misfit-guided noise","MACS: fast inverse problems by measuring the fit","Consistency sampling that checks the data","Inverse problems in a few steps, guided by measurements"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The derivation only shows that the measurement residual is bounded above by the variance term times the operator norm plus a noise floor, and then treats that bound as justification for replacing the variance term with the residual; this requires a single scalar gamma to absorb the operator norm and noise floor across all images, noise levels, and even nonlinear operators.","fun_headline_variants_meta":{"raw":{"variants":["Measurement-aware sampler cuts steps for inverse problems","Few-step inverse imaging with misfit-guided noise","MACS: fast inverse problems by measuring the fit","Consistency sampling that checks the data","Inverse problems in a few steps, guided by measurements"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00074,"raw_usage":{"total_tokens":3127,"prompt_tokens":719,"completion_tokens":2408,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":2347}},"tokens_in":463,"tokens_out":2408,"duration_ms":15486,"temperature":1.0,"reasoning_tokens":2347,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T12:41:49.341204+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a calibration experiment on a fixed linear operator A: set gamma from the derived bound (||A||_2^2 plus the measurement-noise contribution) and compare reconstruction quality against the paper's tuned gamma. If the tuned gamma deviates by more than a small factor from that value, or if the optimal gamma varies strongly across images or noise levels, the substitution is not fully consistent with the aDDIM variance-compensation mechanism.","supporting_citations":[],"review_version":1}