{"id":"f73ec187-dd0c-46b0-bd79-7792b1adec9e","arxiv_id":"2506.16961","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"ResFlow models HQ-to-LQ degradation as a deterministic augmented flow and inverts it via velocity matching, reporting state-of-the-art restoration in fewer than four sampling steps.","lead":"This paper proposes ResFlow, an image restoration method that treats degradation as a deterministic path modeled with continuous normalizing flows and adds an auxiliary variable to handle the ill-posed mapping from degraded to clean images. The method claims state-of-the-art restoration quality in fewer than four sampling steps, which would make fast, high-quality restoration practical.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Auxiliary variable is independent of the image pair, so flow matching cannot learn the claimed disambiguating coupling; entropy-preserving schedule also fails its own derivation.","rationale":"The reader's weakest assumption is exactly the auxiliary-variable coupling: the claim that the velocity network learns a deterministic coupling between y1 and x0 that disambiguates the ill-posed inverse map. I agree, and the problem is even sharper than a missing proof: under the stated training distribution, y_t is independent of (x0,x1), so the flow-matching objective's optimal velocity is provably independent of y_t. The loss therefore provides no learning signal that could create the claimed disambiguation. The entropy-preserving schedule is also internally inconsistent with the paper's own derivation, and the endpoint condition y_0=0 is violated by Eq. (8). These are correctness concerns about the theoretical framing, not about the empirical results per se. The reported gains might still be real and could come from multi-step iterative restoration with a learned residual velocity, rather than from the auxiliary disambiguation mechanism. For that reason the verdict should remain conditional rather than reject: the paper can be salvaged by reframing the method, fixing the schedule derivation, and adding the missing comparisons (InDI, Cold Diffusion, and recent diffusion restoration baselines). The proposed constant-y1 ablation is a direct, inexpensive test of whether the auxiliary variable is load-bearing at all.","tokens_in":22767,"tokens_out":10374,"duration_ms":107612,"concrete_test":"Re-derive the minimizer of Eq. (9) under the paper's independence assumptions: compute E[x1-x0 | x_t, y_t] and verify that y_t cancels, leaving a velocity independent of the auxiliary variable. Then run the authors' training protocol on one benchmark (e.g., Snow100K) with y1 fixed to 0 for all training and inference samples, leaving all other hyperparameters identical. If the resulting PSNR/SSIM matches the reported ResFlow numbers within noise, the auxiliary variable contributes no disambiguating information and the central claim fails. If the random-y1 model is better, measure per-sample output variance for a fixed x1 to show that different y1 produce different valid restorations rather than just learned noise sensitivity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central mechanism is that the auxiliary process y_t 'disambiguates the velocity' and encodes the information lost during degradation. This cannot be learned from the stated objective. In Eq. (9), y1 is sampled as N(0,I) independently of (x0,x1), and y0=0; hence y_t = σ_y^t y1 is independent of (x0,x1). The target x-velocity is ẋ_t = x1 - x0. The minimizer of Eq. (9) is the conditional expectation E[ẋ_t | x_t, y_t]; because y_t is independent of (x0,x1), this equals E[x1 - x0 | x_t], which does not depend on y_t. Thus at the optimum the learned velocity ignores the auxiliary variable, and the ODE from a fixed LQ image x1 is the same for every sampled y1. No deterministic coupling between y1 and x0 is induced by the training loss; any dependence of the trained network on y_t is an optimization artifact, not a learned disambiguation of the ill-posed inverse map. A second internal inconsistency compounds this: the supplementary derivation of the entropy-preserving schedule (Eqs. 33-34) yields σ_y^t = β/(1-t), singular at t=1; the regularized schedule actually used, σ_y^t = β/(1-t+β), does not satisfy the constant-entropy condition (at β=10, ln(1-t)+ln σ_y^t is ln(10/11) at t=0 and ln(5/10.5) at t=0.5). It also gives σ_y^0 = β/(1+β) ≠ 0, contradicting the stated endpoint y_0=0 in Eq. (3). The theoretical centerpiece of the paper is therefore not supported by its own equations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ResFlow, a flow-matching framework for image restoration that models the degradation from high-quality (HQ) to low-quality (LQ) images as a deterministic continuous normalizing flow. To handle the ill-posedness of restoration, the state is augmented with an auxiliary process y_t, claimed to disambiguate the velocity field. The method uses a linear interpolation schedule for x_t and a proposed \"entropy-preserving\" schedule for y_t, trains a velocity network by matching the conditional flow-matching objective (Eq. 9), and restores images by solving the ODE backward from t=1 to t=0 with a sampled y_1. Experiments on desnowing, deraining, dehazing, denoising, defocus deblurring, and JPEG artifact removal report state-of-the-art results with fewer than four sampling steps.","tokens_in":23180,"tokens_out":11670,"duration_ms":115831,"significance":"The empirical contribution is substantial: ResFlow consistently outperforms strong baselines across multiple restoration tasks and datasets, with up to 2.24 dB PSNR gains over Restormer on SIDD and strong results with only four sampling steps. The paper also provides thorough ablations, implementation details, and supplementary visualizations, which support reproducibility. However, the theoretical centerpiece—the entropy-preserving schedule and the claim that the auxiliary variable disambiguates the uncertain inverse map—is contradicted by the paper's own equations. Given that these claims are central to the paper's novelty narrative, the significance of the contribution depends on whether they can be corrected or demoted in revision.","major_comments":[{"comment":"The claim that the trained velocity network \"induces a deterministic coupling\" between y_1 and the HQ image x_0 is not supported by the training objective. In Eq. (9), y_1 is sampled as N(0,I) independently of (x_0,x_1), and y_0 = 0, so y_t = σ^y_t y_1 is independent of (x_0,x_1) for all t. The target x-velocity is \\dot x_t = x_1 - x_0. Minimizing the flow-matching loss (9) for the x-component yields the conditional expectation E[x_1 - x_0 | x_t, y_t], which, by the independence of y_t, reduces to E[x_1 - x_0 | x_t]. Thus, at the optimum, the learned x-velocity is independent of y_t, and the ODE (6) starting from a fixed LQ image x_1 produces the same x_0 for every sampled y_1. The different outputs for different auxiliary variables shown in Fig. 11 are therefore an optimization artifact rather than a learned disambiguation of the ill-posed inverse. This directly undermines the central mechanism claimed in Sec. 3.2 and in the abstract.","section":"Sec. 3.2, Eqs. (9)-(10)"},{"comment":"The \"entropy-preserving\" degradation schedule is internally inconsistent. The derivation in the supplementary sets H(z_t)=H(z_s), leading to Eq. (33), whose solution is σ^y_t = β/(1-t), singular at t=1. The schedule actually used, σ^y_t = β/(1-t+β), does not satisfy Eq. (33): with β=10, ln(1-t)+ln σ^y_t equals ln(10/11) at t=0 but ln(5/10.5) at t=0.5, so the entropy of z_t is not constant. In addition, σ^y_0 = β/(1+β) ≠ 0, violating the boundary condition y_0 = 0 stated in Eq. (3) and the endpoint requirement σ_0 = 0 in Eq. (5). The abstract and introduction claim that ResFlow \"adopts entropy-preserving flow paths,\" but this property is contradicted by the paper's own equations. The schedule may still be a useful heuristic, but it cannot be justified by the constant-entropy derivation as written.","section":"Sec. 3.2, Eq. (8) and Supplementary Sec. 7, Eqs. (33)-(34)"}],"minor_comments":[{"comment":"The text lists \"draining\" as a task name; this should read \"deraining.\"","section":"Sec. 4.1"},{"comment":"The sentence \"Optimizing Eq. (10) is efficient\" appears to reference the wrong equation; the optimization objective is Eq. (9), while Eq. (10) defines the target velocity components.","section":"Sec. 3.3"},{"comment":"The caption of Fig. 11 is identical to that of Fig. 10 and describes defocus deblurring, but the surrounding text and figure content discuss the effect of auxiliary variables on restored outputs. The caption should be corrected.","section":"Supplementary, Fig. 11 caption"},{"comment":"The description of the Snow100K dataset contains the word \"nozomura,\" which appears to be a typographical error; also, \"Outdoor-Rain [49] contains 8100 images for training and 900 images for testing;.\" has an extra semicolon.","section":"Supplementary, Sec. 8"}],"recommendation":"major_revision","confidential_remarks":"The empirical results are strong and the method may be a useful contribution even without the theoretical framing. However, the two major concerns are both load-bearing: the auxiliary variable's disambiguation property is not a consequence of the stated training objective, and the entropy-preserving schedule does not preserve entropy. These cannot be fixed by small edits; the authors must either correct the theory or substantially reframe the claims as empirical heuristics. The editor may want to verify that the revised version does not merely soften the language while keeping the incorrect derivations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid applied flow-matching paper with genuinely strong empirical numbers, but the theoretical story is broken in two places that matter. The auxiliary variable cannot learn the claimed disambiguation from the stated objective, and the entropy-preserving schedule fails its own derivation. I'd still send it to review, but the authors need to fix the claims.\n\nWhat's new and good: ResFlow trains a deterministic ODE from HQ to LQ via simple linear interpolation, matches the velocity as in flow matching, and restores images in as few as four steps. The empirical section is broad and consistent: PSNR/SSIM/LPIPS gains over strong baselines across desnowing, deraining, dehazing, denoising, defocus deblurring, and JPEG artifact removal, including real-world datasets. The idea of adding an auxiliary variable to make the ill-posed inverse well-posed is interesting even if the implementation doesn't do what it claims. They also discuss InDI and Cold Diffusion explicitly, which is more than many such papers do.\n\nNow the soft spots, and they are not minor. The stress-test concern is real. In Eq. (9), y1 is sampled independently of (x0,x1), so y_t is independent of the image pair. The minimizer of the flow-matching loss is the conditional expectation E[x1-x0 | x_t], which does not depend on y_t. At the optimum, the learned velocity ignores the auxiliary variable entirely; any observed dependence is an optimization artifact, not a learned coupling. That breaks the central claim that y_t disambiguates the uncertainty scope. Fig. 11 is anecdotal, not evidence of a statistical coupling.\n\nSecond, the entropy-preserving derivation in the supplementary is self-contradictory. The exact solution is sigma_y = beta/(1-t), which is singular at t=1, and the regularized version beta/(1-t+beta) does not keep ln(1-t)+ln sigma_y constant. It also gives sigma_y(0) != 0, contradicting y_0=0 in Eq. (3). So the name \"entropy-preserving\" is not supported by the paper's own math. The schedule might still be a fine practical choice, but it is not what they call it.\n\nMissing comparisons also hurt: InDI and Cold Diffusion are the closest deterministic-inversion baselines and are not in the tables. No code is provided, so the numbers are not independently reproducible.\n\nBottom line: this paper deserves a serious referee because the empirical results are potentially important and the method is simple enough to be reproduced quickly. A referee should push hard on the theory, and the authors should either fix the derivation or drop the entropy-preserving and disambiguation claims. The experimental work can stand on its own. For readers focused on efficient restoration, the results are valuable; for readers who care about the conceptual story, the paper currently disappoints.","headline":"Strong empirical flow-matching restoration with four-step sampling, but the auxiliary-variable and entropy-preservation claims do not survive the paper's own equations.","tokens_in":23672,"tokens_out":4179,"would_cite":false,"duration_ms":41204,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Image restoration can be done by reversing a learned deterministic degradation flow, cutting sampling to four steps or fewer.","keywords":["image restoration","continuous normalizing flow","flow matching","auxiliary variable","degradation modeling","few-step inference","ill-posed inverse problems"],"falsifier":"For a fixed degraded image, sample two auxiliary vectors $y_1 \\sim N(0,I)$ and run the backward ODE; if the two restored outputs do not differ in the ambiguous regions, or if one is clearly hallucinated, the disambiguation claim fails. Alternatively, compute the differential entropy of $z_t$ along the schedule (8); if it changes with $t$, the entropy-preserving property is violated.","tokens_in":22589,"feed_emoji":"🖼️","tokens_out":7482,"duration_ms":68626,"temperature":0.7,"pith_summary":"This paper proposes ResFlow, a framework that treats image degradation—from clean high-quality image to corrupted low-quality image—as a deterministic path rather than a stochastic diffusion process. Because ordinary deterministic flows preserve information while real degradation destroys it, ResFlow augments the state with an auxiliary variable that is meant to encode what is lost, making the path invertible. The augmented flow is learned by matching the velocity field of a simple linear interpolation between HQ and LQ images, and restoration is done by solving the learned ODE backward from the degraded image in fewer than four steps. If the central claim holds, image restoration becomes much cheaper and faster than diffusion-based methods while matching or exceeding their quality across tasks like denoising, dehazing, deraining, and deblurring.","feed_headline":"Four steps or fewer: reversing a learned flow restores images","feed_subtitle":"ResFlow swaps stochastic diffusion for a reversible deterministic path, cutting sampling cost while keeping quality.","key_machinery":"The central object is the augmented continuous normalizing flow on the concatenated state $z_t = [x_t; y_t]$, governed by the ODE $\\partial z_t/\\partial t = v_\\theta(x_t,y_t,t)$. The data component $x_t$ interpolates linearly between HQ and LQ, while the auxiliary component $y_t$ evolves with the schedule $\\sigma^y_t = \\beta(1-t+\\beta)^{-1}$; together they define the ground-truth velocity $\\dot{z}_t$ used in the flow-matching loss. The auxiliary variable is the mechanism that carries the argument: conceptually it encodes the information lost in degradation, allowing the learned velocity network to pick one HQ image among the many consistent with a given LQ image. Training never simulates the ODE; instead the velocity field is regressed pointwise against the analytic derivative of the interpolation, and inference integrates the learned field backward with Euler steps.","core_discovery":"ResFlow's central claim is that image restoration can be cast as the reverse of a deterministic continuous normalizing flow that is augmented with an auxiliary random variable. The state space is $z_t = [x_t; y_t]$, where $x_t$ moves from the HQ image $x_0$ to the LQ image $x_1$ by linear interpolation $x_t = (1-t)x_0 + t x_1$, and $y_t$ follows the schedule $\\sigma^y_t = \\beta/(1-t+\\beta)$ with $y_0 = 0$ and $y_1$ a standard Gaussian. The velocity field of this augmented flow is learned by minimizing the flow-matching loss $\\mathbb{E}\\int_0^1 \\lambda(t) \\|v_\\theta(x_t,y_t,t) - \\dot{z}_t\\|^2 dt$, with a loss weight that emphasizes times near $t=1$. At inference, the corrupted image $x_1$ is taken as the starting point, an auxiliary sample $y_1\\sim N(0,I)$ is drawn, and the ODE is integrated backward to $t=0$; the auxiliary variable is what the paper argues disambiguates which of several plausible clean images produced the observed degradation, and the schedule is described as entropy-preserving so that the reversal is a well-defined reversible process. The paper reports state-of-the-art PSNR/SSIM/LPIPS on several benchmarks with four or fewer sampling steps.","pith_inferences":["If the deterministic coupling between $y_1$ and $x_0$ is as clean as claimed, the same framework could be adapted for conditional image generation, where the auxiliary variable plays the role of a latent code that selects among plausible restorations; a direct test would be comparing the diversity of outputs on a known one-to-many degradation model.","The paper's own derivation in the supplementary material shows that exact entropy preservation would require $\\sigma^y_t = \\beta/(1-t)$, which is singular at $t=1$; the implemented schedule $\\beta/(1-t+\\beta)$ is a regularized approximation, so the entropy-preserving property should be understood approximately, and its empirical role can be tested by swapping in the singular schedule with a small ","The linear interpolation for $x_t$ ties ResFlow to rectified-flow-style straight paths, which suggests that the four-step budget could be reduced to one step via distillation, a direction the paper does not explore."],"forward_implications":["Restoration can be completed in fewer than four sampling steps, making inference far cheaper than diffusion models that typically need dozens or hundreds of steps.","The same trained velocity field and inference scheme apply across degradation types (snow, rain, haze, noise, blur, JPEG artifacts) without task-specific modifications, since the degradation path is defined by the data pair rather than a hand-designed operator.","The auxiliary variable at $t=1$ can be sampled as a Gaussian, so the method inherently offers a way to produce multiple plausible restored images for one degraded input.","Because training only matches the velocity of a known interpolation, it avoids simulating the ODE and therefore has low per-iteration cost, which the paper argues makes it practical for real-world deployment.","The loss weighting that emphasizes $t$ near 1 is what the paper credits for preserving sharp details that single-step methods average away, so one-step baselines should be dominated on strong degradations."],"supporting_citations":[{"why":"Supplies the flow-matching training objective and the coupling-based pointwise velocity regression that ResFlow adopts.","marker":"[54]"},{"why":"Defines the continuous normalizing flow / ODE formulation that the degradation path and its reversal rely on.","marker":"[15]"},{"why":"Provides the rectified-flow-style linear interpolation and the convex transport cost guarantee that motivates the straight degradation paths.","marker":"[58]"},{"why":"Contributes the U-Net backbone architecture and serves as the stochastic diffusion baseline that ResFlow contrasts against.","marker":"[31]"},{"why":"The prior inversion-by-direct-iteration approach that ResFlow builds on and claims to improve, and the alternative parameterization criticized for high discretization error near $t=0$.","marker":"[20]"},{"why":"Provides the lemma that mutual information is invariant under invertible maps, used in the proof of Proposition 1 that ODE flows preserve mutual information.","marker":"[43]"},{"why":"States the data processing inequality that grounds the argument that real degradation loses information and thus needs the auxiliary variable.","marker":"[11]"}],"fun_headline_variants":["ResFlow: four-step image restoration with a reversible deterministic flow","Fewer than four steps: invert a learned flow to fix images","Deterministic flow reversal beats stochastic diffusion for restoration","Reversible flow: image restoration in under four sampling steps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that the learned velocity network, trained to match the linear-interpolation velocities, produces a deterministic coupling between the random auxiliary variable and the clean image that genuinely disambiguates the ill-posed map—so that every sampled auxiliary input yields a valid, sharp restoration rather than an inconsistent or averaged one.","fun_headline_variants_meta":{"raw":{"variants":["ResFlow: four-step image restoration with a reversible deterministic flow","Fewer than four steps: invert a learned flow to fix images","Deterministic flow reversal beats stochastic diffusion for restoration","Reversible flow: image restoration in under four sampling steps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001049,"raw_usage":{"total_tokens":4435,"prompt_tokens":1000,"completion_tokens":3435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":3375}},"tokens_in":616,"tokens_out":3435,"duration_ms":25557,"temperature":1.0,"reasoning_tokens":3375,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:14:59.976264+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a fixed degraded image, sample two auxiliary vectors $y_1 \\sim N(0,I)$ and run the backward ODE; if the two restored outputs do not differ in the ambiguous regions, or if one is clearly hallucinated, the disambiguation claim fails. Alternatively, compute the differential entropy of $z_t$ along the schedule (8); if it changes with $t$, the entropy-preserving property is violated.","supporting_citations":[],"review_version":1}