{"id":"3f322315-0b74-49e9-b009-6d2789f897d8","arxiv_id":"2608.09405","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MeanSR applies LR-conditioned average-velocity learning from MeanFlow to one-step super-resolution and reports improved no-reference perceptual scores over CTMSR with lower compute.","lead":"MeanSR is a one-step super-resolution method that learns an average velocity field mapping noisy low-resolution latents to high-resolution images in a single network call. The paper reports better CLIPIQA, MUSIQ, and MANIQA scores than the prior one-step method CTMSR while cutting FLOPs and runtime, which could make high-quality super-resolution practical on edge devices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Superiority over CTMSR rests entirely on no-reference metrics from a single training run, with second-stage length and sampling tuned on the same test set; a rerun or human study could overturn it.","rationale":"The reader's weakest assumption correctly identifies the evaluation protocol as the load-bearing point. The paper's own Appendix A concedes a single training run, and Tables 4 and 5 show the final configuration was selected on ImageNet-Test, the same set used for the headline numbers. The 30k collapse in Table 4 demonstrates that the method is highly sensitive to second-stage training length, so the reported 10k result could easily be a favorable fluctuation rather than a stable property. No-reference metrics alone cannot establish perceptual superiority, especially when LPIPS on the synthetic benchmark is worse than CTMSR (0.228 vs 0.197); the paper explains this as better realism, but that interpretation requires human validation. I also examined the mathematical objective: Eq. 5 with the JVP-based target has a known solution family v + C/(t-r) for the MeanFlow self-consistency, but the additional L1/VGG losses and DTM anchor the endpoint, so the theoretical derivation is not the primary concern. The decisive question is whether the empirical superiority survives independent reruns and human evaluation, which is exactly what the reader's conditional verdict requires.","tokens_in":13340,"tokens_out":10171,"duration_ms":115152,"concrete_test":"Obtain the authors' code and training configuration; retrain MeanSR and CTMSR under identical degradation, resolution, and evaluation pipelines with at least 5 independent seeds; report mean and standard deviation of CLIPIQA, MUSIQ, MANIQA, and LPIPS on a held-out ImageNet validation split that was not used for early stopping. Then run a forced-choice human preference study on 100 randomly sampled RealSet65 patches, comparing MeanSR vs CTMSR. If the reported gaps (e.g., RealSR MANIQA +0.09, RealSet65 MANIQA +0.22, CLIPIQA +0.04 to +0.07) are not consistently reproduced or human preference is not significantly above chance, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: MeanSR outperforms CTMSR on CLIPIQA, MUSIQ, and MANIQA while cutting FLOPs and latency. The weakest link is not the flow-matching derivation but the evaluation protocol. Appendix A explicitly states that all quantitative results come from a single training run. Table 4 shows the reported results are a knife-edge early-stopping point: 10k second-stage iterations gives the headline numbers, 20k degrades slightly, and 30k collapses catastrophically (CLIPIQA 0.294, MANIQA 0.2822). That stopping point, and the SATS choices in Table 5, were selected by inspecting ImageNet-Test, the same benchmark used for the final reported numbers; there is no separate validation set. The remaining evidence is no-reference metrics (CLIPIQA, MUSIQ, MANIQA) with no human evaluation, despite the paper claiming 'perceptual' superiority and 'fewer perceptual artifacts.' If these metrics are noisy or weakly correlated with human perception, or if the reported gaps are within run-to-run variance, the claim of superiority over CTMSR is unsupported. The unexplained FLOPs discrepancy (46 G vs 305 G with only 1.3x parameter difference) further suggests the efficiency comparison may not be apples-to-apples.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MeanSR, a one-step perceptual super-resolution method that learns an LR-conditioned average velocity field using a MeanFlow-style formulation, combines it with a reformulated Distribution Trajectory Matching (DTM) objective, and introduces a Stage-Aware Temporal Sampling (SATS) strategy. The method is trained in two stages: first to estimate the average velocity from degraded/noisy latents to HR latents, and second to align the generated trajectory with the target HR trajectory. The authors evaluate on synthetic ImageNet-Test and real-world RealSR/RealSet65, reporting better CLIPIQA, MUSIQ, and MANIQA scores than CTMSR and other one-step methods, with substantially lower FLOPs and inference latency.","tokens_in":13734,"tokens_out":6455,"duration_ms":68172,"significance":"If the empirical claims hold, MeanSR would be a valuable contribution: a distillation-free one-step SR method that explicitly models restoration dynamics and achieves state-of-the-art perceptual quality with a 6x FLOPs reduction relative to CTMSR. The conceptual shift from consistency training to average-velocity trajectory learning is interesting, and the two-stage design with stage-specific temporal sampling is a reasonable idea. However, the current evidence is weakened by the absence of variance information, the use of the test set for hyperparameter selection, contradictory perceptual metrics (worse LPIPS), missing fidelity metrics, and an incompletely explained efficiency comparison. The central mathematical derivation in Eq. (4) is correct, and the paper clearly describes the method, but the empirical claims are not yet fully supported.","major_comments":[{"comment":"The MeanSR training target is self-referential: u_tgt = v - (t-r) * d/dt u_theta uses the network's own Jacobian (via the JVP in Algorithm 1). This makes the loss a fixed-point equation rather than a regression to an external target. For any function satisfying the ODE u = v - (t-r)u' along the sampled path, the loss vanishes, so without a boundary condition (e.g., enforcing u(r) = v(r)) the objective may have many degenerate solutions. While the additional l1 and VGG losses on x_hat_0 provide some grounding, the interpretation of loss_mf as matching the true average velocity is not established. The authors should either provide a convergence/stability argument for this fixed-point objective or replace the target with a network-independent estimate, such as the Monte Carlo average of the instantaneous velocity.","section":"Section 3, Eq. (5) and Algorithm 1"},{"comment":"All quantitative results are reported from a single training run, and the second-stage fine-tuning length is selected by inspecting ImageNet-Test, the same benchmark used for the final tables. Table 4 shows a knife-edge result: 10k iterations give the headline scores, 20k slightly degrade, and 30k collapse catastrophically (LPIPS 0.689, CLIPIQA 0.294, MANIQA 0.2822). Without repeated runs or a proper validation split, the reported superiority over CTMSR could easily be within run-to-run variance or an artifact of early stopping on the test set. The authors should rerun with multiple seeds and report mean and standard deviation, and should hold out a validation set for hyperparameter selection.","section":"Appendix A, Table 4"},{"comment":"The paper claims 'perceptual' superiority over CTMSR but reports worse LPIPS (0.228 vs 0.197), and it omits reference-based PSNR/SSIM entirely despite evaluating on a synthetic benchmark with ground-truth HR images. Since LPIPS is itself a widely used perceptual metric, reporting only CLIPIQA/MUSIQ/MANIQA is selective. The authors should include PSNR/SSIM (or other fidelity metrics) in Table 2 and discuss the LPIPS trade-off explicitly; ideally, they should also include a human reader study to substantiate the claim of 'fewer perceptual artifacts'.","section":"Table 2 and Section 4, 'Evaluation Metrics'"},{"comment":"The efficiency comparison is not sufficiently explained to be convincing. According to Table 1, MeanSR has about 1.3x fewer parameters than CTMSR (131.3 vs 171.5) yet reports 6.6x fewer FLOPs (46.18 vs 305.50 G). Such a large discrepancy cannot be explained by a modest parameter reduction alone and suggests differences in latent resolution, output channels, or architecture choices between the models. The authors need to describe the architecture of MeanSR, the latent resolution used for FLOPs measurement, and the exact hardware/protocol for the runtime comparison; without this, the 'substantial FLOPs reduction' claim may not be comparing like with like.","section":"Table 1 and Section 4.1"},{"comment":"The second-stage pseudo-target tilde z0 = sg(hat z0 - Delta) with Delta = hat z'_0 - z'_0 is computed through a delayed target network, which makes the alignment objective a self-correcting fixed-point procedure. The paper does not analyze the convergence, bias, or stability of this target, nor does it compare with alternative trajectory-alignment objectives beyond the two variants in Table 3. This matters because the authors themselves observe a catastrophic degradation after 30k second-stage iterations (Table 4), suggesting the objective can become unstable. A formal or empirical analysis of the target's behavior would strengthen the method's credibility.","section":"Section 3, Stage 2, Eq. (8)"}],"minor_comments":[{"comment":"The MeanSR row of Table 1 is garbled ('0.65611 131.346.181 26.392'), mixing metric values with parameter/FLOP/runtime numbers; fix the table formatting so that each column is readable.","section":"Table 1"},{"comment":"The phrase 'We alos analyze' contains a typo; it should be 'We also analyze'.","section":"Section 4.2"},{"comment":"The loss is written as 'metric(u - stopgrad(u_tgt))' but the metric is not specified; if it is the L2 norm used in Eq. (5), state this explicitly.","section":"Algorithm 1"},{"comment":"The abbreviations 'DTM-u' and 'DTM-x0' are not defined in the caption or the text; add definitions to clarify what is being ablated.","section":"Table 3 caption"},{"comment":"The caption states the x-axis is CLIPIQA and the y-axis is runtime, but the figure is not shown clearly in the text; if the axes are correct, ensure the figure is legible and includes labeled axes.","section":"Figure 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a plausible and interesting method, but the empirical case is currently under-supported. The self-referential velocity target and the knife-edge second-stage training behavior deserve careful scrutiny in revision. I recommend asking the authors to provide repeated runs with variance, a separate validation set for hyperparameter selection, a direct FLOPs/runtime comparison under matched conditions, and a discussion of the LPIPS trade-off. If the authors cannot address these points within a revision, the central claim of superiority over CTMSR should not be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The method is a reasonable extension of MeanFlow and CTMSR to one-step SR, and the paper is honest about its single-run evaluation. The reported superiority over CTMSR is plausible but not established.\n\nWhat's actually new: the LR-conditioned average velocity field, the image-space DTM variant, and the stage-aware temporal sampling. The equations are coherent and match the cited sources (MeanFlow's Eq. 4, CTMSR's DTM). The two-stage scheme is sensible, and the ablations give real insight—image-space DTM beats velocity-space DTM, and the logit-normal vs uniform split across stages is a nice observation. The paper credits its predecessors clearly; the novelty is incremental, but it is real.\n\nThe soft spot is the evaluation, not the math. Appendix A states all results come from a single training run, with no variance or significance tests. Table 4 shows the headline numbers sit at a knife-edge 10k iterations: 20k degrades slightly, 30k collapses (CLIPIQA 0.294). Tables 4 and 5 were used to select the temporal sampling and second-stage length on ImageNet-Test, the same benchmark reported as the final result—no separate validation set. The no-reference metrics (CLIPIQA, MUSIQ, MANIQA) are the entire basis for claiming superiority over CTMSR, and the paper notably shows worse LPIPS than CTMSR on the synthetic benchmark. The FLOPs discrepancy (46 G vs 305 G with only 1.3x parameter difference) is unexplained and may not be an apples-to-apples comparison. The self-referential training target (u_tgt uses the network's own Jacobian) is a fixed-point consistency objective, not an external ground truth—that is not fatal, but it does weaken the word 'prediction' in the abstract.\n\nThe paper deserves a serious referee. The idea is sensible, the writing is clear, and the authors did not hide their evaluation choices. But the empirical claim needs multiple runs, a held-out validation set, and ideally a human study or at least standard fidelity metrics. If the authors supply code and stronger statistics, this could be a solid applied contribution. I would not cite it yet for the SOTA claim, but I would follow the thread.","headline":"A plausible one-step SR method whose claimed win over CTMSR rests on a single run and no-reference metrics; the method itself is coherent and worth a careful look.","tokens_in":14172,"tokens_out":1578,"would_cite":false,"duration_ms":18456,"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":"A one-step super-resolution model that learns the average velocity of the LR-to-HR restoration path claims better perceptual quality than its closest one-step rival at a fraction of the compute.","keywords":["one-step super-resolution","perceptual super-resolution","average velocity field","MeanFlow","distribution trajectory matching","stage-aware temporal sampling","flow matching","real-world image super-resolution"],"falsifier":"Re-run both MeanSR and CTMSR with several random seeds and compare the distributions of CLIPIQA/MUSIQ/MANIQA on RealSR, RealSet65, and a held-out real-world set; if the inter-seed spread covers the reported gaps, the claimed advantage is not reproducible. Also run a blind human preference study between single-step outputs; if raters do not prefer MeanSR, the metric gains are not perceptual gains.","tokens_in":13172,"feed_emoji":"🖼️","tokens_out":7515,"duration_ms":64154,"temperature":0.7,"pith_summary":"MeanSR tries to make single-image super-resolution both perceptual and fast by learning, in a single network, the average velocity field that carries a degraded and noisy latent image to a credible high-resolution output. The authors argue that this explicit modeling of the restoration trajectory is what one-step methods like CTMSR lack, and they show that adding distribution trajectory matching on top of the learned field, with stage-specific temporal sampling, improves no-reference perceptual scores (CLIPIQA, MUSIQ, MANIQA) on synthetic and real-world benchmarks while cutting FLOPs by about six times and inference latency by roughly half compared with CTMSR. If the claims hold, MeanSR would be a state-of-the-art one-step perceptual SR method that does not need a pretrained diffusion teacher.","feed_headline":"One-step SR via average-velocity learning beats CTMSR at ~1/6 FLOPs","feed_subtitle":"MeanSR does single-step super-resolution by learning the LR-conditioned restoration field, cutting FLOPs ~6x and latency ~2x.","key_machinery":"The central object is the learned LR-conditioned average velocity field $u(z_t, c_{LR}, r, t)$, the time-averaged transport velocity between two latent states on a flow-matching path; it is the quantity that collapses the whole iterative denoising chain into one network evaluation. The paper inherits it from MeanFlow, which expresses the average velocity from an instantaneous velocity $v(z_\\tau, c_{LR}, \\tau) = \\epsilon - z_0$, and the training objective regresses the network output to a stop-gradient target that includes a Jacobian-vector-product term $\\frac{d}{dt}u_\\theta$, computed with automatic differentiation. This same field drives the second stage, where generated and ground-truth trajectories are perturbed in tandem and compared through a delayed target network, with the discrepancy used to build a pseudo-target for LPIPS supervision.","core_discovery":"MeanSR's central claim is that the finite-time transition from a low-resolution input to a high-resolution image can be captured explicitly as an LR-conditioned average velocity, so that a single step $z_{SR} = z_1 - u(z_1, c_{LR}, 0, 1)$ produces the SR latent directly from Gaussian noise. The paper derives this velocity from the MeanFlow identity $u(z_t, c_{LR}, r, t) = v(z_t, c_{LR}, t) - (t-r)\\frac{d}{dt}u(\\cdot)$, learns it with a stop-gradient regression, and then reuses the same field to align the generated trajectory with the ground-truth HR trajectory through Distribution Trajectory Matching in image space. With a logit-normal temporal distribution for velocity estimation and uniform sampling for DTM, the method claims state-of-the-art perceptual quality among one-step SR methods on RealSR, RealSet65, and ImageNet-Test, improving MANIQA over CTMSR by 19.9% and 50.0% on RealSR and RealSet65 respectively while using roughly 6x less FLOPs and 2x less runtime.","pith_inferences":["The same LR-conditioned average-velocity objective could transfer to other restoration tasks (denoising, deblurring, face enhancement) because it only requires defining a condition and a target distribution, though the paper does not test this.","Because the main comparison is entirely metric-based, a forced-choice human preference study on RealSR and RealSet65 outputs would be a direct check of whether the higher CLIPIQA/MUSIQ/MANIQA numbers translate into visible perceptual gains.","The stage-2 hyper-parameters (10k iterations, logit-normal vs uniform sampling) were selected on the same ImageNet-Test used for final numbers, so the reported gains could partly reflect benchmark overfitting; evaluating on a freshly degraded hold-out set would quantify this.","The large MANIQA improvements (19.9% on RealSR, 50.0% on RealSet65) may be inflated by metric sensitivity; pairwise perceptual tests with human raters would show whether the perceived gain is as large as the numbers suggest."],"forward_implications":["One-step SR with a single network evaluation: sampling is $z_{SR} = z_1 - u(z_1, c_{LR}, 0, 1)$, so no iterative denoising or numerical ODE solving is needed.","No dependence on pretrained diffusion teachers or distillation, unlike SinSR or OSEDiffR; training starts from synthetic LR-HR pairs and a VAE.","Substantially lower compute: about 6x fewer FLOPs and 2x lower latency than CTMSR while improving all three no-reference perceptual metrics on both real-world benchmarks.","Faster training: roughly 13x fewer iterations to reach the same CLIPIQA level as CTMSR on ImageNet.","Stage-aware temporal sampling is a general training recipe: logit-normal sampling for velocity estimation and uniform sampling for trajectory alignment."],"supporting_citations":[{"why":"Supplies the MeanFlow average-velocity identity that lets the model predict the whole trajectory in one step.","marker":"Geng et al. 2025"},{"why":"CTMSR, the main baseline; contributes Distribution Trajectory Matching, which MeanSR reformulates for average-velocity generation.","marker":"You et al. 2025"},{"why":"Provides the conditional flow-matching interpolation path $z_t = (1-t)z_0 + t\\epsilon$ used to define states.","marker":"Lipman et al. 2023"},{"why":"Provides the VAE encoder/decoder that maps images to and from the latent space where trajectory learning happens.","marker":"Rombach et al. 2022a"},{"why":"RealESRGAN degradation pipeline generates the synthetic LR-HR training pairs.","marker":"Wang et al. 2021"},{"why":"CLIPIQA, one of the three no-reference perceptual metrics used for the main comparisons.","marker":"Wang, Chan, and Loy 2023"},{"why":"MUSIQ, no-reference image quality metric used for the main comparisons.","marker":"Ke et al. 2021"},{"why":"MANIQA, no-reference perceptual metric on which MeanSR reports the largest gains.","marker":"Yang et al. 2022"},{"why":"ImageNet supplies training data and the 3,000-image ImageNet-Test constructed for synthetic evaluation.","marker":"Deng et al. 2009"}],"fun_headline_variants":["One-step SR via average-velocity learning: 6x less FLOPs","Average-velocity SR does one step, beats CTMSR on perceptual quality","MeanSR: one-step SR, 6x faster, beats CTMSR on perceptual metrics","Learns LR-to-HR velocity field for one-step perceptual SR","One-step SR: explicit average-velocity model cuts FLOPs 6x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire superiority claim rests on no-reference perceptual metrics (CLIPIQA, MUSIQ, MANIQA) and a single unrepeated training run, with temporal-sampling choices tuned on the same ImageNet-Test benchmark; if these metrics do not track human perception or the reported gaps are within run-to-run noise, the advantage over CTMSR collapses.","fun_headline_variants_meta":{"raw":{"variants":["One-step SR via average-velocity learning: 6x less FLOPs","Average-velocity SR does one step, beats CTMSR on perceptual quality","MeanSR: one-step SR, 6x faster, beats CTMSR on perceptual metrics","Learns LR-to-HR velocity field for one-step perceptual SR","One-step SR: explicit average-velocity model cuts FLOPs 6x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000701,"raw_usage":{"total_tokens":3175,"prompt_tokens":964,"completion_tokens":2211,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":2120}},"tokens_in":580,"tokens_out":2211,"duration_ms":15558,"temperature":1.0,"reasoning_tokens":2120,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:51:23.270217+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run both MeanSR and CTMSR with several random seeds and compare the distributions of CLIPIQA/MUSIQ/MANIQA on RealSR, RealSet65, and a held-out real-world set; if the inter-seed spread covers the reported gaps, the claimed advantage is not reproducible. Also run a blind human preference study between single-step outputs; if raters do not prefer MeanSR, the metric gains are not perceptual gains.","supporting_citations":[],"review_version":1}