{"id":"b3c549c9-c718-47d8-a9f9-9e10bbe751ab","arxiv_id":"2608.00978","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"QC-FM constructs flow-matching source samples by mapping projected data ranks to Gaussian quantiles, without solving a batch assignment, and reports FID improvements up to 12.9% over the independent-coupling baseline.","lead":"This paper introduces a cheaper way to build the noise-data pairings used to train flow-based image generators, using ranks along random directions instead of solving an assignment problem. On four standard image benchmarks it reports lower FID scores than the independent-coupling baseline and than optimal-transport coupling, pointing to a scalable alternative for training generative models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Frame resampling gap: per-frame variance elimination does not transfer to the U-blind averaged objective; empirical FID gains are the sole support.","rationale":"After reading the full text, the fixed-frame theory appears internally consistent: the proofs that slice variance vanishes and slice marginals are generated exactly rely only on comonotonicity and monotone m_t, and the complement-variance comparison in Theorem 2 Step 3 is sound. The hybrid construction and the calibrated k heuristic are clearly presented as empirical. The main soft spot is the transfer from per-frame guarantees to the deployed U-blind, resampled, hybrid training objective. The paper acknowledges this (Prop. A.8, Discussion), so it is not a hidden flaw, but it is load-bearing because the theoretical motivation for the method would be weakened if D_t cancels the variance reduction. The proposed fixed-frame experiment would directly measure the magnitude of D_t's effect. Since the empirical evidence is consistent across four datasets and the concern is addressable, I do not move the reader's CONDITIONAL verdict.","tokens_in":34492,"tokens_out":12552,"duration_ms":138216,"concrete_test":"On CIFAR-10 with the same hyperparameters and budget, train QC-FM-Mixture (p=0.2, k=16) once with the projection frame U resampled at every step and once with a single fixed random Haar frame U held constant throughout training (also evaluate a fixed PCA frame for reference). Report FID-50K for at least 3 seeds per condition. If the fixed-frame FID is substantially lower (e.g., >5% relative), the resampling term D_t is a major penalty and the per-frame guarantees do not transfer; if FID is within run-to-run noise, frame resampling is not the bottleneck and the concern is minor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The theoretical guarantees (Theorem 2, Theorem A.4) apply to a fixed frame U and to the ideal field of the per-frame QC coupling. In actual training, U is redrawn every step and is not observed by the velocity network. Proposition A.8 decomposes the frame-averaged irreducible variance into the mean per-frame variance plus a nonnegative resampling term D_t = E||E[Y|x_t,U] − E[Y|x_t]||^2. The paper provides no upper bound on D_t for real data that would guarantee the averaged training floor remains below the Baseline floor; Proposition A.9's bound is loose, and Proposition A.10's exact analysis is restricted to isotropic Gaussian data. The hybrid variants dilute the mechanism further: with p=0.2, only 20% of pairs are QC anchors, so the slice variance of the mixed objective is at best (1−p) times the Baseline slice variance. Thus the theory does not establish that the actual QC-FM training objective is better or that FID improves. The empirical FID gains are the only support, but they lack error bars, multiple seeds, and released code, so the central claim rests on an unquantified gap between per-frame theory and frame-averaged practice.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Quantile Coupling Flow Matching (QC-FM), a one-sided coupling for flow matching that constructs source samples by mapping within-batch data ranks along k random orthonormal projections to Gaussian quantiles and completing the remaining directions by conditional Gaussian sampling. The method avoids pairwise cost matrices and batchwise assignment, with an O(B log B) per-batch cost. The paper develops a population-level theory for the QC primitive: for a fixed frame, the irreducible regression variance vanishes along the selected slices and the ideal per-slice flow is straight; it also analyzes transport cost, direction resampling, finite-batch approximation, and the copula-type dependence among slice codes. For training, the paper introduces two hybrid variants, QC-FM-Mixture and QC-FM-Adjacency, which apply QC to an anchor subset and pair the Gaussian remainder differently. Experiments on CIFAR-10, CelebA-64, FFHQ-64, and ImageNet-64 report FID improvements over an independent Baseline and over Hungarian OT-CFM under matched budgets, with the Mixture variant reducing FID by 6.3–12.9% relative to Baseline.","tokens_in":34822,"tokens_out":4658,"duration_ms":56738,"significance":"If the theoretical claims are taken as scoped to the fixed-frame primitive, they are a clean and useful contribution: the per-frame variance decomposition and slice-straightness results are carefully stated and the proofs in Appendix A are coherent, and the explicit treatment of the frame-resampling gap (Proposition A.8) and finite-batch consistency is more honest than is typical in this area. The algorithmic idea is simple, efficient, and plausibly of practical interest as a lightweight alternative to mini-batch OT. However, the central empirical claim—that QC-FM-Mixture improves FID across four datasets—rests on single runs without error bars or multiple seeds, and the theory does not directly cover the frame-averaged, U-blind training setup used in the experiments. The paper explicitly acknowledges several of these limitations, which is a strength, but the gap between the per-frame theory and the deployed hybrid objective remains a load-bearing issue.","major_comments":[{"comment":"The central empirical claim is supported only by single FID values with no error bars, confidence intervals, or multiple seeds. For example, QC-FM-Adjacency improves on CelebA by only 0.6% (1.76 vs 1.75) and OT-CFM on ImageNet is within 0.1% (8.80 vs 8.81); these differences could be within run-to-run noise. Please report mean and standard deviation over at least three independent training runs (or equivalent uncertainty quantification) for all main FID comparisons, and state how many seeds Table 2 and Table 3 use. The toy diagnostics in Table 1 also report a single seed ('same seed'). Without this, the headline improvement over Baseline and OT-CFM is not statistically grounded.","section":"Table 2 and Section 'Image Generation Results'"},{"comment":"The theoretical guarantees (Theorem 2, Theorem A.4) are explicitly per-frame, but training redraws U every step and the network does not observe U. Proposition A.8 shows the frame-averaged irreducible variance equals the mean per-frame variance plus a nonnegative resampling term D_t, and the paper provides no practical upper bound on D_t for real data. The manuscript itself states that 'These results concern the QC primitive and do not bound the floor of the full hybrid coupling used in the experiments.' This is an important limitation: the theory does not establish that the U-blind training objective is better than Baseline, and the empirical FID gains are the sole support for that claim. Please either provide a usable bound or, at minimum, add an empirical diagnostic measuring D_t (e.g., the norm of the difference between per-frame ideal fields on real data) to quantify the magnitude o","section":"Appendix A.8, Proposition A.8 and surrounding text"},{"comment":"All image experiments use the hybrids, not the pure QC primitive, and for Mixture the population pair law is π_mix = p π_qc + (1-p) π_ind with p=0.2. This means the mixed objective is L_mix = p L_qc + (1-p) L_ind, so even a perfect per-frame QC floor leaves a (1-p) fraction of the Baseline floor in the hybrid objective. Proposition A.6 bounds only the source marginal, not the hybrid irreducible variance. The paper should state more prominently that Theorem 2 does not transfer to the hybrid objective, and discuss what the dilution implies for interpreting the FID gains as evidence for the QC mechanism rather than for the particular hybrid completion.","section":"Section 'Gaussian-Remainder Hybrid Extensions'"}],"minor_comments":[{"comment":"The text says 'the QC coupling above is the primitive analyzed in Section .' with a missing section number. Please fill in the reference.","section":"Section 'Gaussian-Remainder Hybrid Extensions'"},{"comment":"Algorithm 1 takes an input batch {x(i)}_{i=1}^B, but in the hybrids it is called on the anchor subset Bx,anc of size M. This is clear from the prose but could be confusing in the pseudocode; consider adding a comment that B may be M in the hybrid call.","section":"Algorithm 1 and Algorithm 2"},{"comment":"The notation N(0, P_U^⊥) for the complement Gaussian is nonstandard; writing N(0, I_d - UU^T) or specifying that the covariance is the projection matrix would improve readability.","section":"Appendix A.1 and throughout"},{"comment":"The calibrated heuristic Q(k) depends on ω fitted to CIFAR-10, and Table A transfers this to other datasets. This is a free parameter; the paper should explicitly state that the transferred k values are heuristic and that the FID gains could be sensitive to this calibration, even though the heuristic is data-driven.","section":"Appendix B"},{"comment":"Table D is referenced as 'Table D in Appendix C.5' and appears in the appendix; please ensure the in-text reference is to the correct appendix label (C.5 contains Table D, but C.5 is titled 'Batch-Size Ablation' while Table D is in the same section).","section":"Appendix C.5, Table D"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid, honestly written contribution, but the empirical core needs more rigor before publication. The lack of multiple seeds and error bars is a standard requirement for a claim of consistent FID improvement across four datasets. The frame-resampling gap is acknowledged, but the paper's framing in the abstract and introduction could overstate the transfer from per-frame theory to the deployed hybrid; the authors should either add a real-data diagnostic for D_t or consistently qualify the theoretical claims as applying to the primitive only. I do not see a fundamental correctness error in the main derivations; the issues are substantial enough to require major revision but are fixable within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nYou should know this paper for its core idea: instead of matching two batches, QC-FM constructs each source directly from within-batch ranks along k random projections, mapped to Gaussian quantiles, with the complement filled by conditional Gaussian sampling. That one-sided construction is new to me and it's clean, with no pairwise cost matrix and O(B log B) complexity. The authors also do something increasingly rare: they state clearly what the theory does and does not cover. The fixed-frame results (per-slice irreducible variance vanishes, ideal slice flow straight) are derived carefully, and Appendix A.8 openly acknowledges that the frame is redrawn and not observed by the network, decomposing the averaged objective into a per-frame term plus a nonnegative resampling term D_t. They then bound D_t and analyze it exactly for isotropic Gaussian data. That is honest, and it's more than most papers in this area do.\n\nThe weaknesses are real but mostly empirical. The FID comparisons in Table 2 are single-point numbers, with no error bars or multiple seeds. Hyperparameters (k, p) and the calibrated weight ω are selected on CIFAR-10 and transferred to the other datasets, which is reasonable as a protocol but would benefit from a per-dataset ablation. Code is not released. Also, since the Mixture variant uses p=0.2, only 20% of pairs are QC anchors; the theoretical per-frame guarantees apply to that fraction, so the empirical gain is doing the heavy lifting. The stress-test note about D_t is fair in spirit, but I'd soften it: the paper doesn't claim the per-frame theory bounds the hybrid's floor—it says so explicitly in the main text—and it provides partial bounds plus an exact Gaussian analysis. The gap is real for general data, not a hidden flaw.\n\nWho should read this: anyone working on flow-matching couplings or efficient structured transport. The method is simple enough to implement and the theory gives a principled way to think about rank-based bias. I'd want to see reproducibility (code, seeds, FID distributions) before relying on the quantitative claims, but the idea and the analysis deserve refereeing. My verdict: accept for serious review, with the empirical section requiring revision.","headline":"One-sided quantile coupling for flow matching is a genuinely new, clearly explained idea with honest theory; the empirical wins are plausible but single-run, so treat the FID numbers with caution.","tokens_in":35246,"tokens_out":2807,"would_cite":true,"duration_ms":32707,"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":"A one-sided rank-based coupling straightens flow-matching training and cuts FID by up to 12.9%.","keywords":["flow matching","quantile coupling","one-sided coupling","comonotone transport","sliced projections","generative modeling","FID","source–target coupling"],"falsifier":"On a held-out CIFAR-10 batch at the paper's settings ($k=16$, $p=0.2$), compute $D_t = \\mathbb{E}\\lVert E[Y\\mid x_t,U] - E[Y\\mid x_t]\\rVert^2$ by Monte Carlo over resampled frames. If $D_t$ is large relative to the per-frame slice variance it replaces, the fixed-frame variance-elimination theorem cannot explain the FID gains; if the FID gap between QC-FM and Baseline persists when $D_t$ is made negligible by conditioning the network on $U$, then the mechanism is the frame-averaged coupling rather than the per-frame one.","tokens_in":34415,"feed_emoji":"🧮","tokens_out":9586,"duration_ms":89914,"temperature":0.7,"pith_summary":"Flow matching trains a generative model by teaching a network to reproduce the velocity of straight paths from a Gaussian source to data. The paper claims that the hard part—choosing which noise sample goes with which data point—can be done one-sidedly, without ever matching two batches. QC-FM ranks each data point's projections along $k$ random orthogonal directions, maps the within-batch ranks to Gaussian quantiles, and fills the remaining directions with Gaussian noise. For a fixed projection frame this removes the per-slice regression variance that no network could reduce, and it makes the ideal flow exactly straight along those slices. Under matched training budgets, QC-FM-Mixture improves FID by 6.3--12.9\\% over the independent Baseline on CIFAR-10, CelebA, FFHQ, and ImageNet-64, and outperforms Hungarian OT-CFM on all four, at a per-batch cost of $O(B\\log B)$ instead of a batchwise assignment.","feed_headline":"Projected ranks cut flow-matching FID by up to 12.9%","feed_subtitle":"No cost matrix, no assignment: a one-sided quantile coupling straightens per-slice flows and beats OT-CFM.","key_machinery":"The central object is the one-sided quantile coupling: a random orthonormal frame $U=[u_1,\\dots,u_k]$ projects each data point; within-batch ranks per slice are mapped through the inverse Gaussian CDF to codes $z^{(i)}$, and the source is $U z^{(i)}$ plus independent Gaussian noise in the orthogonal complement. Along each slice this realizes the comonotone coupling between the projected data law and the standard Gaussian, which is what makes the per-slice velocity single-valued (zero irreducible variance) and the per-slice flow a straight $\\mathrm{W}_2$-geodesic. The frame-averaged analysis adds the resampling term $D_t = \\mathbb{E}\\lVert \\mathbb{E}[Y \\mid x_t, U] - \\mathbb{E}[Y \\mid x_t]\\rV","core_discovery":"QC-FM pairs each data point $x^{(i)}$ with a source $e^{(i)} = U z^{(i)} + (I_d - U U^{\\top})\\varepsilon^{(i)}$, where $U$ is a random orthonormal frame, $z^{(i)}$ holds the Gaussian quantiles of the within-batch ranks of the projections $u_j^{\\top} x^{(i)}$, and $\\varepsilon^{(i)}$ is Gaussian noise in the complement. For a fixed frame this is the comonotone optimal transport along each slice: slice velocity is a deterministic function of the state, so irreducible regression variance vanishes along each selected slice, the ideal flow is exactly straight, and the slice marginal is generated exactly. The only deviation from the standard Gaussian prior is a $k$-dimensional dependence among sli","pith_inferences":["The one-sided construction suggests a natural online adaptive scheme: compute the debiased dependence statistic $\\hat b_B$ during training and adjust $k$ on the fly, rather than calibrating $k$ per dataset before training.","Because QC-FM only needs projected order statistics, it should transfer to modalities where Euclidean distances are not meaningful—text, graphs, or audio—provided one can define a projection.","The per-slice straightness guarantee may make QC-FM a useful initialization for distillation or for flow models that aim to learn near-straight trajectories, since the learned map can concentrate its capacity on the inter-slice dependence rather than on path curvature."],"forward_implications":["Structured-coupling benefits no longer require solving a batch assignment: QC-FM's per-batch cost is $O(k B \\log B + d k^2)$ plus $O(B d k)$ for projections, against $O(B^2 d + B^3)$ for exact Hungarian OT-CFM.","Because the source marginal is exactly Gaussian at $k=1$ and only mildly perturbed for $k\\ge 2$, QC-FM is a drop-in change to existing flow-matching training loops: only the endpoint pairing changes.","The trade-off between slice count $k$ and source-prior mismatch is explicit and measurable: transport cost falls linearly in $k$ while the covariance defect grows like $k(k-1)\\varrho$, explaining the paper's non-monotone $k$-ablation (FID improves from $k=1$ to $k=16$, degrades at $k=64$).","The Mixture hybrid retains the Baseline objective on a $1-p$ fraction of the batch, so it interpolates between independent coupling and full QC structure; the paper's anchor-ratio sweeps show the best mixture point depends on how the remainder is paired."],"supporting_citations":[{"why":"Defines the flow-matching objective and the irreducible-variance decomposition that QC-FM's theorems target.","marker":"Lipman et al. 2023"},{"why":"Provides the Hungarian OT-CFM minibatch-coupling baseline that QC-FM is compared against and whose quadratic cost it avoids.","marker":"Tong et al. 2024"},{"why":"Supplies the one-dimensional comonotone coupling optimality that makes rank-to-quantile mapping the per-slice optimal transport.","marker":"Villani 2009"},{"why":"Gives the variance split used in Theorem 2 to separate network approximation error from irreducible per-point velocity spread.","marker":"Helbling et al. 2026"},{"why":"Provides the empirical-measure Wasserstein rates used to prove finite-batch consistency of the rank-grid source.","marker":"Bobkov and Ledoux 2019"},{"why":"The batchwise sort/quantile coupling used as a controlled comparison to isolate the effect of one-sided construction in Table 3.","marker":"Groom et al. 2026"}],"fun_headline_variants":["Ranks straighten flows, FID down 12.9%","One-sided quantile coupling beats OT-CFM","No cost matrix: quantile coupling lowers FID 12.9%","Straight slices: quantile coupling improves FID","QC-FM: one-sided, no assignment, better FID"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The fixed-frame guarantees—zero slice variance and exactly straight slice flow—are assumed to survive actual training, where the projection frame is redrawn every step and the velocity network never observes it; the paper's own decomposition shows the frame-averaged objective contains an unmeasured nonnegative resampling term $D_t$, so the per-frame theory alone does not bound the trained model's loss or final FID.","fun_headline_variants_meta":{"raw":{"variants":["Ranks straighten flows, FID down 12.9%","One-sided quantile coupling beats OT-CFM","No cost matrix: quantile coupling lowers FID 12.9%","Straight slices: quantile coupling improves FID","QC-FM: one-sided, no assignment, better FID"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001065,"raw_usage":{"total_tokens":4372,"prompt_tokens":883,"completion_tokens":3489,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":3402}},"tokens_in":627,"tokens_out":3489,"duration_ms":27692,"temperature":1.0,"reasoning_tokens":3402,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:34:16.615004+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out CIFAR-10 batch at the paper's settings ($k=16$, $p=0.2$), compute $D_t = \\mathbb{E}\\lVert E[Y\\mid x_t,U] - E[Y\\mid x_t]\\rVert^2$ by Monte Carlo over resampled frames. If $D_t$ is large relative to the per-frame slice variance it replaces, the fixed-frame variance-elimination theorem cannot explain the FID gains; if the FID gap between QC-FM and Baseline persists when $D_t$ is made negligible by conditioning the network on $U$, then the mechanism is the frame-averaged coupling rather than the per-frame one.","supporting_citations":[],"review_version":1}