{"id":"b7b2b6d6-03d6-4e67-a1d7-d34ea8be4051","arxiv_id":"2504.20179","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Integration Flow learns the integrated denoising map of an ODE generative model and reports competitive one-step FID on CIFAR-10 and ImageNet for VE diffusion, rectified flow, and PFGM++.","lead":"This paper proposes Integration Flow, a way to train generative models by learning the whole ODE-based denoising path in one step rather than solving the ODE numerically. It reports one-step image generation on CIFAR-10 and ImageNet with FIDs around 2.9 to 4.4 for diffusion, rectified flow, and PFGM++.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The printed training loop feeds a stale estimate from a different sample into every update, so the optimal network must ignore that input and the reported one-step results are not reproducible from the described procedure.","rationale":"I read the paper as proposing a unified one-step estimator for ODE-based generative models, with the central empirical claim being the reported one-step FIDs. For that claim to hold, the training loop must produce a network whose iterative update rule is effective at inference. The reader identified the under-specification of the training/inference loop as the weakest assumption, and my reading agrees: the pseudo-code for training is not merely underspecified but, taken literally, uses an input x0^(n) that is independent of the current target and observation. In that setting the optimal squared-error or Pseudo-Huber predictor is constant in x0^(n), so the iterative refinement shown at inference cannot have been learned by the printed procedure. This is a correctness risk for the central empirical claim, not merely a missing implementation detail. The theoretical results in the paper do not rescue the procedure: Theorem 1's variance decomposition assumes the additional conditioning variable is informative, which fails under the printed sampling scheme. The other theoretical issues noted by the reader, such as the inconsistency in Theorem 2 and the Jensen-inequality argument in Theorem 3, are real but secondary; they concern the proofs' validity rather than the reproducibility of the headline numbers. My recommended verdict is unchanged from the reader's CONDITIONAL: the paper should not be accepted as-is until the training loop is clarified and the experiment is reproduced under the corrected procedure.","tokens_in":16190,"tokens_out":7844,"duration_ms":81621,"concrete_test":"Train the VE variant on CIFAR-10 with identical hyperparameters under two regimes: (i) the printed Algorithm 1 using the global stale x0^(n); (ii) a per-sample unrolled loop where, for each sampled (x0, xt, t), one initializes x0^(0)~N(0,I) and runs K=8 updates x0^(k+1)=g_theta(x0^(k), xt, t) with the same (x0, xt, t) before one gradient step, accumulating the loss over the K outputs. Then evaluate one-step FID under Algorithm 2. If regime (i) reproduces the claimed 2.86 and regime (ii) differs, the printed procedure is self-contained and the concern is resolved; if regime (i) fails to train or gives a much worse FID while (ii) reproduces the claim, the paper omits a required unrolling step.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 (and the analogous loops for Rectified Flow and PFGM++) maintains a single global counter n and at each training step computes x0^(n+1)=g_theta(x0^(n), xt, t), where x0^(n) is the iterate produced at the previous training step for a previously sampled (x0_prev, xt_prev, t_prev). Because the current (x0, xt, t) is drawn independently, x0^(n) is statistically independent of the current x0 and xt. The Bayes-optimal predictor E[x0 | xt, x0^(n)] therefore equals E[x0 | xt]; conditioning on x0^(n) carries no information, and any dependence of g_theta on x0^(n) can only increase the expected loss. Theorem 1, which motivates the anchor state by comparing E[A|B,C] with E[A|B], does not apply because its C is uninformative under this training procedure. At inference, however, Algorithm 2 iterates on the same xT, so after the first application x0^(n) becomes correlated with xT and the network is evaluated in a regime the printed training loop never creates. Section 3.3 describes progressively refined estimates as if iterations are unrolled on a fixed (x0, xt, t), but no algorithm specifies such an inner loop. The reported one-step FIDs (2.86, 3.36, 2.91) thus rest on a training dynamic that is either omitted from the paper or, if implemented exactly as printed, degenerates to a map that must ignore its iterative input.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Integration Flow, a framework that directly learns the integrated effect of an ODE-based generative trajectory rather than solving the ODE numerically. The method defines g(x0, xt, t) = xt - G(x0, xt, t), incorporating the target state x0 as an \"anchor,\" and uses an iterative refinement x0^(n+1) = at xt + bt G_theta(x0^(n), xt, t). The authors provide three theorems (stability, non-intersection, optimality), derive specialized forms for VE diffusion, Rectified Flow, and PFGM++, and report one-step FIDs on CIFAR-10 and ImageNet 64x64 (e.g., 2.86, 3.36, 2.91 for VE, RF, PFGM++ on CIFAR-10). The paper claims to be the first unified structure for ODE-based generative models and the first to show exact straightness of 1-Rectified Flow without reflow.","tokens_in":2229,"tokens_out":2840,"duration_ms":92355,"significance":"If the empirical results were reproducible and the theory sound, the paper would make a valuable contribution: a single framework enabling one-step generation across several ODE-based generative models with competitive FIDs would be practically significant. The related work is reasonably comprehensive and the use of Pseudo-Huber loss and preconditioning is sensible. However, the manuscript does not include code, model checkpoints, or error bars, and the theoretical support contains serious internal inconsistencies. The central training algorithm as printed makes the iterative input statistically independent of the target, so the reported one-step results cannot be attributed to the described method. The paper's strengths are its clear organization and the breadth of empirical comparisons, but these do not compensate for the load-bearing flaws in the method description and theory.","major_comments":[{"comment":"The training loop in Algorithm 1 (and Algorithms 3 and 5) carries x0^(n) across independent training steps. At step n, x0^(n) is the output produced for a previously sampled (x0, xt, t), and the current (x0, xt, t) is drawn independently. Hence x0^(n) carries no information about the current x0, and the Bayes-optimal predictor E[x0 | xt, x0^(n)] equals E[x0 | xt]; any dependence of g_theta on x0^(n) can only increase the loss. Theorem 1, which justifies the anchor state by comparing E[A|B,C] with E[A|B], does not apply because its C is uninformative under this procedure. At inference (Algorithm 2), iterations are run on the same xT, so x0^(n) becomes correlated with xT, a regime the training loop never creates. No inner loop unrolling iterations for the same sample is specified. As written, the training procedure forces the network to ignore x0^(n), and the reported one-step FIDs (Sections 5.2, Tables 2 and 3) are not reproducible from the description. Please provide a correct training algorithm (e.g., unrolling k steps per sample with backpropagation through the unrolled graph, or a clarified procedure) and re-run the experiments under that procedure.","section":"Section 3.3, Eq. (11); Algorithm 1"},{"comment":"Theorem 2 is internally inconsistent with Eq. (7). Equation (7) defines g(x0, xt, t) := x0, independently of xt for a fixed x0. Then for any two distinct states xt != yt with the same x0, the left-hand side of Eq. (13) is ||g(x0, xt, t) - g(x0, yt, t)|| = ||x0 - x0|| = 0, which violates the claimed lower bound e^{-Lt}||xt - yt|| > 0. The proof in Appendix B.2 treats g as the flow map of the reverse ODE, i.e., the solution at time 0 starting from xt, but the earlier definition identifies g with x0 itself. The bi-Lipschitz property and the non-intersection conclusion therefore do not follow from the stated definitions; Theorem 2 must be reformulated or removed.","section":"Section 3.2, Eq. (7); Section 3.4, Theorem 2; Appendix B.2"},{"comment":"Theorem 1 is a conditional-variance inequality: it states that E[(A - E[A|B,C])^2] <= E[(A - E[A|B])^2]. This is a standard fact, but the proof in Appendix B.1 asserts Var(A|B,C) <= Var(A|B) without proving it, and more importantly the theorem is used to justify the anchor state x0^(n). Under the training procedure in Algorithm 1, x0^(n) is not informative about x0, so the inequality is vacuous for the method as described. The theorem also assumes the estimators g'_theta and g_theta can represent the exact conditional expectations, which is not discussed. The theoretical justification of the anchor state is therefore not established.","section":"Section 3.4, Theorem 1; Appendix B.1"},{"comment":"The claim that Integration Flow is \"optimal\" for Flow Matching/Rectified Flow is not supported by the proof. The proof uses Jensen's inequality to show that the time-integrated objective is a lower bound on the flow-matching objective, i.e., min_v E ||∫(x1 - x0 - v(xt,t))dt||^2 <= min_v E ∫||(x1 - x0) - v(xt,t)||^2 dt. This inequality alone does not imply that the integrated approach achieves the best possible value; it only shows that the integrated objective is a relaxation. No argument is given that the minimum of the relaxed objective equals the minimum of the original objective, or that the learned G attains that minimum. The theorem's conclusion, \"a direct one-step method (via G) achieves the best possible value,\" is therefore not proven. Additionally, the theorem's notation (x0 for the source, x1 for the target) conflicts with the rest of the paper, where x0 is the data sample.","section":"Section 3.4, Theorem 3; Appendix B.3"},{"comment":"The empirical evaluation is not independently verifiable: no code, checkpoints, or random seeds are provided, and Tables 2 and 3 report FIDs without error bars. The claim that \"the flow has already become exactly straight\" for 1-Rectified Flow (Section 5.2, Figure 2) is not supported by any quantitative straightness metric or by a comparison of trajectory curvature before and after training. The reported one-step results also depend on the training/inference mismatch described above, so as they stand they do not validate the method. Please provide code/checkpoints or at least detailed experimental logs, and either provide a direct measurement of path straightness or temper the claim.","section":"Section 5; Figure 2"}],"minor_comments":[{"comment":"In Theorem 3 and its proof, the notation x0 and x1 is swapped relative to the main text, where x0 is the target data sample and xT (or z1) is the noise. This should be fixed to avoid confusion.","section":"Appendix B.3"},{"comment":"Both algorithms contain a duplicated initialization line: \"initialize x0^(0) ~ N(0,I)\" appears twice in Algorithm 4, and similar duplication occurs in Algorithm 6. One initialization suffices.","section":"Algorithms 4 and 6"},{"comment":"The caption says panels (a)-(b) are \"regenerated from (Liu et al., 2022)\" but it is unclear which panels are original and which are produced by the proposed method; add axis labels and a clear legend.","section":"Figure 2"},{"comment":"The row for \"1-Rectified flow\" reports NFE=1, FID=378; this seems unusually high even for one-step rectified flow and may be a typographical error (e.g., 37.8 or 3.78). Please verify.","section":"Table 2"},{"comment":"Equation (32) defines g_theta(x0, xt, t) = κ(σt) xt + 1/(1-κ(σt)) G_theta(x0, xt, t), but the text and Table 1 specify b_t = 1 - κ(σt), so the coefficient in front of G_theta appears inconsistent. Please check the derivation.","section":"Appendix A.1, Eq. (32)"}],"recommendation":"major_revision","confidential_remarks":"The concerns in this report are substantial and go to the core of the method. If the authors can provide a corrected training algorithm and re-run all experiments, plus fix Theorem 2 and Theorem 3, the paper could become publishable. Given the absence of code and the formal incoherence of the printed training loop, I would not accept the current version; however, the underlying idea is not obviously invalid, so I am not recommending rejection outright. I would ask the editor to weight the authors' response carefully: an 'it was a typo' response without corrected experiments and code should not be sufficient."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper reports one-step FID numbers for VE diffusion, rectified flow, and PFGM++ that are good enough to matter if real, but the procedure in Algorithm 1 cannot produce them, and the theoretical sections contain clear errors. I read the stress-test and agree with it.\n\nWhat is genuinely new: extending the DDDM idea to a unified integration-flow framework across three ODE-based generative models, with preconditioning for VE and PFGM++, and reporting one-step results without reflow for rectified flow. The numbers (2.86/3.36/2.91 on CIFAR-10) are competitive with consistency-model distillations. The writing is clear.\n\nWhere it falls apart:\n1. Training loop. Algorithm 1 (and 3, 5) maintains one global x0^(n) that is updated from a previously sampled (x0, xt, t). Because the current sample is independent, x0^(n) carries no information about the current x0 given xt. The optimal network must ignore its first argument, so the iterative refinement at inference is never trained. The paper mentions storing x0^(n) in a buffer, but the algorithm does not specify per-sample storage or unrolling. As printed, the method cannot work.\n2. Theorem 2. Eq. (7) says g(x0, xt, t) = x0. For fixed x0, this map is constant in xt, so it cannot be bi-Lipschitz with the stated lower bound. The proof also applies Grönwall to the ODE flow map but assumes the network is exactly the true g, which makes the conclusion trivially false unless x0^(n)=x0.\n3. Theorem 3. The Jensen step gives a lower bound on the flow-matching objective. That does not show the integrated approach is optimal; it shows the integrated objective is a relaxation.\n\nNo code, no checkpoints, no error bars. The experiments are described in enough detail that a careful reader could reimplement, but the missing training-loop specification is not a detail: it is the load-bearing part.\n\nWho this is for: researchers doing fast ODE sampling/distillation who want to see the DDDM recipe carried further. The empirical claims are worth a look if the authors ever release code.\n\nRecommendation: I would not send this to peer review as is. The contradictions and the undefined training procedure are too fundamental. I'd ask the authors for a corrected algorithm and code, then look again.","headline":"One-step results worth a look, but the printed training loop is degenerate and Theorem 2 contradicts the paper's own Eq. (7); not ready for serious scrutiny.","tokens_in":17098,"tokens_out":5245,"would_cite":false,"duration_ms":50972,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Integration Flow is a generative framework that learns the integral of an ODE trajectory directly, so diffusion, rectified flow, and PFGM++ can generate in one network evaluation.","keywords":["integration flow","generative modeling","one-step generation","ODE-based generative models","diffusion models","rectified flow","PFGM++","trajectory integral"],"falsifier":"Take an ODE-based generative model with a closed-form velocity, such as VE diffusion on a Gaussian mixture whose score is known exactly; if a trained Integration Flow's one-step output deviates from the exact ODE solution by more than the discretization error of a standard solver at one step, then the claim that it learns the integral without solver error would be falsified.","tokens_in":15981,"feed_emoji":"⚡","tokens_out":7634,"duration_ms":65227,"temperature":0.7,"pith_summary":"Integration Flow is a generative modeling framework that learns the total effect of an ODE trajectory directly, rather than solving the ODE step by step. The paper claims that this makes one-step generation possible for a unified family of ODE-based models, specifically the VE diffusion model, rectified flow, and PFGM++. It reports CIFAR-10 FIDs of 2.86, 3.36, and 2.91, and ImageNet 64x64 FIDs of 4.09, 4.35, and 4.15, all with a single function evaluation. A reader would care because sampling speed is the main practical bottleneck of ODE-based generative models, and the paper presents a single mechanism that removes the iterative solver from all three of them.","feed_headline":"One learned integral makes diffusion-style models one-step generators","feed_subtitle":"A single network evaluation replaces iterative ODE solves: CIFAR-10 FID 2.86 for VE diffusion.","key_machinery":"The load-bearing identity is the antiderivative relation $\\int_t^0 v(x_s, s)\\,ds = V(x_0, 0) - V(x_t, t)$, rewritten as $x_0 = g(x_0, x_t, t) = x_t - G(x_0, x_t, t)$. A neural network $G_\\theta$ replaces $G$, and the time-dependent coefficients $a_t$ and $b_t$ reparameterize the map to keep training stable when noise variance is large, as it is in VE diffusion and PFGM++. The iterative update $x_0^{(n+1)} = a_t x_t + b_t G_\\theta(x_0^{(n)}, x_t, t)$ is what lets the estimate converge, and the Pseudo-Huber loss $d(x, y) = \\sqrt{\\|x-y\\|_2^2 + c^2} - c$ is what the network is trained against.","core_discovery":"Integration Flow defines the map $g_\\theta(x_0, x_t, t) = a_t x_t + b_t G_\\theta(x_0, x_t, t)$, where $G_\\theta$ approximates the integrated reverse-time dynamics $G(x_0, x_t, t) = \\int_t^0 v(x_s, s)\\,ds$ from an intermediate state $x_t$ to the target $x_0$, with $x_0$ used as an explicit anchor. The paper argues that after training, $g_\\theta(x_0^{(n)}, x_t, t)$ refines an estimate $x_0^{(n)}$ toward the true $x_0$, so that one pass of the network is enough to generate a sample. The authors claim that this single structure covers VE diffusion, rectified flow, and PFGM++, that it makes 1-rectified flow exactly straight without reflow, and that the anchor-state formulation is provably at least as accurate in mean squared error as conditioning on $x_t$ alone.","pith_inferences":["If the antiderivative identity generalizes beyond the three implementations in the paper, the same $g_\\theta$ construction should apply to any well-defined ODE-based generative model, including stochastic interpolants and flow matching variants, by setting $a_t$ and $b_t$ from the noise scheduler.","The stability theorem's conditioning argument suggests that feeding the current estimate $x_0^{(n)}$ back into the network should make error shrink with each iteration; a direct test would be to measure whether the multistep FID improvement matches the contraction predicted by the bi-Lipschitz bound.","The reported exact straightness of 1-rectified flow implies the training objective implicitly regularizes trajectory curvature; this could be tested by measuring the expected path curvature of learned trajectories against a baseline rectified flow trained with the usual matching loss.","The paper notes extra training memory comes from carrying $x_0^{(n)}$; an off-GPU buffer would make the method cheaper to train, so the memory cost is an engineering constraint rather than a fundamental one."],"forward_implications":["For the three models tested, one function evaluation replaces an ODE solve: VE diffusion reaches FID 2.86 on CIFAR-10 and 4.09 on ImageNet 64x64 at NFE 1.","Rectified flow reaches competitive one-step quality without reflow, with FID 3.36 on CIFAR-10 and 4.35 on ImageNet, making the learned flow exactly straight in the 1-rectified case.","PFGM++ becomes a one-step generator, with FID 2.91 on CIFAR-10 and 4.15 on ImageNet, which the authors state has not been shown before.","The same architecture, loss, and training loop apply to all three models, with only the noise scheduler and the $a_t, b_t$ coefficients changing, so the framework unifies models that previously required separate treatment.","A second network evaluation further improves quality, such as the CIFAR-10 VE FID dropping from 2.86 to 2.62, so the learned integral also serves as the initialization for cheap few-step refinement."],"supporting_citations":[{"why":"Supplies the probability-flow ODE and VE diffusion setup that Integration Flow learns the integral of.","marker":"(Song et al., 2020b)"},{"why":"Defines rectified flow and the reflow procedure that the paper claims to bypass with exactly straight one-step paths.","marker":"(Liu et al., 2022)"},{"why":"Defines PFGM++ and the ODE/alignment relation $r = \\sigma\\sqrt{D}$ that Integration Flow reuses for its PFGM++ variant.","marker":"(Xu et al., 2023)"},{"why":"Introduces the directly denoising formulation that the paper presents as a VP-only special case of Integration Flow and whose parameterization motivates $G_\\theta$.","marker":"(Zhang et al., 2024)"},{"why":"Provides the improved consistency-training techniques and loss design that the paper adapts for stable one-step training.","marker":"(Song & Dhariwal, 2023)"},{"why":"Supplies the noise schedule and preconditioning analysis that informs the $a_t, b_t$ choice in the VE case.","marker":"(Karras et al., 2022)"},{"why":"Provides the ADM architecture that all Integration Flow experiments use.","marker":"(Dhariwal & Nichol, 2021)"}],"fun_headline_variants":["Learn the ODE integral, not the solver: one-step generative models","Integration Flow: anchor to x0, one pass to generate","First unified ODE-generative model with exact one-step generation","Learn the integral, skip the solve: one-step diffusion on CIFAR-10","Integration Flow: first exact straight 1-rectified flow, no reflow"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported one-step results rest on the iterative training update $x_0^{(n+1)} = a_t x_t + b_t G_\\theta(x_0^{(n)}, x_t, t)$ being a well-defined and effective procedure, but the paper specifies reusing an iterate from a previous, independently sampled data point and time rather than unrolling iterations for the same sample, so the training dynamic that produces the reported accuracy is left underspecified.","fun_headline_variants_meta":{"raw":{"variants":["Learn the ODE integral, not the solver: one-step generative models","Integration Flow: anchor to x0, one pass to generate","First unified ODE-generative model with exact one-step generation","Learn the integral, skip the solve: one-step diffusion on CIFAR-10","Integration Flow: first exact straight 1-rectified flow, no reflow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00115,"raw_usage":{"total_tokens":4811,"prompt_tokens":1031,"completion_tokens":3780,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":3683}},"tokens_in":647,"tokens_out":3780,"duration_ms":27392,"temperature":1.0,"reasoning_tokens":3683,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:35:52.460682+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an ODE-based generative model with a closed-form velocity, such as VE diffusion on a Gaussian mixture whose score is known exactly; if a trained Integration Flow's one-step output deviates from the exact ODE solution by more than the discretization error of a standard solver at one step, then the claim that it learns the integral without solver error would be falsified.","supporting_citations":[],"review_version":1}