{"id":"2543ac59-5422-4fd0-a0bf-294b30893f1e","arxiv_id":"2506.22304","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Koopman-operator linearization of Conditional Flow Matching preserves teacher trajectories, enabling one-step sampling, image inversion, and linear spectral control of generative flows.","lead":"This paper trains a linear operator that, in a lifted space, reproduces the full non-linear trajectory of a pre-trained flow-based generative model, enabling one-step generation. It also uses the operator's spectral decomposition to edit, invert, and analyze images generated by the original model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Trajectory-preserving claim rests on latent-space MSE computed in the learned lifting itself; pixel-space trajectory fidelity on teacher rollouts is never reported, so intermediate-state preservation is not yet established.","rationale":"The paper is read in good faith: the Koopman linearization of CFM dynamics is a solid and novel idea, the consistency loss clearly improves trajectory fidelity relative to the endpoint-only ablation, and the code is released, which makes verification feasible. The reader's CONDITIONAL verdict is appropriate, and my concern does not move it. The reader identified the finite-dimensional invariant subspace existence as the weakest assumption; I agree that this is a real and acknowledged limitation. My stress-test points to a complementary, more immediately testable gap: the quantitative trajectory evidence is collected in the learned latent space and the training objective is evaluated on conditional-path samples, so the central 'trajectory-preserving' claim is not yet supported by pixel-space trajectory measurements on the teacher's actual rollouts. This is not an accusation of error; it is a request for a direct measurement that the released code can produce. If the pixel-space trajectory check passes, the central claim is substantially strengthened. If it fails, the contribution would be better described as endpoint-preserving latent linearization with an encouraging but unverified link to intermediate image states. The reader's conditional verdict already anticipates the need for additional validation, so no change to the verdict is recommended.","tokens_in":25679,"tokens_out":15104,"duration_ms":174077,"concrete_test":"Using the released code, generate 1000 held-out teacher rollouts {x_t} by integrating the OT-CFM teacher. For t in {0.25, 0.5, 0.75}, compute the mean LPIPS and MSE between the teacher state x_t and the decoded Koopman state g^{-1}(exp(tL)g(0,x_0)), and compare with teacher self-reconstruction LPIPS at the same t. Also evaluate the consistency residual ||Lg(t,x_t)-Dg(t,x_t)[1,v(t,x_t)]|| on these true rollout states and compare it with the same residual evaluated on conditional-path samples used in training. If the rollout residual is within the same order of magnitude as the conditional-sample residual and the LPIPS is comparable to teacher self-reconstruction, the trajectory-preserving claim is directly supported; if not, the current evidence only establishes latent-space linearity and endpoint fidelity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the Koopman system reproduces the teacher's full generative trajectory, not just endpoints. The consistency loss (Eq. 13) is the load-bearing term: it is what separates the method from endpoint distillation. Two gaps weaken the inference from this loss to the claim. First, Proposition 3's unbiasedness is relative to the mixture path p_t(x)=∫p_t(x|x1)q(x1)dx1; it is an unbiased estimator of the residual evaluated on conditional-path samples, not of the residual on the teacher's actual ODE rollouts. The two coincide only if the teacher's velocity field exactly matches the marginal velocity of p_t, which is not guaranteed for a fixed pretrained CFM and is precisely what finite-capacity training limits. Second, the reported trajectory-fidelity metrics (Table 1, Table 6, Fig. 10) compare encoded teacher rollouts g(t,x_t) with Koopman rollouts exp(tL)z0 in the learned latent space, which is the same representation in which linearity is imposed. Since L_recon is applied only at t=1, intermediate decoded states are not constrained by any reported quantitative metric. The semantic-editing and inversion results are qualitative and do not settle whether the decoded intermediate states match the teacher's intermediate images. The manuscript therefore does not yet verify trajectory preservation in the data space where generative modeling is defined; it verifies linearity of the learned latent encoding and endpoint accuracy, which is a weaker statement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Koopman-operator framework that linearizes the full generative dynamics of a pre-trained Continuous Flow Matching (CFM) model. The authors learn an encoder g_phi, a decoder g_psi^{-1}, and a linear generator matrix L such that the lifted coordinates evolve as dg/dt = L g, with an affine augmentation to handle the non-autonomous time dependence. The training objective combines phase, target, reconstruction, and a trajectory consistency loss; the consistency loss (Eq. 13) is claimed to be a simulation-free, unbiased estimator of the marginal consistency objective. Sampling is one-step via x1 = g^{-1}(exp(L) g(0, x0)). The paper further demonstrates applications in spectral mode analysis, semantic editing, inversion, and class-conditional spectral signatures, and reports competitive FID scores on MNIST, FFHQ, and CIFAR-10.","tokens_in":25999,"tokens_out":6163,"duration_ms":71260,"significance":"If the central trajectory-preservation claim is fully verified, this would be a meaningful advance: it would provide a single linear surrogate for the non-autonomous dynamics of a generative flow, enabling one-step parallel sampling while retaining fidelity to the teacher's intermediate states, and it would make spectral-analysis and control tools available for generative models. The paper is transparent about its main theoretical limitation (finite-dimensional invariant Koopman subspace), releases code, and includes an ablation clearly showing that the consistency loss is responsible for trajectory fidelity in the learned latent space. However, the current empirical verification is incomplete in a way that is load-bearing for the main claim: trajectory fidelity is only measured in the learned latent space, and the theoretical estimator is unbiased with respect to a mixture path rather than with respect to the teacher's actual ODE rollouts. The significance is therefore conditional on additional pixel-space trajectory evaluation.","major_comments":[{"comment":"The trajectory-fidelity claim is supported only by latent-space comparisons: the authors encode teacher rollouts g(t, x_t) and compare them to exp(tL) g(0, x0) in the learned Koopman space. Because the reconstruction loss L_recon (Eq. 10) is applied only at t=1, no reported quantity constrains the decoder's output at intermediate times. A model could have small latent MSE while decoding to pixel-space states that do not match the teacher's intermediate images. Since the abstract and introduction explicitly claim full-trajectory preservation, the manuscript must report pixel-space trajectory fidelity, e.g., the MSE or LPIPS between g^{-1}(exp(tL)g(0,x0)) and the teacher ODE rollout x_t on shared noise seeds, at several times t (and with and without the consistency loss). Absent such a measurement, the central claim is not empirically established.","section":"Sec. 6.2, Table 6"},{"comment":"Proposition 3 shows that Eq. (13) is an unbiased estimator of the marginal consistency loss with respect to the mixture path p_t(x) = ∫ p_t(x|x1) q(x1) dx1. However, the teacher's actual ODE trajectories are generated by the vector field v_t, and unless v_t is exactly the marginal velocity of p_t, the conditional-path samples are not distributed as teacher rollouts. For a pretrained CFM this holds only approximately, and it is precisely the approximation error that the linearization must tolerate. The paper should either justify that Eq. (13) controls the residual along true teacher rollouts, or quantify the discrepancy by evaluating the consistency residual on stored teacher trajectories and comparing it with the conditional-path estimate. Without this, 'trajectory-preserving' is only established relative to a surrogate path.","section":"Sec. 4.3, Eq. (13)"},{"comment":"The paper acknowledges that a finite-dimensional invariant Koopman subspace is 'a strong condition that need not hold for arbitrary nonlinear flows' and cannot be guaranteed a priori. This assumption is load-bearing: without it, the consistency loss cannot reach zero and the linearization is approximate. The current evidence (latent rollout MSE ~ 5e-6, competitive FID, and an intermediate optimal dimension in Fig. 5) is suggestive but indirect. A direct diagnostic would strengthen the claim significantly: report the held-out consistency-loss value L_cons (Eq. 13) and, ideally, the residual norm evaluated on actual teacher rollouts. If these residuals are not near zero, the paper should soften the trajectory-preservation claim accordingly.","section":"Sec. 6.3"},{"comment":"The semantic-mode evaluation is partially circular: modes are selected using CLIP embedding similarity, and then coherence (Eq. 19) is measured with the same CLIP embeddings. This selection-evaluation loop inflates the reported coherence scores, especially for the 'with consistency' model. The interpretability claims should be validated with an independent attribute classifier or human evaluation, or the paper should explicitly state that the reported coherence is relative to the CLIP space used for discovery rather than an independent semantic measure.","section":"Sec. 5.2 and Sec. 6.4, Table 2"}],"minor_comments":[{"comment":"The notation in Eq. (11) omits the time argument in the Jacobian: it should read ∇_x g_phi(t, x_t) · v_t(x_t), consistent with Eq. (13) and the surrounding text.","section":"Eq. (11)"},{"comment":"The caption says 'N=400, 800, 100, full'; the third value is likely meant to be 1000 rather than 100.","section":"Fig. 3"},{"comment":"The text states that 'the optimal FID value [is] at an intermediary Koopman dimension as seen in Figure 5', but the Figure 5 caption says 'the higher the dimension, the lower the FID'. These statements are inconsistent and should be reconciled.","section":"Sec. 6.3 vs Fig. 5"},{"comment":"The block structure in Eq. (7) would be clearer if the dimensions of the blocks (b_g, A_gt, A_gg) were specified, since the augmented state mixes scalar time with vector-valued observables.","section":"Sec. 4.1, Eq. (7)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the consistency loss in Eq. 13 is a genuine advance over prior Koopman distillation: it is simulation-free, unbiased for the marginal path, and clearly separates this work from Berman et al.'s endpoint-only matching. Second, the trajectory-preserving claim is not yet fully supported: all quantitative trajectory fidelity is measured in the learned latent space where linearity is assumed, and the paper never reports pixel-space intermediate rollouts against the teacher. That is the gap that makes me call the paper conditional rather than accept.\n\nCredit where due: the affine lift (Eq. 7) for the time coordinate is elegant, Prop. 2 identifies a real trap in using conditional velocities, Prop. 3 gives a correct unbiased estimator, and the authors are honest in Sec. 6.3 and Limitations that finite-dimensional invariant Koopman subspaces are not guaranteed. The code is available. The spectral decomposition and semantic editing results are interesting, though as demonstrations rather than rigorous evidence.\n\nThe soft spots, in size order. Largest: the latent-space MSE (Table 6, Fig. 10) shows the learned g makes the teacher's path approximately linear, but it does not show that exp(tL)z0 decoded into images matches the teacher's intermediate images. The stress-test note is right that this is endpoint accuracy plus latent linearity. Second: the consistency loss is evaluated on conditional path samples p_t(x_t|x_1); it is unbiased for the marginal path only if the teacher's velocity matches the marginal velocity of the mixture path. For a fixed pretrained teacher, that is not given, so the loss can be small while the teacher's ODE rollouts drift. This is not fatal, but it means 'trajectory-preserving' is currently a hypothesis supported by indirect evidence. Third: FID is 10-16 on FFHQ/CIFAR, clearly behind Rectified Flow and MeanFlow (2-5). The paper says sample quality is not the point, but that makes 'competitive sample quality' an overstatement. Minor: semantic modes are selected by CLIP coherence and then evaluated by CLIP coherence, so the validation is partly circular; no comparison against existing editing methods.\n\nWho this is for: anyone working on flow distillation, Koopman methods, or interpretability of generative models. The theoretical part is worth reading. The empirical claims need tightening rather than rejection.\n\nRecommendation: send to peer review. The core idea is novel and the theory holds; the missing pixel-space trajectory metric and the FID position are addressable in revision.","headline":"The consistency loss is a real step beyond endpoint distillation, but the paper's 'trajectory-preserving' claim currently rests on latent-space evidence; worth a serious review with a request for pixel-space trajectory metrics.","tokens_in":26569,"tokens_out":3096,"would_cite":true,"duration_ms":34374,"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":"The paper claims that a single linear operator in a learned lifted space reproduces the full non-linear trajectory of a pre-trained flow-matching model, so sampling becomes one step and the operator's spectrum becomes a tool for…","keywords":["Koopman operator theory","Continuous Normalizing Flows","Conditional Flow Matching","trajectory-preserving linearization","one-step generative sampling","spectral analysis of generative dynamics","simulation-free consistency loss","generative model inversion and editing"],"falsifier":"Run the identical pipeline on a teacher whose dynamics are known to have no finite-dimensional Koopman invariant subspace — for instance a chaotic or polynomial flow of the kind discussed in the literature the paper cites on finite-dimensional Koopman forms — and record the consistency-loss residual and trajectory MSE at a large latent dimension. If the residual stays bounded well away from zero while endpoints still match, the trajectory-preserving claim is refuted: the method would be fitting boundaries and only approximating interiors. A complementary test that applies to the paper's own models is to compare intermediate marginals: integrate the teacher to a time $t \\in (0,1)$ and compare the distribution of $x_t$ with one-step Koopman rollouts $\\exp(tL)$ from the same noises; agreement at $t=1$ with divergence in the middle would falsify full-trajectory linearization.","tokens_in":25458,"feed_emoji":"🌀","tokens_out":15379,"duration_ms":143245,"temperature":0.7,"pith_summary":"This paper tries to show that the non-linear, time-dependent dynamics of a pre-trained Continuous Normalizing Flow (CFM) can be replaced by a single fixed linear operator without losing the teacher's behavior along the way. Its claim is that a learned encoder lifts each state into a 1024-dimensional Koopman coordinate space in which the teacher's vector field becomes exactly the linear law $\\mathrm{d}z/\\mathrm{d}t = Lz$, matching the teacher infinitesimally at every point of the trajectory and not merely at its endpoints. If that holds, sampling collapses to one step — encode the noise, apply one matrix exponential, decode — and the eigenvalues and modes of $L$ become handles for analyzing, editing, and inverting the generative process. This matters because existing one-step distillation methods treat the teacher as a black box and discard its intermediate dynamics, whereas a trajectory-preserving linearization keeps those dynamics and makes them readable through linear algebra alone.","feed_headline":"One matrix exponential replaces the full generative sampling path","feed_subtitle":"A single linear operator reproduces a flow model's entire trajectory, enabling one-step sampling and spectral editing.","key_machinery":"The central object is a learned finite-dimensional Koopman representation: a diffusion-style SongUNet encoder–decoder pair that maps the data space to a 1026-dimensional lifted space (1024 observables plus time and a constant coordinate), together with a dense affine generator matrix $L$ whose constrained block structure forces $\\dot{t} = 1$ and $\\dot{1} = 0$. The load-bearing identity is the Koopman generator equation $Lg = \\nabla g \\cdot v_t$, adapted from the autonomous-system generator formula to non-autonomous flow matching; the consistency loss is exactly the square of this equation averaged over the marginal path distribution. The mechanism that makes the objective trainable is the unbiased-estimator identity of Proposition 3, which rewrites the expectation over the unknown marginal path as an expectation over data samples and conditional paths — the simulation-free trick that makes CFM itself trainable, applied here with the teacher's known marginal velocity. Finally, the real Schur decomposition of $L$ splits the flow into independent exponential and spiral modes, and the paper orders these modes by their eigenvalues to reveal a coarse-to-fine hierarchy in generation.","core_discovery":"The central discovery, stated on the paper's own terms, is that a trajectory-preserving linearization of a pre-trained Conditional Flow Matching model exists and is learnable. For a teacher with velocity field $v_t$, the paper constructs an encoder $g_\\phi$, a decoder $g_\\psi^{-1}$, and a fixed generator matrix $L$ such that the lifted coordinate $z_t = [1, t, g_\\phi(t, x_t)]$ evolves linearly, $\\mathrm{d}z/\\mathrm{d}t = Lz$, with the block structure of $L$ forcing time to advance at unit rate. The decisive ingredient is the consistency loss $L_{\\mathrm{cons}} = \\mathbb{E}\\,\\lVert L g(t,x_t) - \\nabla g(t,x_t)\\cdot v_t(x_t)\\rVert^2$, which is an unbiased, simulation-free estimate of the true marginal objective (Proposition 3); the paper proves that the naive conditional substitute carries a positive gap (Proposition 2) and that the decoder's reconstruction loss fixes the coordinate freedom left by linear gauge invariance (Proposition 1). With this loss, one-step rollouts $x_1 \\approx g^{-1}(\\exp(L)z_0)$ reproduce teacher trajectories to mean squared error near $5\\times10^{-6}$, whereas the same pipeline without the loss matches endpoints only, with error near $1.3\\times10^{-3}$. The paper further reports that the spectrum of $L$ then carries meaning: Schur modes arrange coarse-to-fine, single-mode perturbations yield coherent attribute edits (sunglasses coherence 0.97), and images can be inverted back to noise through $\\exp(-L)$.","pith_inferences":["Editorial inference: if real-world flow models admit such faithful finite-dimensional Koopman representations, then editing, interpolation, and inversion of generative models reduce to linear algebra in one fixed basis — a learned Fourier analysis of generation — and the eigenvalue ordering itself becomes a compressed description of how a dataset's images are produced.","Editorial inference: the consistency-loss residual is a measurable 'Koopman defect' per model, so it could be used as a diagnostic that ranks teachers or datasets by how far their dynamics are from linearizable; the paper builds this instrument but does not deploy it that way.","Editorial inference: a direct testable extension is to run the same linearization inside the latent space of a high-resolution diffusion transformer, where the operator dimension stays near 1024 while image resolution grows; the paper's quality gap on harder datasets would then either close, supporting the invariant-subspace story, or persist, pointing to the encoder as the bottleneck.","Editorial inference: the class-conditioned spectral signatures indicate that modes split into shared coarse structure and class-specific fine structure, suggesting a control scheme — steering class or identity by rescaling mode coefficients — that the paper does not itself propose."],"forward_implications":["Sampling from a CFM becomes one step — encode the noise, apply the matrix exponential once, decode — with the paper reporting FID 10.1 on FFHQ and 16.7 on CIFAR-10 at about 37 ms per image.","The linearization is not a boundary fit: Koopman rollouts track the teacher's full trajectories to mean squared error near $5\\times10^{-6}$ with the consistency loss, versus roughly $1.3\\times10^{-3}$ without it.","The spectrum of $L$ becomes a control surface: moving along a single Koopman mode edits attributes (CLIP coherence 0.97 for sunglasses, 0.94 for brown hair), and discovered directions transfer back into the original CFM's noise space.","Inversion becomes a matrix exponential: $\\exp(-L)$ sends images back to the teacher's noise space, which the paper uses for reconstruction and for downstream inpainting, super-resolution, and denoising.","Because the consistency estimator needs only the teacher's velocity and marginals, the same linearization extends to any flow with an evaluable marginal velocity, including diffusion models via their probability-flow ODE."],"supporting_citations":[{"why":"Foundational source of the operator-on-observables construction; the premise that non-linear dynamics become linear when lifted.","marker":"Koopman, 1931"},{"why":"Supplies the generator identity Lg = ∇g·v and the invariant-subspace framework that the consistency loss instantiates.","marker":"Brunton et al., 2022"},{"why":"Defines Neural ODEs and Continuous Normalizing Flows, the dynamical system the paper linearizes.","marker":"Chen et al., 2018"},{"why":"Introduces the Conditional Flow Matching objective whose simulation-free sampling strategy the consistency estimator mirrors.","marker":"Lipman et al., 2023"},{"why":"Provides the OT-CFM teacher used in experiments and the conditional/marginal path formalism behind Propositions 2 and 3.","marker":"Tong et al., 2024"},{"why":"Closest prior work, a Koopman-based one-step distillation of diffusion; the no-consistency ablation reduces to it and isolates the paper's new infinitesimal-consistency term.","marker":"Berman et al., 2025"},{"why":"Rectified Flow, the trajectory-straightening baseline that motivates the contrast between distilling endpoints and linearizing full dynamics.","marker":"Liu et al., 2023"},{"why":"MeanFlow, the one-step generative baseline compared in the FID tables.","marker":"Geng et al., 2025"},{"why":"The cited result that finite-dimensional Koopman invariant subspaces need not exist; it anchors the paper's own caveat on its strongest assumption.","marker":"Iacob et al., 2023"},{"why":"Probability-flow ODE formulation cited to support the claim that the linearization extends beyond flow matching to diffusion models.","marker":"Song et al., 2021"}],"fun_headline_variants":["Flow models go linear: one operator, one-step sampling","Koopman lens turns generative flows into a single matrix","One-step sampling via Koopman-linearized flows","Trajectory-preserving linearization unlocks spectral editing of flows"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that the teacher's flow admits a finite-dimensional invariant Koopman subspace — that roughly 1024 learned coordinates are enough to make the teacher's dynamics exactly linear. The paper itself flags this as a strong condition that need not hold for arbitrary non-linear flows; if it fails, the consistency loss cannot approach zero and the linearization is only approximate.","fun_headline_variants_meta":{"raw":{"variants":["Flow models go linear: one operator, one-step sampling","Koopman lens turns generative flows into a single matrix","One-step sampling via Koopman-linearized flows","Trajectory-preserving linearization unlocks spectral editing of flows"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000281,"raw_usage":{"total_tokens":1755,"prompt_tokens":1124,"completion_tokens":631,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":740,"completion_tokens_details":{"reasoning_tokens":564}},"tokens_in":740,"tokens_out":631,"duration_ms":7255,"temperature":1.0,"reasoning_tokens":564,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:07:22.270836+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical pipeline on a teacher whose dynamics are known to have no finite-dimensional Koopman invariant subspace — for instance a chaotic or polynomial flow of the kind discussed in the literature the paper cites on finite-dimensional Koopman forms — and record the consistency-loss residual and trajectory MSE at a large latent dimension. If the residual stays bounded well away from zero while endpoints still match, the trajectory-preserving claim is refuted: the method would be fitting boundaries and only approximating interiors. A complementary test that applies to the paper's own models is to compare intermediate marginals: integrate the teacher to a time $t \\in (0,1)$ and compare the distribution of $x_t$ with one-step Koopman rollouts $\\exp(tL)$ from the same noises; agreement at $t=1$ with divergence in the middle would falsify full-trajectory linearization.","supporting_citations":[{"cited_title":"Hamiltonian systems and transformation in hilbert space","cited_arxiv_id":null,"evidence_quote":"Foundational source of the operator-on-observables construction; the premise that non-linear dynamics become linear when lifted."},{"cited_title":"Neural ordinary differential equations","cited_arxiv_id":null,"evidence_quote":"Defines Neural ODEs and Continuous Normalizing Flows, the dynamical system the paper linearizes."},{"cited_title":"Flow matching for generative modeling","cited_arxiv_id":null,"evidence_quote":"Introduces the Conditional Flow Matching objective whose simulation-free sampling strategy the consistency estimator mirrors."},{"cited_title":"One-step offline distillation of diffusion-based models via koopman modeling","cited_arxiv_id":null,"evidence_quote":"Closest prior work, a Koopman-based one-step distillation of diffusion; the no-consistency ablation reduces to it and isolates the paper's new infinitesimal-consistency term."}],"review_version":1}