Pith. sign in

REVIEW 3 major objections 3 minor 42 references

Few-step integration error in generative ODEs is injected where the model's velocity or prediction field varies most, then transported far from that source by the learned dynamics; a signed propagation of per-step truncation residuals recon

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 21:56 UTC pith:RZJI2PZ6

load-bearing objection A careful, well-scoped empirical study: per-region integration error is predicted by Flow Complexity and is transported nonlocally, but the reconstruction's reference-path assumption needs a divergence check before the strongest claims fully land. the 3 major comments →

arxiv 2607.16361 v1 pith:RZJI2PZ6 submitted 2026-07-17 cs.CV

Spatial Transport of Integration Error in Generative ODEs

Songheng Yin This is my paper
classification cs.CV
keywords generative ODEsflow matchingdiffusion modelsintegration errorspatial error transportFlow Complexityerror predictionODE solvers
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish where integration error enters when a trained flow or diffusion model is run with few solver steps, and how that error reaches the final sample. It claims the error is not noise: each image has a stable, learned spatial difficulty landscape, concentrated in a few regions and reproducible across budgets and retraining, and this landscape can be read from the model's own velocity or prediction field. The reading is only partial because error is injected locally but then carried far by the model's dynamics; a signed accounting of per-step truncation residuals propagated through the linearized dynamics reconstructs the endpoint error's direction and regional structure. If correct, this separates 'where error is injected' from 'where it lands' and shows both are readable from the model, with the injection landscape modifiable by a training penalty.

Core claim

The paper's central claim is that few-step integration error in generative ODEs has a learned, spatially structured source-and-transport story. A per-region statistic, Flow Complexity, defined as the accumulated variation of the model's velocity (or prediction) field along a fine reference trajectory, bounds the truncation the solver injects in each region, and a leading-order bound (Proposition 1) relates final per-region error to Flow Complexity values carried through the state-transition operator of the linearized dynamics. Empirically, Flow Complexity predicts per-region error within images at Spearman rho 0.57-0.70 on five 256px models, but only partially: a perturbation protocol shows

What carries the argument

The argument rests on two objects. Flow Complexity (FC) is a per-region total variation: for flow models it is the sum of the magnitudes of the learned velocity field's changes along the trajectory, and for diffusion models the corresponding variation of the prediction field. FC is the quantity the Euler truncation bound ties to local injection: each region's accumulated variation bounds the magnitude of the local truncation residual it injects. The second object is the state-transition operator Phi of the linearized dynamics, which carries each injected residual to the endpoint; Proposition 1 bounds final per-region error by a Phi-weighted sum of all regions' FC values. The reconstruction p

Load-bearing premise

The reconstruction's per-step truncation residuals are evaluated along the near-exact reference trajectory, not along the path the cheap solver actually visits; if the cheap trajectory diverges from the reference, the residuals injected during real sampling differ from the ones being propagated.

What would settle it

Measure the divergence between an NFE-6 Euler path and the fine reference path on the paper's own models, and recompute the reconstruction with residuals evaluated along the cheap path instead of the reference path: if the endpoint cosine (0.81-0.87) and per-region rho (0.80-0.82) drop to the no-propagation level, the transport account describes reference-path error rather than the error experienced by the actual solve.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the paper is right, per-region difficulty maps are not noise: they are stable, model-specific structures that emerge during training, so a solver or sampler could treat them as part of the model's signature.
  • Early-window Flow Complexity, even computed only from the first three steps of a cheap solve, localizes a meaningful share of final error (about a quarter, against a 14.8% random baseline in the paper's protocol-frozen tests), so adaptive refinement can in principle start before the trajectory is complete.
  • Because a region's final error is dominated by error transported in from other regions, improving few-step sampling by refining only the locally hardest regions will be incomplete; an allocation rule must account for propagation through the linearized dynamics.
  • A training-time penalty on Flow Complexity reduces both injected error and coarse-solve error and improves few-step FID at matched optimization steps (57-65% integration-error reduction; about 7% FID gain at NFE 6 on ImageNet), so the difficulty landscape is something training can change.
  • The paper's own compute-matched analysis shows the training penalty wins at moderate scale but not at ImageNet scale when the extra forward per step is charged, so the mechanism's production value remains open.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial: If injected error is transported nonlocally, the natural next test is an adaptive step allocator that assigns budget to projected 'landing' regions (computed from a cheap propagated-residual estimate) rather than to local FC; the paper does not offer such a sampler, but its transport kernel suggests it.
  • Editorial: The direction-, sign-, and timing-dependence of the reconstruction suggests that a penalty on signed cumulative variation or a horizon-weighted variant might straighten trajectories more efficiently than the magnitude penalty the paper tests; the paper reports such variants predict error as well as or better than FC, so a training objective built on them is a plausible extension.
  • Editorial: The prediction horizon is model-dependent (early for flow models, late for diffusion), which suggests that 'when difficulty becomes readable' is itself learned; an online probe that detects the error-determining window during sampling could turn the diagnostic into a stopping rule, though the paper does not establish that the cheap-solve signal is strong enough to drive such a rule.
  • Editorial: The paper's finding that refining the lowest-FC regions improves FID while refining the highest-FC regions reduces integration error implies numerical integrability and perceptual quality are different objectives; a testable extension is to measure whether propagation-aware correction changes FID differently than injection-aware correction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper studies where integration error is injected and how it spreads when a trained generative ODE is solved with very few steps. It defines a per-region diagnostic, Flow Complexity (FC), as the total variation of the model's velocity/prediction field along the trajectory, and shows that FC predicts the per-region reference-solver discrepancy (cheap solve vs. near-exact solve) with within-image Spearman rho 0.57–0.70 across five 256px models. It then uses perturbation experiments and a signed-residual reconstruction (e ≈ Σ Φ τ_k) to argue that endpoint error is mostly transported from other regions rather than locally injected, with off-diagonal contributions carrying more association than a region's own injection. A training penalty on FC lowers reference-solver discrepancy and improves few-step FID at matched optimization steps on the authors' own models, though not at matched training compute on ImageNet. The paper is notably explicit about its limitations and scope.

Significance. If the central transport claim holds, the paper is a genuinely new spatial account of numerical error in generative ODEs: it separates where error is injected from where it lands, shows that the source can be read from the model itself, and demonstrates that the source structure can be modified by training. The paper's strengths include convergence checks (rank correlation ≥0.986 against finer references), image-level bootstrap CIs, protocol-frozen nulls, matched-budget selector comparisons, and explicit acknowledgment of single-run and compute-matched caveats. The empirical apparatus is unusually careful, and the FC landscape result alone is a useful contribution. However, the load-bearing reconstruction and the diagonal-vs-transported contrast rest on a reference-path linearization whose validity is not demonstrated against actual cheap-trajectory divergence.

major comments (3)
  1. [A.2, Eq. (4)–(5)] The reconstruction defines the signed per-step truncation τ_k as the teacher-forced Euler defect evaluated at the reference state x_ref_k, and propagates it via linearized dynamics evaluated at x_ref_{k+1}. The target is the actual error of the cheap Euler solve, but the manuscript never reports how far the cheap trajectory diverges from the reference trajectory, nor the magnitude of the neglected nonlinear remainder. If the cheap trajectory diverges substantially over NFE 4–6, τ_k can differ from the defect actually injected during real sampling, and the 0.81–0.87 endpoint cosines and the diagonal-vs-off-diagonal contrast could be properties of the reference path rather than of the true error-injection process. The manuscript itself lists "evaluation along the reference rather than the perturbed trajectory" as a residual source (A.2), but gives no quantitative estimate. Please report pe
  2. [§6 / Fig. 6B] The headline contrast — own injection ρ≈0.33–0.42, transported-in partial ρ≈0.66–0.69 — is computed entirely from the same reference-path decomposition described above. The source-count, sign-randomization, and destination-permutation nulls are well designed, but they all operate on the reference-path objects. If those objects are not faithful to the actual cheap trajectory, the contrast may be a linearization artifact rather than evidence of real transport. The paper should provide a check in which the actual cheap trajectory is used, or at least show that the reconstruction's conclusions are stable when the propagation path is perturbed by the observed level of cheap-reference divergence. Absent that, the mechanistic interpretation of the partial correlations is under-supported.
  3. [A.2 ("first-order closure")] The paper says the decomposition is "only worth the name if it holds quantitatively," and uses "closes" language, yet the median relative error at the operational NFE=6 budget is 0.65 on CelebA and 0.50 on ImageNet. The cosines of 0.81–0.87 show that the first-order reconstruction captures direction and regional ordering, but roughly half to two-thirds of the error vector remains unexplained. Please state this directly in §6 as well as in A.2, and either decompose the residual (nonlinear defect interactions, finite-difference error, reference-solver error, reference-path mismatch) or soften the "closes" wording. This is relevant because the off-diagonal partial correlation is presented as the quantitative foundation of the main conclusion.
minor comments (3)
  1. [§6 / Fig. 6B / A.2] The ImageNet reconstruction sample size is inconsistent: §6 and Fig. 6B say n=24, while A.2 first reports n=32 before later switching to n=24. Please harmonize the reported n and update the CIs if needed.
  2. [Fig. 7A] The legend reads "CelebA DiT (flow, latent)" but DiT is described elsewhere as a diffusion model and SiT as the flow model; this is likely a typo for the authors' FM-DiT model. Please correct the legend to avoid confusion.
  3. [B.11] The reproducibility section states that code and configs "will be released upon publication." Given the number of committed artifacts (results/sit/, results/celebahq/, nulls, baselines), a public repository link or archive with the analysis code would materially strengthen the paper.

Circularity Check

0 steps flagged

No significant circularity: FC-error correlation is empirical, and the residual reconstruction is explicitly a post-hoc first-order check with null controls.

full rationale

The paper's two central chains are not circular. First, Flow Complexity is defined independently of the target: FC(i)=Σ||u_{t+Δt}(i)-u_t(i)|| (Def. 1), while the target is the per-region reference-solver discrepancy ||e_i||^2 from a cheap solve vs. a near-exact reference. The link between them is a leading-order bound (Prop. 1), not an equality, and the paper explicitly says 'a bound is not a ranking' and that the ρ values are empirical, not derived. No parameter is fitted to the error to produce FC, and the paper reports held-out confirmations for the cheap-trajectory variant. Second, the reconstruction \hat e=Σ Φ τ_k (Appendix A.2) is the standard first-order global-error expansion: τ_k is the teacher-forced Euler defect at reference states (Eq. 4), Φ is estimated by finite differences of the near-exact solver along those residuals (Eq. 5), and e=x_euler_T−x_ref_T is the actual cheap-solve discrepancy. This is a post-hoc consistency test, not an independent prediction; the paper labels it 'post-hoc mechanistic account' and identifies its non-circular content as the diagonal-vs-transported contrast and the sign/destination/time null ladder. The high cosine therefore does not function as a fitted prediction, and the transport conclusion is supported by the nulls (destination permutation destroys the reconstruction) and by the matched-horizon perturbation controls. The unmeasured cheap-vs-reference trajectory divergence noted by the skeptic is a validity/robustness concern about the first-order remainder, not a demonstration that any equation reduces by definition to its own input. No self-citation chain is load-bearing.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 1 invented entities

The central claims rest mostly on measurement choices and empirical protocol, not on a free-parameter fit: FC is computed directly from the model, and the reconstruction uses true residuals plus the model's Jacobian. The main data-dependent numbers are lambda, epsilon, and the early-prefix k; the main domain assumptions are smoothness, reference-solve fidelity, and the reference-state residual definition. No new physical entities are postulated; FC is a new statistic with independent predictive evidence.

free parameters (3)
  • lambda (FCM penalty weight) = 0.3
    Selected by a controlled CIFAR-32 lambda sweep (Appendix B.5); used for all FCM training arms. Not fitted to the ImageNet/CelebA target, but a hand-chosen hyperparameter the intervention results depend on.
  • epsilon (perturbation scale) = 0.05
    Chosen for the central-difference estimate of Phi tau_k in Eq. (5); checked to lie in a linear-response regime (factor-of-4 linearity), but the reconstruction cosine depends on this choice.
  • early-prefix k = 2 (chosen on exploratory seed, then frozen)
    The early-localization claim uses the first two cheap velocity evaluations; the prefix was selected after inspecting an exploratory seed, then frozen for confirmatory seeds. It is a data-dependent protocol choice disclosed by the authors.
axioms (5)
  • domain assumption Learned drift smoothness: u in C^2 with bounded derivatives / L-Lipschitz, so Euler truncation is O(H^3) and the variational linearization is valid.
    Used in Proposition 1 and the reconstruction (Appendix A.1); plausible for trained networks but not proved.
  • domain assumption Near-exact reference solve (Heun-200 or fine DDIM) stands in for the exact endpoint.
    The discrepancy target e = x_cheap - x_ref is computed against finite-step references; convergence is checked by rank-correlation invariance under refinement (Section 3, B.2), but the reference is not the true endpoint.
  • domain assumption Per-step truncation residual tau_k evaluated at the reference state equals the error the cheap solver actually injects.
    Eq. (4) in Appendix A.2 teacher-forces the Euler step at reference states; if the cheap trajectory diverges from the reference, the injected residual differs. No divergence measurement is reported.
  • domain assumption The state-transition operator estimated by central finite differences along the reference trajectory approximates the true variational operator.
    Central to the reconstruction of Section 6 and A.2; linearity is checked over a factor of four in epsilon, but the Jacobian is evaluated on the reference path, not the actual cheap path.
  • standard math Gronwall's inequality and standard Euler/Taylor remainder bounds.
    Used to derive Proposition 1 in Appendix A.1; standard numerical-analysis background.
invented entities (1)
  • Flow Complexity (FC) independent evidence
    purpose: Per-region diagnostic statistic: total variation of the velocity/prediction field along a trajectory window; intended to measure local truncation injection and predict per-region integration error.
    FC is a new quantity with a falsifiable handle: it predicts per-region reference-solver discrepancy on held-out images and external models (Section 4.4), so its predictive content is testable outside the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 21638 in / 13749 out tokens · 150675 ms · 2026-08-01T21:56:44.670628+00:00 · methodology

0 comments
read the original abstract

A trained flow or diffusion model is usually run with only a handful of solver steps, and the integration error this leaves behind is unevenly distributed across the image. We ask where that error is injected and how it reaches the endpoint, and answer with a signed source-and-transport accounting of few-step integration error, tested to first order. A perturbation experiment on five models at 256^2 resolution shows the learned dynamics spread local disturbances widely: near the start of sampling, under 10% of the summed endpoint response remains at the source. Signed one-step truncation residuals, propagated through the model's own linearized dynamics, reconstruct much of the endpoint error's direction and regional structure (cosine 0.81-0.87), and a region's error owes more to what arrives from elsewhere than to its own injection. Structure-destroying nulls, with protocols frozen before evaluation, locate what carries the account: randomizing contribution signs halves it, and reassigning which region receives each contribution, with content, norms, and signs intact, destroys it entirely. Where the injections land is readable from the model itself. The variation of its velocity or prediction field along the trajectory, a structure that emerges during training, predicts the final per-region gap (within-image rho of 0.57-0.70 on fine trajectories, weaker from the cheap solve alone). The prediction is partial because endpoint error depends not only on injected magnitude but on its sign, timing, and transport through the learned dynamics. A training penalty on the injected variation lowers few-step error, so the structure is one a model can be trained to change.

Figures

Figures reproduced from arXiv: 2607.16361 by Songheng Yin.

Figure 1
Figure 1. Figure 1: The spatial integrability landscape. For SiT-XL/2 ImageNet-256 samples: the gener￾ated image, the model-intrinsic Flow Complexity map (computed from the model’s own velocities, without seeing the final error), and the per-region reference-solver discrepancy (cheap vs. near-exact solve). FC concentrates where the cheap solver actually errs (within-image Spearman ρ per row), on real 16 × 16 regions. Both are… view at source ↗
Figure 2
Figure 2. Figure 2: The landscape is concentrated. Lorenz curves of per-region reference-solver discrepancy (cumulative error share vs. fraction of regions, hardest first) for the three external models; the dashed diagonal is uniform (no concentration). The top decile of regions carries 24–26% of the error across models, 2.4–2.6× a uniform allocation, a moderate but reproducible concentration. 4.2 Stable The landscape is a pr… view at source ↗
Figure 3
Figure 3. Figure 3: The difficulty landscape is learned and stable (controlled CIFAR model; early￾complexity map; within-image Spearman on shared noise). (A) agreement between the map at training step t and a converged reference. Measured against an independently seeded run’s final map (orange, no self-correlation), the map is uninformative at initialization and rises to the cross￾seed ceiling ρ≈0.90 (dashed), so the identity… view at source ↗
Figure 4
Figure 4. Figure 4: FC predicts error for the typical image, not just on average. Per-image within-image Spearman ρ between FC and the per-region reference-solver discrepancy (violins; white bar = me￾dian) for the three external models. Medians (0.62–0.73) exceed the means and only 0–5% of images fall below ρ = 0 (dashed), so the aggregate is not carried by a few high-correlation images. 6 [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 5
Figure 5. Figure 5: Prediction horizon. Within-image correlation between early-window FC (observed up to cutoff f) and the final per-region integration error, for SiT (flow), ADM (diffusion, pixel), and DiT (diffusion, latent). Every model exposes a predictive window (open circles mark each peak); its location depends on the model, early for the flow model (already ρ = 0.44 at f = 0.1) and later for the diffusion models. Wher… view at source ↗
Figure 6
Figure 6. Figure 6: Reconstructing endpoint error from signed truncation residuals (our own 256px mod￾els; post-hoc, using the true residuals and the model’s Jacobian). (A) the endpoint error vector, re￾built by summing the residuals with vs. without propagation through the linearized dynamics: propa￾gation lifts the cosine from 0.71/0.77 to 0.81/0.87. (B) per-region error, within-image ρ against the actual per-region error: … view at source ↗
Figure 7
Figure 7. Figure 7: Integration error is spatially transported; its apparent temporal localization is a hori￾zon effect. A region is perturbed at trajectory time t ∗ and the endpoint response is read per region. (A) to the endpoint, the self-retained response rises toward data on a flow transformer (SiT-XL), a diffusion U-Net (ADM, pixel), and our own model. (B) under a matched rollout horizon (∆ = 0.1) the trend flattens or … view at source ↗
Figure 8
Figure 8. Figure 8: Minimizing Flow Complexity improves few-step FID at matched optimization steps (256px). FID (50k samples, cfg = 1.0) vs. sampling budget (NFE), baseline vs. FCM (λ = 0.3), trained from scratch in lockstep on class-conditional ImageNet-256 (left) and CelebA-HQ-256 (right). The gain is largest in the few-step regime and present at every NFE we test, with no high￾NFE cost. The compute-matched verdict differs … view at source ↗
Figure 9
Figure 9. Figure 9: The penalty’s benefit lands on the difficult landscape. Regions binned by baseline FC percentile (CelebA, n=32768 regions, λ=0.3 vs. baseline at matched 100k steps). (A) the penalty lowers FC by a near-uniform 3.4–4.8% across every difficulty bin. (B) the resulting absolute integration-error reduction climbs about 7.5× from the easy half to the hardest 5%, concentrating the improvement on the high-FC regio… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 7 linked inside Pith

  1. [1]

    European Conference on Computer Vision (ECCV) , year=

    SiT: Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformers , author=. European Conference on Computer Vision (ECCV) , year=

  2. [2]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Diffusion Models Beat GANs on Image Synthesis , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  3. [3]

    International Conference on Learning Representations (ICLR) , year=

    Score-Based Generative Modeling through Stochastic Differential Equations , author=. International Conference on Learning Representations (ICLR) , year=

  4. [4]

    International Conference on Learning Representations (ICLR) , year=

    Denoising Diffusion Implicit Models , author=. International Conference on Learning Representations (ICLR) , year=

  5. [5]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  6. [6]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  7. [7]

    International Conference on Learning Representations (ICLR) , year=

    Flow Matching for Generative Modeling , author=. International Conference on Learning Representations (ICLR) , year=

  8. [8]

    International Conference on Learning Representations (ICLR) , year=

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author=. International Conference on Learning Representations (ICLR) , year=

  9. [9]

    International Conference on Learning Representations (ICLR) , year=

    Building Normalizing Flows with Stochastic Interpolants , author=. International Conference on Learning Representations (ICLR) , year=

  10. [10]

    IEEE/CVF International Conference on Computer Vision (ICCV) , year=

    Scalable Diffusion Models with Transformers , author=. IEEE/CVF International Conference on Computer Vision (ICCV) , year=

  11. [11]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Denoising Diffusion Probabilistic Models , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  12. [12]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Elucidating the Design Space of Diffusion-Based Generative Models , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  13. [13]

    International Conference on Machine Learning (ICML) , year=

    Consistency Models , author=. International Conference on Machine Learning (ICML) , year=

  14. [14]

    International Conference on Learning Representations (ICLR) , year=

    Progressive Distillation for Fast Sampling of Diffusion Models , author=. International Conference on Learning Representations (ICLR) , year=

  15. [15]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    DeepCache: Accelerating Diffusion Models for Free , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  16. [16]

    arXiv preprint arXiv:2404.02258 , year=

    Mixture-of-Depths: Dynamically Allocating Compute in Transformer-Based Language Models , author=. arXiv preprint arXiv:2404.02258 , year=

    Pith/arXiv arXiv

  17. [17]

    International Conference on Machine Learning (ICML) , year=

    Minimizing Trajectory Curvature of ODE-based Generative Models , author=. International Conference on Machine Learning (ICML) , year=

  18. [20]

    IEEE/CVF International Conference on Computer Vision (ICCV) , year=

    Efficient Diffusion Training via Min-SNR Weighting Strategy , author=. IEEE/CVF International Conference on Computer Vision (ICCV) , year=

  19. [22]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Region-Adaptive Sampling for Diffusion Transformers , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  20. [26]

    arXiv preprint arXiv:2603.06454 , year=

    Training Flow Matching: The Role of Weighting and Parameterization , author=. arXiv preprint arXiv:2603.06454 , year=

    Pith/arXiv arXiv

  21. [27]

    Solving Ordinary Differential Equations I: Nonstiff Problems , author=

  22. [28]

    Albergo and Eric Vanden-Eijnden

    Michael S. Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In International Conference on Learning Representations (ICLR), 2023

    2023

  23. [29]

    Foveated diffusion: Efficient spatially adaptive image and video generation

    Brian Chao, Lior Yariv, Howard Xiao, and Gordon Wetzstein. Foveated diffusion: Efficient spatially adaptive image and video generation. arXiv preprint arXiv:2603.23491, 2026

    arXiv 2026

  24. [30]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  25. [31]

    N rsett, and Gerhard Wanner

    Ernst Hairer, Syvert P. N rsett, and Gerhard Wanner. Solving Ordinary Differential Equations I: Nonstiff Problems. Springer, 2nd edition, 1993

    1993

  26. [32]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  27. [33]

    Training-free mixed-resolution latent upsampling for spatially accelerated diffusion transformers

    Wongi Jeong, Kyungryeol Lee, Hoigi Seo, and Se Young Chun. Training-free mixed-resolution latent upsampling for spatially accelerated diffusion transformers. arXiv preprint arXiv:2507.08422, 2025

    arXiv 2025

  28. [34]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  29. [35]

    Isokinetic flow matching for pathwise straightening of generative flows

    Tauhid Khan. Isokinetic flow matching for pathwise straightening of generative flows. arXiv preprint arXiv:2604.04491, 2026

    Pith/arXiv arXiv 2026

  30. [36]

    Minimizing trajectory curvature of ode-based generative models

    Sangyun Lee, Beomsu Kim, and Jong Chul Ye. Minimizing trajectory curvature of ode-based generative models. In International Conference on Machine Learning (ICML), 2023. arXiv:2301.12003

    Pith/arXiv arXiv 2023

  31. [37]

    Sdit: Semantic region-adaptive for diffusion transformers

    Bowen Lin, Fanjiang Ye, Yihua Liu, Zhenghui Guo, Boyuan Zhang, Weijian Zheng, Yufan Xu, Tiancheng Xing, Yuke Wang, and Chengming Zhang. Sdit: Semantic region-adaptive for diffusion transformers. arXiv preprint arXiv:2601.12283, 2026. arXiv:2601.12283

    arXiv 2026

  32. [38]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. In International Conference on Learning Representations (ICLR), 2023

    2023

  33. [39]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In International Conference on Learning Representations (ICLR), 2023

    2023

  34. [40]

    Region-adaptive sampling for diffusion transformers

    Ziming Liu, Yifan Yang, Chengruidong Zhang, Yiqi Zhang, Lili Qiu, Yang You, and Yuqing Yang. Region-adaptive sampling for diffusion transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026. arXiv:2502.10389

    arXiv 2026

  35. [41]

    Curveflow: Curvature-guided flow matching for image generation

    Yan Luo, Drake Du, Hao Huang, Yi Fang, and Mengyu Wang. Curveflow: Curvature-guided flow matching for image generation. arXiv preprint arXiv:2508.15093, 2025

    Pith/arXiv arXiv 2025

  36. [42]

    Albergo, Nicholas M

    Nanye Ma, Mark Goldstein, Michael S. Albergo, Nicholas M. Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In European Conference on Computer Vision (ECCV), 2024. arXiv:2401.08740

    Pith/arXiv arXiv 2024

  37. [43]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    2023

  38. [44]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations (ICLR), 2022

    2022

  39. [45]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021 a

    2021

  40. [46]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021 b

    2021

  41. [47]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In International Conference on Machine Learning (ICML), 2023

    2023

  42. [48]

    Consistency flow matching: Defining straight flows with velocity consistency

    Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining straight flows with velocity consistency. arXiv preprint arXiv:2407.02398, 2024

    Pith/arXiv arXiv 2024