pith. machine review for the scientific record. sign in

arxiv: 2605.11547 · v1 · submitted 2026-05-12 · 💻 cs.LG · cs.AI

Recognition: 2 theorem links

· Lean Theorem

Sharpen Your Flow: Sharpness-Aware Sampling for Flow Matching

Aditi Gupta , Soon Hoe Lim , Annan Yu , N. Benjamin Erichson
Authors on Pith no claims yet

Pith reviewed 2026-05-13 02:16 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords flow matchingsharpness-aware samplingEuler integrationnon-uniform timesteppinggenerative modelsdiscretization errorsample quality
0
0 comments X

The pith

A sharpness profile from offline finite differences lets non-uniform Euler steps improve flow matching sample quality at any fixed budget.

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

The paper introduces SharpEuler, a sampler that spends integration steps more densely where a pretrained flow model's velocity field changes rapidly. It estimates this sharpness offline via finite differences along calibration trajectories, smooths the profile, and converts it through a quantile transform into a non-uniform timestep grid. At inference the method performs ordinary Euler integration but on this adaptive schedule, using exactly the same number of model calls as a uniform grid. Justification rests on a numerical link between trajectory acceleration and discretization error, a variational derivation of power-law step densities, and a statistical guarantee of stability at the terminal distribution. Experiments indicate gains in mode coverage and reduced inter-mode leakage.

Core claim

SharpEuler constructs a solver-aware sharpness profile by finite-difference estimation of velocity-field changes along calibration trajectories, applies smoothing and a quantile transform to obtain a timestep grid for any chosen budget, and demonstrates that Euler integration on this grid produces higher-quality samples than uniform spacing while preserving the same evaluation count.

What carries the argument

The solver-aware sharpness profile: a smoothed finite-difference estimate of velocity acceleration along calibration paths, quantile-transformed into a non-uniform timestep schedule.

If this is right

  • Sample quality improves at fixed budgets through reduced inter-mode leakage and increased mode coverage.
  • The sampler remains training-free and works on any pretrained flow matching model.
  • The quantile transform accommodates arbitrary inference budgets while keeping the same total number of model evaluations.
  • Numerical, variational, and statistical principles together ensure the non-uniform schedule is stable at the terminal distribution.

Where Pith is reading between the lines

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

  • The same offline profiling approach could be tested with higher-order integrators or adaptive step-size controllers during sampling.
  • Calibration trajectories might need to be drawn from a distribution closer to the test-time starting measure for more complex data.
  • Analogous sharpness estimation could be applied to diffusion or other continuous-time generative models that rely on numerical integration.

Load-bearing premise

The sharpness profile estimated offline via finite differences on calibration trajectories accurately identifies regions of high discretization error and generalizes to the actual sampling trajectories at test time.

What would settle it

If samples generated with the sharpness-derived timestep grid show no improvement over uniform-grid samples in quality metrics such as mode coverage or inter-mode leakage on held-out data, the claimed benefit is refuted.

read the original abstract

Flow matching models generate samples by numerically integrating a learned velocity field, with each integration step requiring a neural network evaluation. Fast generation therefore requires using a small fixed evaluation budget effectively: the key question is not only how to integrate the flow, but where the sampler should spend its steps. We propose SharpEuler, a training-free sampler that profiles a pretrained model offline by estimating where the learned velocity field changes most rapidly along calibration trajectories. This finite-difference estimate defines a solver-aware sharpness profile, which is smoothed and converted by a quantile transform into a timestep grid for any desired inference budget. At test time, sampling remains ordinary Euler integration with the same number of model evaluations as a uniform schedule. We justify SharpEuler using three principles: a numerical principle identifying trajectory acceleration as the leading source of Euler discretization error, a variational principle deriving sharpness-based power-law timestep densities, and a statistical guarantee showing that the finite-sample calibrated sampler is stable at the terminal distribution level. Our experiments show that SharpEuler improves sample quality at fixed budgets, reducing inter-mode leakage and increasing mode coverage.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The paper proposes SharpEuler, a training-free sampler for flow matching models. It profiles a pretrained velocity field offline via finite-difference estimates of sharpness (trajectory acceleration) along uniform calibration trajectories, smooths the profile, and applies a quantile transform to derive a non-uniform timestep grid for any target number of Euler steps. Sampling at test time uses standard Euler integration on this grid. The method is justified by a numerical argument that acceleration dominates Euler truncation error, a variational derivation of power-law timestep densities from sharpness, and a statistical stability guarantee at the terminal measure. Experiments claim improved sample quality, reduced inter-mode leakage, and better mode coverage at fixed budgets.

Significance. If the central claims hold, SharpEuler provides a practical, model-agnostic way to allocate a fixed inference budget more effectively in flow matching and related ODE-based generative models. The combination of an offline sharpness profile with a quantile-based grid offers a concrete, reproducible improvement over uniform schedules without retraining or architectural changes. The three-principle justification (numerical, variational, statistical) and the explicit handling of discretization error sources are strengths that could influence sampler design beyond flow matching.

major comments (3)
  1. [§4 (statistical guarantee) and §3.2 (quantile transform)] The statistical stability guarantee (abstract and §4) rests on the claim that the offline sharpness profile estimated on uniform calibration trajectories remains representative under the non-uniform schedule produced by the quantile transform. Because the learned velocity field is nonlinear, the locations of high curvature can shift when local step sizes change; the manuscript provides no direct verification (e.g., comparison of sharpness profiles or integrated error on the final vs. calibration trajectories) that this invariance holds at the budgets used in experiments.
  2. [Table 2, Figure 4, and §5.2] Table 2 and Figure 4 report gains in FID and mode coverage, but the ablation isolating the effect of the sharpness-derived grid versus a simple non-uniform schedule (e.g., linear or exponential) is missing. Without this control, it is unclear whether the observed reduction in inter-mode leakage is attributable to the sharpness profile or to any non-uniform allocation.
  3. [§3.1 (numerical principle) and Eq. (7)] The numerical principle (§3.1) identifies trajectory acceleration as the dominant Euler truncation source and uses finite differences on calibration paths. However, the finite-difference stencil and smoothing parameter are treated as fixed hyperparameters; no sensitivity analysis shows how variation in these choices affects the final timestep grid or sample quality, which is load-bearing for the claim of a “solver-aware” profile.
minor comments (3)
  1. [§3.2] Notation for the sharpness profile S(t) and the quantile mapping Q(·) is introduced without an explicit equation linking them to the final timestep grid; adding a single displayed equation would improve clarity.
  2. [Related Work] The manuscript cites prior work on adaptive step-size methods for ODEs but does not discuss why those adaptive schemes were not used as baselines; a short paragraph contrasting offline quantile allocation with online error-estimate adaptation would strengthen the positioning.
  3. [Figure 3] Figure 3 (sharpness profiles) would benefit from error bars or multiple calibration runs to indicate variability of the finite-difference estimate.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which help clarify the presentation and strengthen the empirical support for SharpEuler. We address each major comment below, indicating the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: [§4 (statistical guarantee) and §3.2 (quantile transform)] The statistical stability guarantee (abstract and §4) rests on the claim that the offline sharpness profile estimated on uniform calibration trajectories remains representative under the non-uniform schedule produced by the quantile transform. Because the learned velocity field is nonlinear, the locations of high curvature can shift when local step sizes change; the manuscript provides no direct verification (e.g., comparison of sharpness profiles or integrated error on the final vs. calibration trajectories) that this invariance holds at the budgets used in experiments.

    Authors: We agree that explicit verification of profile invariance under the quantile-derived non-uniform schedule would strengthen the statistical guarantee in §4. In the revised manuscript we will add a direct comparison: for each experimental budget we recompute the sharpness profile along the adaptive trajectories (using the same finite-difference estimator) and report both the L2 distance to the original calibration profile and the integrated truncation error accumulated under the final schedule. These quantities will be tabulated alongside the existing FID and mode-coverage results to confirm that high-curvature regions remain stable at the budgets used in Tables 2 and 4. revision: yes

  2. Referee: [Table 2, Figure 4, and §5.2] Table 2 and Figure 4 report gains in FID and mode coverage, but the ablation isolating the effect of the sharpness-derived grid versus a simple non-uniform schedule (e.g., linear or exponential) is missing. Without this control, it is unclear whether the observed reduction in inter-mode leakage is attributable to the sharpness profile or to any non-uniform allocation.

    Authors: We acknowledge that the current experiments do not isolate the contribution of the sharpness profile from the mere use of a non-uniform grid. In the revision we will add an ablation in §5.2 (and corresponding rows in Table 2) that compares SharpEuler against two simple non-uniform baselines: (i) a linear ramp schedule and (ii) an exponential schedule whose density matches the average power-law exponent derived in §3.2. All three schedules will use identical numbers of function evaluations; we will report FID, mode coverage, and inter-mode leakage so that readers can see the incremental benefit attributable to the sharpness-derived quantile transform. revision: yes

  3. Referee: [§3.1 (numerical principle) and Eq. (7)] The numerical principle (§3.1) identifies trajectory acceleration as the dominant Euler truncation source and uses finite differences on calibration paths. However, the finite-difference stencil and smoothing parameter are treated as fixed hyperparameters; no sensitivity analysis shows how variation in these choices affects the final timestep grid or sample quality, which is load-bearing for the claim of a “solver-aware” profile.

    Authors: We agree that the robustness of the solver-aware profile to the finite-difference stencil and smoothing bandwidth is important to document. In the revised §3.1 and appendix we will include a sensitivity study that varies the stencil width (1-, 2-, and 3-step central differences) and the Gaussian smoothing bandwidth over a factor of four. For each combination we will show the resulting timestep grids and the downstream FID and mode-coverage numbers on the same models and budgets used in the main experiments. This will demonstrate that the reported gains are stable across reasonable choices of these hyperparameters. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; derivation is self-contained

full rationale

The paper derives its timestep grid from an offline finite-difference sharpness profile computed on uniform calibration trajectories of the pretrained velocity field, then applies a quantile transform and invokes a numerical principle (acceleration as leading Euler error), a variational principle (power-law allocation), and a statistical stability guarantee at the terminal measure. None of these steps reduce by construction to the inputs: the profile is an empirical measurement used to adapt the schedule, the principles are stated as independent justifications rather than tautologies, and no self-citation chain or fitted parameter is renamed as a prediction. The claimed improvement is supported by experiments rather than being definitionally forced. Potential mismatch between calibration and test trajectories under the non-uniform grid is a validity or generalization issue, not a circularity in the derivation chain itself.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on the transferability of an offline finite-difference sharpness estimate and on the validity of the three stated principles. No new physical entities are postulated. Free parameters are implicit in the smoothing and quantile steps but not quantified in the abstract.

free parameters (2)
  • smoothing parameter for sharpness profile
    The profile is smoothed before the quantile transform; the choice of smoother or bandwidth is a free parameter not specified in the abstract.
  • quantile mapping for target budget
    Conversion of the smoothed profile into a timestep grid for any desired number of steps requires a quantile transform whose exact mapping parameters are not detailed.
axioms (2)
  • domain assumption Trajectory acceleration is the leading source of Euler discretization error
    Invoked as the numerical principle justifying focus on sharpness.
  • domain assumption Sharpness-based power-law timestep densities optimize integration accuracy
    Derived from the variational principle mentioned in the abstract.

pith-pipeline@v0.9.0 · 5490 in / 1570 out tokens · 65168 ms · 2026-05-13T02:16:05.934872+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages · 5 internal anchors

  1. [1]

    Stochastic Interpolants: A Unifying Framework for Flows and Diffusions

    Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.arXiv preprint arXiv:2303.08797, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    MIT Press, 2024

    Francis Bach.Learning Theory from First Principles. MIT Press, 2024

    work page 2024

  3. [3]

    Nearly d-linear convergence bounds for diffusion models via stochastic localization

    Joe Benton, Valentin De Bortoli, Arnaud Doucet, and George Deligiannidis. Nearly d-linear convergence bounds for diffusion models via stochastic localization. InThe Twelfth International Conference on Learning Representations, 2024

    work page 2024

  4. [4]

    FLUX.1, 2024

    Black Forest Labs. FLUX.1, 2024. 11

    work page 2024

  5. [5]

    John Wiley & Sons, 2016

    John Charles Butcher.Numerical Methods for Ordinary Differential Equations. John Wiley & Sons, 2016

    work page 2016

  6. [6]

    On the trajectory regularity of ODE-based diffusion sampling.arXiv preprint arXiv:2405.11326, 2024

    Defang Chen, Zhenyu Zhou, Can Wang, Chunhua Shen, and Siwei Lyu. On the trajectory regularity of ODE-based diffusion sampling.arXiv preprint arXiv:2405.11326, 2024

    work page arXiv 2024

  7. [7]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. InAdvances in Neural Information Processing Systems (NeurIPS), 2013

    work page 2013

  8. [8]

    Scaling atomistic protein binder design with generative pretraining and test-time compute

    Kieran Didi, Zuobai Zhang, Guoqing Zhou, Danny Reidenbach, Zhonglin Cao, Sooyoung Cha, Tomas Geffner, Christian Dallago, Jian Tang, Michael M Bronstein, et al. Scaling atomistic protein binder design with generative pretraining and test-time compute. InThe Fourteenth International Conference on Learning Representations

    work page

  9. [9]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. InForty-First International Conference on Machine Learning, 2024

    work page 2024

  10. [10]

    Springer Science & Business Media, 2010

    David F Griffiths and Desmond J Higham.Numerical Methods for Ordinary Differential Equations: Initial Value Problems, volume 5. Springer Science & Business Media, 2010

    work page 2010

  11. [11]

    Quantifying epistemic uncertainty in diffusion models.arXiv preprint arXiv:2602.09170, 2026

    Aditi Gupta, Raphael A Meyer, Yotam Yaniv, Elynn Chen, and N Benjamin Erichson. Quantifying epistemic uncertainty in diffusion models.arXiv preprint arXiv:2602.09170, 2026

    work page arXiv 2026

  12. [12]

    Geometric numerical integration

    Ernst Hairer, Marlis Hochbruck, Arieh Iserles, and Christian Lubich. Geometric numerical integration. Oberwolfach Reports, 3(1):805–882, 2006

    work page 2006

  13. [13]

    Springer, 1993

    Ernst Hairer, Gerhard Wanner, and Syvert P Nørsett.Solving ordinary differential equations I: Nonstiff problems. Springer, 1993

    work page 1993

  14. [14]

    Flowts: Time series generation via rectified flow.arXiv preprint arXiv:2411.07506, 2024

    Yang Hu, Xiao Wang, Zezhen Ding, Lirong Wu, Huatian Zhang, Stan Z Li, Sheng Wang, Jiheng Zhang, Ziyun Li, and Tianlong Chen. Flowts: Time series generation via rectified flow.arXiv preprint arXiv:2411.07506, 2024

    work page arXiv 2024

  15. [15]

    FluxPipeline documentation, 2024

    Hugging Face Diffusers. FluxPipeline documentation, 2024. Default inference budget: 50 steps

    work page 2024

  16. [16]

    Pyramidal flow matching for efficient video generative modeling

    Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. InThe Thirteenth International Conference on Learning Representations

    work page

  17. [17]

    Elucidating the design space of diffusion-based generative models.Advances in Neural Information Processing Systems, 35:26565–26577, 2022

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models.Advances in Neural Information Processing Systems, 35:26565–26577, 2022

    work page 2022

  18. [18]

    Improved precision and recall metric for assessing generative models

    Tuomas Kynkäänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. InAdvances in Neural Information Processing Systems (NeurIPS), 2019

    work page 2019

  19. [19]

    Minimizing trajectory curvature of ODE-based generative models

    Sangyun Lee, Beomsu Kim, and Jong Chul Ye. Minimizing trajectory curvature of ODE-based generative models. InInternational Conference on Machine Learning, pages 18957–18973. PMLR, 2023

    work page 2023

  20. [20]

    A kinetic-energy perspective of flow matching.arXiv preprint arXiv:2602.07928, 2026

    Ziyun Li, Huancheng Hu, Soon Hoe Lim, Xuyu Li, Fei Gao, Enmao Diao, Zezhen Ding, Michalis Vazirgiannis, and Henrik Bostrom. A kinetic-energy perspective of flow matching.arXiv preprint arXiv:2602.07928, 2026

    work page arXiv 2026

  21. [21]

    Is Flow Matching Just Trajectory Replay for Sequential Data?

    Soon Hoe Lim, Shizheng Lin, Michael W Mahoney, and N Benjamin Erichson. Is flow matching just trajectory replay for sequential data?arXiv preprint arXiv:2602.08318, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  22. [22]

    Elucidating the design choice of probability paths in flow matching for forecasting.Transactions on Machine Learning Research, 2024

    Soon Hoe Lim, Yijin Wang, Annan Yu, Emma Hart, Michael W Mahoney, Sherry Li, and N Benjamin Erichson. Elucidating the design choice of probability paths in flow matching for forecasting.Transactions on Machine Learning Research, 2024

    work page 2024

  23. [23]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. InThe Eleventh International Conference on Learning Representations, 2023

    work page 2023

  24. [24]

    Flow Matching Guide and Code

    Yaron Lipman, Marton Havasi, Peter Holderrieth, Neta Shaul, Matt Le, Brian Karrer, Ricky TQ Chen, David Lopez-Paz, Heli Ben-Hamu, and Itai Gat. Flow matching guide and code.arXiv preprint arXiv:2412.06264, 2024. 12

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  25. [25]

    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. InThe Eleventh International Conference on Learning Representations (ICLR), 2023

    work page 2023

  26. [26]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in Neural Information Processing Systems, 35:5775–5787, 2022

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in Neural Information Processing Systems, 35:5775–5787, 2022

    work page 2022

  27. [27]

    Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference.arXiv preprint arXiv:2310.04378, 2023

    work page internal anchor Pith review arXiv 2023

  28. [28]

    Zatom-1: A Multimodal Flow Foundation Model for 3D Molecules and Materials

    Alex Morehead, Miruna Cretu, Antonia Panescu, Rishabh Anand, Maurice Weiler, Tynan Perez, Samuel Blau, Steven Farrell, Wahid Bhimji, Anubhav Jain, Hrushikesh Sahasrabuddhe, Pietro Liò, Tommi Jaakkola, Rafael Gómez-Bombarelli, Rex Ying, N. Benjamin Erichson, and Michael W. Mahoney. Zatom-1: A multimodal flow foundation model for 3d molecules and materials....

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  29. [29]

    Reliable fidelity and diversity metrics for generative models

    Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. Reliable fidelity and diversity metrics for generative models. InProceedings of the 37th International Conference on Machine Learning (ICML), pages 7176–7185, 2020

    work page 2020

  30. [30]

    Entropic Flow Matching for Optimal Time Scheduling, 2024

    NVIDIA BioNeMo. Entropic Flow Matching for Optimal Time Scheduling, 2024. BioNeMo Framework Documentation, MoCo tutorial

    work page 2024

  31. [31]

    Align your steps: optimizing sampling schedules in diffusion models

    Amirmojtaba Sabour, Sanja Fidler, and Karsten Kreis. Align your steps: optimizing sampling schedules in diffusion models. InProceedings of the 41st International Conference on Machine Learning, pages 42947–42975, 2024

    work page 2024

  32. [32]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. InInternational Conference on Learning Representations, 2022

    work page 2022

  33. [33]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InInternational Conference on Learning Representations, 2021

    work page 2021

  34. [34]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. InProceedings of the 40th International Conference on Machine Learning, pages 32211–32252, 2023

    work page 2023

  35. [35]

    Flowllm: Flow matching for material generation with large language models as base distributions.Advances in Neural Information Processing Systems, 37:46025–46046, 2024

    Anuroop Sriram, Benjamin K Miller, Ricky T Chen, and Brandon M Wood. Flowllm: Flow matching for material generation with large language models as base distributions.Advances in Neural Information Processing Systems, 37:46025–46046, 2024

    work page 2024

  36. [36]

    Entropic time schedulers for generative diffusion models

    Dejan Stancevic, Florian Handke, and Luca Ambrogioni. Entropic time schedulers for generative diffusion models. InAdvances in Neural Information Processing Systems 38 (NeurIPS 2025), 2025

    work page 2025

  37. [37]

    Optimal scheduling of dynamic transport

    Panos Tsimpos, Ren Zhi, Jakob Zech, and Youssef Marzouk. Optimal scheduling of dynamic transport. In The Thirty Eighth Annual Conference on Learning Theory, pages 5441–5505, 2025

    work page 2025

  38. [38]

    Springer, 2008

    Cédric Villani.Optimal Transport: Old and New. Springer, 2008

    work page 2008

  39. [39]

    Learning fast samplers for diffusion models by differentiating through sample quality

    Daniel Watson, William Chan, Jonathan Ho, and Mohammad Norouzi. Learning fast samplers for diffusion models by differentiating through sample quality. InInternational Conference on Learning Representations, 2022

    work page 2022

  40. [40]

    Watson, J

    Daniel Watson, Jonathan Ho, Mohammad Norouzi, and William Chan. Learning to efficiently sample from diffusion probabilistic models.arXiv preprint arXiv:2106.03802, 2021

    work page arXiv 2021

  41. [41]

    Score-optimal diffusion schedules.Advances in Neural Information Processing Systems, 37:107960–107983, 2024

    Christopher Williams, Andrew Campbell, Arnaud Doucet, and Saifuddin Syed. Score-optimal diffusion schedules.Advances in Neural Information Processing Systems, 37:107960–107983, 2024

    work page 2024

  42. [42]

    Accelerating diffusion sampling with optimized time steps

    Shuchen Xue, Zhaoqiang Liu, Fei Chen, Shifeng Zhang, Tianyang Hu, Enze Xie, and Zhenguo Li. Accelerating diffusion sampling with optimized time steps. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8292–8301, 2024

    work page 2024

  43. [43]

    Lipschitz singularities in diffusion models

    Zhantao Yang, Ruili Feng, Han Zhang, Yujun Shen, Kai Zhu, Lianghua Huang, Yifei Zhang, Yu Liu, Deli Zhao, Jingren Zhou, et al. Lipschitz singularities in diffusion models. InICLR, 2024. 13

    work page 2024

  44. [44]

    Unipc: A unified predictor-corrector framework for fast sampling of diffusion models.Advances in Neural Information Processing Systems, 36:49842–49869, 2023

    Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models.Advances in Neural Information Processing Systems, 36:49842–49869, 2023

    work page 2023

  45. [45]

    On numerical integration in neural ordinary differential equations

    Aiqing Zhu, Pengzhan Jin, Beibei Zhu, and Yifa Tang. On numerical integration in neural ordinary differential equations. InInternational Conference on Machine Learning, pages 27527–27547. PMLR, 2022. 14 Appendix This appendix is organized as follows. In App. A we discuss related work in detail and position SharpEuler relative to recent works. In App. B we...

    work page 2022