arxiv: 2605.07513 · v1 · submitted 2026-05-08 · 💻 cs.LG

Recognition: no theorem link

Tessellations of Semi-Discrete Flow Matching

Emile Pierret , Johannes Hertrich , Samuel Hurault , Julie Delon

Authors on Pith no claims yet

Pith reviewed 2026-05-11 02:33 UTC · model grok-4.3

classification 💻 cs.LG

keywords flow matchingsemi-discrete transportassignment regionsterminal flow mapgenerative modelingoptimal transport comparisonnon-convex cells

0 comments

The pith

Semi-discrete flow matching produces open simply connected assignment regions that can be non-convex with curved boundaries unlike Laguerre cells.

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

The paper examines the geometry of terminal flow maps when a Gaussian source measure is transported to a finite discrete target in flow matching. It establishes that the preimages of each target atom under the terminal flow map form regions that are open and simply connected, and homeomorphic to the unit ball under one further assumption. A concrete four-point planar counterexample demonstrates that these regions can be non-convex, bounded by curves, and exhibit adjacency and boundedness patterns distinct from the convex cells of semi-discrete optimal transport. The analysis isolates the intrinsic geometry created by the exact closed-form velocity field before any neural approximation is introduced, which clarifies the partitioning behavior that generative models inherit when trained on finite datasets.

Core claim

The terminal assignment regions are the preimages of the discrete target points under the terminal flow map of semi-discrete flow matching. These regions are open and simply connected. Under an additional assumption they are homeomorphic to the unit ball. A planar example with four target points shows the regions can nevertheless be non-convex, possess curved boundaries, and follow different boundedness and adjacency rules than the Laguerre cells of semi-discrete optimal transport.

What carries the argument

Terminal assignment regions: the preimages of individual atoms of the discrete target measure under the terminal flow map induced by the exact semi-discrete velocity field.

If this is right

The exact flow-matching objective partitions space into regions whose topology and shape are independent of later neural approximation.
These regions remain open and simply connected even when they are non-convex, so the flow map still defines a well-behaved assignment for every point in the source space.
The difference from Laguerre cells means semi-discrete flow matching can induce adjacency patterns and boundedness that optimal transport does not produce.
Because the velocity field is available in closed form, the geometry can be studied without optimization artifacts.

Where Pith is reading between the lines

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

If the non-convexity observed in the four-point example persists in higher dimensions, flow matching may separate modes of multimodal targets more flexibly than convex-cell methods.
The topological regularity (openness and simple connectedness) could be used to guarantee that neural approximations of the velocity field still produce invertible maps almost everywhere.
One could test whether the homeomorphism-to-ball property holds generically by sampling random discrete targets and checking the topology of the computed regions.
The curved boundaries suggest that interpolation or sampling procedures inside each region may require different numerical handling than the linear boundaries of Voronoi or Laguerre tessellations.

Load-bearing premise

The analysis uses the exact closed-form velocity field and invokes an additional unspecified assumption to conclude that the regions are homeomorphic to the unit ball.

What would settle it

Explicitly compute the terminal flow map for the four-point planar configuration given in the paper and inspect whether the resulting preimage regions are non-convex with curved boundaries; separately, identify a concrete instance where the additional assumption fails and check whether the regions cease to be homeomorphic to a ball.

Figures

Figures reproduced from arXiv: 2605.07513 by Emile Pierret, Johannes Hertrich, Julie Delon, Samuel Hurault.

**Figure 1.** Figure 1: Semi-discrete assignment cells Γk for 10 points in dimension 2, for (a) Optimal Transport, (b) Flow Matching with closed-form velocity and (c) Flow Matching with velocity approximated by a neural network. In this last case, each grid point is assigned the color of the target point closest to its image under the terminal flow. and Santambrogio [2022]. This raises a basic question: what geometry is intrinsic… view at source ↗

**Figure 2.** Figure 2: Illustration of Corollary 1. For n = 10 points (ak)1≤k≤n, we display the associated semi-discrete flow matching cells of (cak)1≤k≤n for c = 0.1, 0.5, and 1.5, which coincide. This also shows that a point may or may not belong to its associated cell. Reduction to the Affine Hull Next, we show that the FM dynamics can be completely reduced to the dynamics on the affine hull of the points ak. More precisely, … view at source ↗

**Figure 3.** Figure 3: We plot the generalized assignment regions [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Illustration of the centers of the cells. For n = 10 points (ak)1≤k≤n, we depict the associated semi-discrete flow matching cells and the corresponding centers limt→1 γ −1 t (ak), represented by star symbols. The dotted lines indicate the curves (γ −1 t (ak))0≤t<1. We have seen in the previous subsection that the cells Γk are simply connected, meaning they do not contain “onedimensional” holes. In orde… view at source ↗

**Figure 5.** Figure 5: Illustration of the counterexample analyzed in Section [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

**Figure 6.** Figure 6: Illustration of the non-monoticity of the flow map, in the four-points configuration of Section [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

**Figure 7.** Figure 7: shows that the qualitative differences between OT and FM persist in this higher-dimensional latent setting. Compared with Laguerre cells, the FM cells have curved boundaries, are often non-convex, and display different locations and neighborhood relations. Moreover, the trained FM cells closely resemble the exact FM cells, suggesting that even a simple learned velocity can recover the main geometric featur… view at source ↗

read the original abstract

We study Flow Matching in a semi-discrete setting where a Gaussian source is transported toward a discrete target supported on finitely many points. This semi-discrete regime is the theoretical setting behind the use of Flow Matching for generative modeling, where the target distribution is represented by a finite dataset. In this semi-discrete regime, the exact Flow Matching velocity field is available in closed form, which makes it possible to analyze the geometry induced by the terminal flow map independently of optimization and approximation effects. We investigate the terminal assignment regions, namely the preimages of the target atoms under the terminal flow. We show that these regions are open, simply connected and, under an additional assumption, homeomorphic to the unit ball. At the same time, a planar four-point example shows that these cells can differ sharply from Laguerre cells arising in semi-discrete optimal transport: they may be non-convex, have curved boundaries, and exhibit different boundedness and adjacency patterns. These results clarify the geometry intrinsically induced by the exact semi-discrete Flow Matching objective before neural approximation enters the picture.

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.

Desk Editor's Note private letter to a colleague

Semi-discrete Flow Matching produces assignment regions that are topologically simple but geometrically unlike those from optimal transport, as shown by a four-point example.

read the letter

The main point is that semi-discrete Flow Matching creates assignment regions that are open and simply connected, and a simple four-point example shows these regions can be non-convex with curved boundaries unlike the Laguerre cells from optimal transport. The paper does a good job by working with the exact closed-form velocity field. This lets them describe the terminal flow map exactly and analyze its geometry without any neural network or optimization noise. The topological claims follow from that setup, and the counterexample is concrete enough to see the differences in convexity, boundary shape, and adjacency patterns. They separate the unconditional properties from the ones that need an extra assumption, which is a clear way to present the results. The work also ties back to the generative modeling setting where the target is a finite dataset, so the analysis has direct relevance. One soft spot is the unspecified additional assumption needed for the regions to be homeomorphic to the unit ball. Without knowing what it is, it is difficult to judge whether the stronger topological result applies in typical cases or if the four-point example even meets it. The example itself is only in the plane, so it would help to see if similar non-convexity appears in higher dimensions. The paper does not appear to explore how the regions behave under changes in source variance or point spacing. This kind of geometric analysis is for readers interested in the theoretical foundations of flow matching and its relation to optimal transport. Someone working on understanding inductive biases in generative models with discrete data would find it useful. The results are specific enough that the paper deserves a serious referee rather than a desk reject.

Referee Report

2 major / 1 minor

Summary. The paper analyzes the geometry of terminal assignment regions (preimages of target atoms under the terminal flow map) in semi-discrete flow matching, where a Gaussian source is transported to a finite discrete target. Using the exact closed-form velocity field, it claims these regions are open and simply connected unconditionally, and homeomorphic to the unit ball under an additional (unspecified) assumption. A planar four-point example is provided to show that the regions can be non-convex with curved boundaries and differ from Laguerre cells in semi-discrete optimal transport in boundedness and adjacency.

Significance. If the results hold, this clarifies the intrinsic geometry of exact semi-discrete flow matching before neural approximation, which is directly relevant to generative modeling with finite datasets. The closed-form velocity field and the concrete four-point counter-example to Laguerre cells are strengths, providing falsifiable geometric distinctions independent of optimization effects.

major comments (2)

[Abstract] Abstract and main theorem statement: the additional assumption required for homeomorphism to the unit ball is left unspecified. This assumption is load-bearing for the stronger topological claim, yet its precise statement, necessity, and verification against the four-point example (which shows non-convexity) are absent, preventing assessment of applicability to the generative modeling setting.
[Results on four-point example] Four-point planar example (presumably in the results section): while it successfully demonstrates non-convexity, curved boundaries, and differences from Laguerre cells, there is no check whether the additional assumption holds here or how it interacts with the unconditional claims of openness and simple connectedness. The example is central to distinguishing the flow-matching tessellation from optimal transport.

minor comments (1)

Notation for the terminal flow map and assignment regions should be introduced earlier and used consistently to improve readability of the geometric claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thorough reading and valuable feedback on our work. We address the major comments point by point below, indicating where revisions will be made to improve clarity and completeness.

read point-by-point responses

Referee: [Abstract] Abstract and main theorem statement: the additional assumption required for homeomorphism to the unit ball is left unspecified. This assumption is load-bearing for the stronger topological claim, yet its precise statement, necessity, and verification against the four-point example (which shows non-convexity) are absent, preventing assessment of applicability to the generative modeling setting.

Authors: We agree that the additional assumption for the homeomorphism result should be stated explicitly in the abstract and highlighted in the main theorem for better accessibility. The assumption appears in the body of the paper but was not carried forward to the abstract or sufficiently emphasized. We will revise the abstract and theorem statement to include its precise formulation, discuss its necessity relative to the unconditional openness and simple connectedness results, and confirm that it is satisfied by the four-point example (where the regions remain homeomorphic to the disk). This will be a straightforward clarification with no change to the underlying claims. revision: yes
Referee: [Results on four-point example] Four-point planar example (presumably in the results section): while it successfully demonstrates non-convexity, curved boundaries, and differences from Laguerre cells, there is no check whether the additional assumption holds here or how it interacts with the unconditional claims of openness and simple connectedness. The example is central to distinguishing the flow-matching tessellation from optimal transport.

Authors: The openness and simple connectedness of the terminal assignment regions are established unconditionally in the main theorem, without requiring the additional assumption used only for the homeomorphism claim. The four-point example is constructed to satisfy the assumption, which can be verified directly from the closed-form velocity field and resulting flow map. We will add an explicit verification paragraph in the results section showing that the assumption holds and explaining the separation between the unconditional topological properties and the conditional homeomorphism. This addresses the interaction and strengthens the distinction from Laguerre cells. revision: yes

Circularity Check

0 steps flagged

No circularity; topological claims derived directly from closed-form velocity field and model definition

full rationale

The paper's core results on assignment regions being open, simply connected, and (under an extra assumption) homeomorphic to the unit ball follow from the semi-discrete flow matching definition and its exact closed-form velocity field. No fitted parameters are renamed as predictions, no self-citation chain supplies the load-bearing uniqueness or ansatz, and the four-point counter-example is presented as an independent geometric illustration rather than a self-referential fit. The derivation remains self-contained against the stated assumptions without reducing to its inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The analysis rests on the existence of a closed-form velocity field in the semi-discrete regime and on one unspecified additional assumption needed for the homeomorphism result. No free parameters or invented entities are mentioned.

axioms (2)

domain assumption The exact Flow Matching velocity field is available in closed form for a Gaussian source and discrete target.
Invoked in the abstract to enable analysis independent of optimization effects.
ad hoc to paper An additional (unspecified) assumption holds that makes the assignment regions homeomorphic to the unit ball.
Required for the strongest topological claim but not detailed in the abstract.

pith-pipeline@v0.9.0 · 5485 in / 1521 out tokens · 83199 ms · 2026-05-11T02:33:27.601353+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

299 extracted references · 299 canonical work pages · 1 internal anchor

[1]

The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

Gradient Variance Reveals Failure Modes in Flow-Based Generative Models , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

work page
[2]

Memorization and regularization in generative diffusion models.arXiv preprint arXiv:2501.15785,

Memorization and Regularization in Generative Diffusion Models , author =. arXiv:2501.15785 , year =

work page arXiv
[3]

The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

work page
[4]

The Eleventh International Conference on Learning Representations , year=

Building Normalizing Flows with Stochastic Interpolants , author=. The Eleventh International Conference on Learning Representations , year=

work page
[5]

, title =

Bishop, Christopher M. , title =. 2006 , isbn =

work page 2006
[6]

The Eleventh International Conference on Learning Representations , year=

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

work page
[7]

The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

On the Relation between Rectified Flows and Optimal Transport , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

work page
[8]

Vansh Bansal and Saptarshi Roy and Purnamrita Sarkar and Alessandro Rinaldo , year=. On the. 2410.14949 , archivePrefix=

work page arXiv
[9]

2024 , eprint=

Flow Matching Guide and Code , author=. 2024 , eprint=

work page 2024
[10]

The Eleventh International Conference on Learning Representations , year=

Flow Matching for Generative Modeling , author=. The Eleventh International Conference on Learning Representations , year=

work page
[11]

Foundations and Trends in Machine Learning , volume=

Computational Optimal Transport , author=. Foundations and Trends in Machine Learning , volume=. 2019 , doi=

work page 2019
[12]

Algorithmica , volume=

Minkowski-Type Theorems and Least-Squares Clustering , author=. Algorithmica , volume=. 1998 , doi=

work page 1998
[13]

Computer Graphics Forum , volume=

A Multiscale Approach to Optimal Transport , author=. Computer Graphics Forum , volume=. 2011 , doi=

work page 2011
[14]

Journal of the European Mathematical Society , volume=

Convergence of a Newton Algorithm for Semi-discrete Optimal Transport , author=. Journal of the European Mathematical Society , volume=. 2019 , doi=

work page 2019
[15]

arXiv preprint arXiv:2301.13731 , year=

A relaxed proximal gradient descent algorithm for convergent plug-and-play with proximal denoiser , author=. arXiv preprint arXiv:2301.13731 , year=

work page arXiv
[16]

Proximal Denoiser for Convergent Plug-and-Play Optimization with Nonconvex Regularization , author =

work page
[17]

Advances in Neural Information Processing Systems , year=

Convergent Bregman Plug-and-Play Image Restoration for Poisson Inverse Problems , author=. Advances in Neural Information Processing Systems , year=

work page
[18]

Solv- ing inverse problems by joint posterior maximization with a vae prior.arXiv preprint arXiv:1911.06379, 2019

Solving Inverse Problems by Joint Posterior Maximization with a VAE Prior , author=. arXiv preprint arXiv:1911.06379 , year=

work page arXiv 1911
[19]

arXiv preprint arXiv:1410.8479 , year=

Metric selection in Douglas-Rachford splitting and ADMM , author=. arXiv preprint arXiv:1410.8479 , year=

work page arXiv
[20]

arXiv preprint arXiv:1405.7741 , year=

Properties of Pseudocontractive Updates in Convex Optimization , author=. arXiv preprint arXiv:1405.7741 , year=

work page arXiv
[21]

2011 , publisher=

Distributed optimization and statistical learning via the alternating direction method of multipliers , author=. 2011 , publisher=

work page 2011
[22]

SIAM Journal Imaging Sci

Optimal transport with proximal splitting , author=. SIAM Journal Imaging Sci. , volume=. 2014 , publisher=

work page 2014
[23]

Primer on monotone operator methods , author=. Appl. Comput. Math , volume=

work page
[24]

Physica D: nonlinear phenomena , volume=

Nonlinear total variation based noise removal algorithms , author=. Physica D: nonlinear phenomena , volume=. 1992 , publisher=

work page 1992
[25]

Learning deep

Zhang, Kai and Zuo, Wangmeng and Gu, Shuhang and Zhang, Lei , booktitle=CVPR, pages=. Learning deep

work page
[26]

SIAM Journal Optimization , volume=

The resolvent average of monotone operators: dominant and recessive properties , author=. SIAM Journal Optimization , volume=. 2016 , publisher=

work page 2016
[27]

Proximit\'e et dualit\'e dans un espace hilbertien , url =

Moreau, Jean-Jacques , journal =. Proximit\'e et dualit\'e dans un espace hilbertien , url =

work page
[28]

Foundations and Trends in optimization , volume=

Proximal algorithms , author=. Foundations and Trends in optimization , volume=. 2014 , publisher=

work page 2014
[29]

SIAM Journal Optimization , volume=

Global convergence of splitting methods for nonconvex composite optimization , author=. SIAM Journal Optimization , volume=. 2015 , publisher=

work page 2015
[30]

Doklady akademii nauk , volume=

On the solution of ill-posed problems and the method of regularization , author=. Doklady akademii nauk , volume=. 1963 , organization=

work page 1963
[31]

The flow map of the

Hugo Lavenant and Filippo Santambrogio , keywords =. The flow map of the. Applied Mathematics Letters , volume =. 2022 , issn =. doi:https://doi.org/10.1016/j.aml.2022.108225 , url =

work page doi:10.1016/j.aml.2022.108225 2022
[32]

Combettes and Jean-Christophe Pesquet

Lipschitz certificates for neural network structures driven by averaged activation operators , author=. arXiv preprint arXiv:1903.01014 , year=

work page arXiv 1903
[33]

On decomposing the proximal map , author=

work page
[34]

2019 , publisher=

Charles Dossal , title=. 2019 , publisher=

work page 2019
[35]

Mathematical Programming , volume=

On the ergodic convergence rates of a first-order primal--dual algorithm , author=. Mathematical Programming , volume=. 2016 , publisher=

work page 2016
[36]

and Pock, T

Chambolle, A. and Pock, T. , title =. Journal of Math. Imaging and Vision , pages =

work page
[37]

IEEE Access , volume=

A survey on nonconvex regularization-based sparse and low-rank recovery in signal processing, statistics, and machine learning , author=. IEEE Access , volume=

work page
[38]

Douglas--Rachford Splitting and

Themelis, Andreas and Patrinos, Panagiotis , journal=. Douglas--Rachford Splitting and

work page
[39]

FFDNet: Toward a fast and flexible solution for

Zhang, Kai and Zuo, Wangmeng and Zhang, Lei , journal=TIP, volume=. FFDNet: Toward a fast and flexible solution for. 2018 , publisher=

work page 2018
[40]

Beyond a gaussian denoiser: Residual learning of deep

Zhang, Kai and Zuo, Wangmeng and Chen, Yunjin and Meng, Deyu and Zhang, Lei , journal=TIP, volume=. Beyond a gaussian denoiser: Residual learning of deep

work page
[41]

arXiv preprint arXiv:2003.04595 , year=

Nonlinear Power Method for Computing Eigenvectors of Proximal Operators and Neural Networks , author=. arXiv preprint arXiv:2003.04595 , year=

work page arXiv 2003
[43]

Nonlinear Analysis: Theory, Methods & Applications , volume=

Right Bregman nonexpansive operators in Banach spaces , author=. Nonlinear Analysis: Theory, Methods & Applications , volume=. 2012 , publisher=

work page 2012
[44]

SIAM Journal on Optimization , volume=

First order methods beyond convexity and Lipschitz gradient continuity with applications to quadratic inverse problems , author=. SIAM Journal on Optimization , volume=. 2018 , publisher=

work page 2018
[45]

Bolte, J. The. SIAM Journal on Optimization , volume=. 2007 , publisher=

work page 2007
[46]

2009 , publisher=

Variational analysis , author=. 2009 , publisher=

work page 2009
[47]

2011 , author=

Fixed-Point Algorithms for Inverse Problems in Science and Engineering , title=. 2011 , author=

work page 2011
[48]

IEEE Journal of Selected Topics in Signal Processing , volume=

A Douglas--Rachford splitting approach to nonsmooth convex variational signal recovery , author=. IEEE Journal of Selected Topics in Signal Processing , volume=. 2007 , publisher=

work page 2007
[49]

Journal of Mathematical Analysis and Applications , volume=

Compositions and convex combinations of averaged nonexpansive operators , author=. Journal of Mathematical Analysis and Applications , volume=. 2015 , publisher=

work page 2015
[50]

arXiv preprint arXiv:2012.05703 , year=

TFPnP: Tuning-free Plug-and-Play Proximal Algorithm with Applications to Inverse Imaging Problems , author=. arXiv preprint arXiv:2012.05703 , year=

work page arXiv 2012
[51]

Bauschke, H. H. and Combettes, P. L. , PUBLISHER =

work page
[52]

Nonlinear Analysis: Theory, Methods & Applications , volume=

Strong convergence theorems for a common zero of a countably infinite family of -inverse strongly accretive mappings , author=. Nonlinear Analysis: Theory, Methods & Applications , volume=. 2009 , publisher=

work page 2009
[53]

Journal of Mathematical Analysis and Applications , volume=

Strong convergence theorems for resolvents of accretive operators in Banach spaces , author=. Journal of Mathematical Analysis and Applications , volume=. 1980 , publisher=

work page 1980
[54]

P. L. Lions and B. Mercier. Splitting Algorithms for the Sum of Two Nonlinear Operators. Siam Journal Numer. Anal. 1979

work page 1979
[55]

Journal Eckstein and D. P. Bertsekas , title =. Mathematical Programming , year =

work page
[56]

Gabay and B

D. Gabay and B. Mercier. A dual algorithm for the solution of nonlinear variational problems via finite element approximation. Computers & Mathematics with Applications. 1976

work page 1976
[57]

and Marroco, A

Glowinski, R. and Marroco, A. , journal =. Sur l'approximation, par éléments finis d'ordre un, et la résolution, par pénalisation-dualité d'une classe de problèmes de Dirichlet non linéaires , volume =

work page
[58]

Hurwicz and H

K.Journal Arrow and L. Hurwicz and H. Uzawa , title=

work page
[59]

2013 , journal=

Convergence of descent methods for semi-algebraic and tame problems: proximal algorithms, forward–backward splitting, and regularized Gauss–Seidel methods , author=. 2013 , journal=

work page 2013
[60]

Characterizations of

Bolte, J. Characterizations of. Transactions of the American Mathematical Society , volume=

work page
[61]

IEEE Transactions on image processing , volume=

Image denoising by sparse 3-D transform-domain collaborative filtering , author=. IEEE Transactions on image processing , volume=. 2007 , publisher=

work page 2007
[62]

2005 , organization=

A non-local algorithm for image denoising , author=. 2005 , organization=

work page 2005
[63]

Global convergence of

Wang, Yu and Yin, Wotao and Zeng, Jinshan , journal=. Global convergence of

work page
[64]

and Deledalle, C.-A

Sutour, C. and Deledalle, C.-A. and Aujol, J.-F. , journal=. Adaptive Regularization of the NL-Means: Application to Image and Video Denoising , year=

work page
[65]

IEEE Transactions Computational Imaging , volume=

Scalable plug-and-play ADMM with convergence guarantees , author=. IEEE Transactions Computational Imaging , volume=. 2021 , publisher=

work page 2021
[66]

IEEE Journal of Selected Topics in Signal Processing , volume=

Rare: Image reconstruction using deep priors learned without groundtruth , author=. IEEE Journal of Selected Topics in Signal Processing , volume=. 2020 , publisher=

work page 2020
[67]

arXiv preprint arXiv:1905.05113 , year=

Block coordinate regularization by denoising , author=. arXiv preprint arXiv:1905.05113 , year=

work page arXiv 1905
[68]

Diffusion models beat gans on image synthesis , author=

work page
[69]

IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

Diffusion models in vision: A survey , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2023 , publisher=

work page 2023
[70]

arXiv preprint arXiv:2306.09251 , year=

Towards faster non-asymptotic convergence for diffusion-based generative models , author=. arXiv preprint arXiv:2306.09251 , year=

work page arXiv
[71]

International Conference on Algorithmic Learning Theory , pages=

Convergence of score-based generative modeling for general data distributions , author=. International Conference on Algorithmic Learning Theory , pages=. 2023 , organization=

work page 2023
[72]

Analyzing and improving the training dynamics of diffusion models , author=

work page
[73]

2024 , url=

Dustin Podell and Zion English and Kyle Lacey and Andreas Blattmann and Tim Dockhorn and Jonas M. 2024 , url=

work page 2024
[74]

2023 , url=

DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking , author=. 2023 , url=

work page 2023
[75]

ACM Computing Surveys , volume=

Diffusion models: A comprehensive survey of methods and applications , author=. ACM Computing Surveys , volume=. 2023 , publisher=

work page 2023
[76]

2024 , month = nov, journal =

Dynamical Regimes of Diffusion Models , author =. 2024 , month = nov, journal =

work page 2024
[77]

2018 , organization=

prDeep: robust phase retrieval with a flexible deep network , author=. 2018 , organization=

work page 2018
[78]

Async-RED: A provably convergent asynchronous block parallel stochastic method using deep denoising priors,

Async-RED: A provably convergent asynchronous block parallel stochastic method using deep denoising priors , author=. arXiv preprint arXiv:2010.01446 , year=

work page arXiv 2010
[79]

Weighted nuclear norm minimization with application to image denoising , author=

work page
[80]

Rudin, Leonid I and Osher, Stanley and Fatemi, Emad , title =. Phys. D

work page
[81]

IEEE Transactions on Image Processing , volume=

Image restoration by iterative denoising and backward projections , author=. IEEE Transactions on Image Processing , volume=. 2018 , publisher=

work page 2018

Showing first 80 references.