Pith. sign in

REVIEW 2 major objections 5 minor 160 references

Neural networks that learn conditional means and covariances turn multi-step image generators into stronger one-step models.

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 · grok-4.5

2026-07-30 18:56 UTC pith:EASKERXM

load-bearing objection Clean theory-to-practice post-training method: polynomial denoiser projections justify conditional AMFD, and the GenEval one-step wins over multi-step FLUX teachers are the real headline. the 2 major comments →

arxiv 2607.26860 v1 pith:EASKERXM submitted 2026-07-29 cs.LG

Amortized Moment Matching for Visual Generation

Wenze Liu , Xintao Wang , Pengfei Wan , Xiangyu Yue This is my paper
classification cs.LG
keywords amortized moment matchingAmortized Fréchet Distancediffusion denoisersone-step generationconditional momentstext-to-image post-trainingpolynomial projectionflow matching
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 argues that diffusion denoisers secretly encode the moments of the data distribution, and that you can exploit that fact to train generators by matching those moments with neural networks instead of hand-computed statistics. Projecting the optimal denoiser onto polynomials of degree n recovers data moments up to order n+1; the affine (degree-1) case yields a practical loss called Amortized Fréchet Distance (AMFD). Unlike prior Fréchet losses that compute global means and covariances over huge populations, AMFD trains small neural amortizers that predict conditional moments on the fly, using a matrix-free Jacobian trick so the method scales to high-dimensional features and text conditioning. Used as a post-training objective on frozen representation spaces, AMFD produces more stable one-step generators on ImageNet than exact moment matching and, on text-to-image models, lets one-step students beat their multi-step teachers on instruction-following while matching human-preference scores. The work also shows that matching only the first two moments works well only in spaces that already carry strong semantics.

Core claim

An n-th degree polynomial projection of the optimal diffusion or flow-matching denoiser uniquely identifies the conditional data moments up to order n+1. Instantiating the affine case produces the Amortized Fréchet Distance loss, whose neural amortizers learn conditional means and covariance operators and supply generator gradients that convert multi-step models into one-step generators that surpass their teachers on instruction-following benchmarks.

What carries the argument

Denoising Projection Moment Hierarchy (Theorem 3.1): the orthogonal projection of the optimal velocity field onto degree-n polynomials matches moment tensors up to order n, and a triangular recurrence recovers the data moments themselves. The affine specialization yields the AMFD loss, trained by alternating amortizer updates (denoising regression) with generator updates that pull means and Jacobi-adjusted covariances together.

Load-bearing premise

That neural amortizers trained by affine denoising, plus a local Jacobi pullback calibrated only at isotropic covariance, give generator gradients faithful enough to the true conditional Fréchet discrepancy that matching just the first two moments identifies the target distribution.

What would settle it

Train the same generator with exact population conditional means and covariances (or a higher-order polynomial amortizer) versus AMFD; if AMFD’s one-step GenEval/FDr6 gains disappear or reverse while exact matching still succeeds, the amortizer-plus-Jacobi approximation is the failure point.

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

If this is right

  • One-step post-training of large text-to-image models can improve instruction following beyond the multi-step teacher without hurting preference scores.
  • Matching only mean and covariance is sufficient for generation once the feature space is semantically rich (e.g., RAE latents), but insufficient in raw pixels or weak VAEs.
  • Condition-aware amortizers become essential as soon as the conditioning signal is high-dimensional (text), while unconditional amortizers can suffice for simple class labels.
  • Matrix-free JVP parameterization lets second-order moment matching run in native high-dimensional generative spaces where explicit covariance matrices are intractable.

Where Pith is reading between the lines

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

  • If higher-degree polynomial amortizers can be made tractable, the same framework could supply third- and fourth-order training signals without leaving the diffusion-projection theory.
  • The robustness of neural amortization over exact batch statistics suggests other distribution-matching objectives (MMD, energy distance) may also benefit from learned rather than empirical estimators.
  • The finding that two moments suffice only in strong semantic spaces implies future one-step methods should co-design the representation encoder with the moment matcher rather than treat encoders as frozen black boxes.

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

2 major / 5 minor

Summary. The paper introduces amortized moment matching: neural networks learn conditional data moments as distributional training signals for generators. By projecting optimal diffusion/flow denoisers onto polynomial classes, Theorem 3.1 (formalized as B.1) shows that a degree-n projection uniquely identifies conditional moments up to order n+1 via normal equations and a triangular recurrence. The tractable affine case (Corollary 3.2) yields an injective map from covariance to an affine operator R_t, instantiated as the Amortized Fréchet Distance (AMFD) loss. AMFD uses neural amortizers with matrix-free JVP parameterization and alternating generator–amortizer optimization to match conditional means and covariances in frozen representation spaces. Empirically, as a post-training objective it improves one-step ImageNet generation over FD-loss (FDr6 1.79/1.75 on JiT-H/pMF-H), probes native generative spaces, and scales to text-to-image, where one-step AMFD-post-trained FLUX.2 [klein] 4B models exceed multi-step teachers on GenEval (0.846 vs 0.794) while matching PickScore.

Significance. If the results hold, the work gives a clean theoretical bridge from diffusion denoisers to moment hierarchies and a practical, condition-aware alternative to population FD matching that scales to high-dimensional and text-conditioned settings. Strengths include a full orthogonal-projection proof with an injective Σ↔R_t map (Appendix B), a matrix-free JVP pipeline, released code and checkpoints, and broad empirics spanning ImageNet scales, native-space probes, PixelGen, and FLUX T2I against FD-loss, DMD2, and iRDM. The condition-aware gains on GenEval and the robustness of the neural formulation versus exact FD-loss (Figures 6–7) are concrete contributions to one-step post-training.

major comments (2)
  1. [§4.2, Eqs. 17–18; Appendix C.2] §4.2 and Appendix C.2 derive the covariance weight ω_cov(t) and the Jacobi pullback δa_J under the isotropic equilibrium Σ=I, where operators commute and scalar calculus applies. The generator is then trained far from this point, especially early in post-training. The manuscript shows empirical robustness versus FD-loss (Figures 6–7) and that the non-Jacobi variant is weaker, but does not quantify how well δa_J tracks the true Fréchet covariance force away from Σ=I. A short diagnostic (e.g., cosine similarity or relative error between δa_J and the analytic FD gradient on held-out features across training) would make the central practical claim—that the amortizer+Jacobi pipeline supplies faithful second-order signals—much firmer.
  2. [Table 1; §5.3; Table 3] Table 1 reports a large estimation gap (AMFD-C vs AMFD-U) on Inception features, yet Table 2 prefers AMFD-U on ImageNet while Table 3 shows AMFD-C is essential for multi-step FLUX/PixelGen GenEval. The narrative explanation (class labels vs rich text; estimation difficulty vs conditional signal) is plausible but not stress-tested with a controlled sweep (e.g., number of classes/prompt diversity vs conditional benefit, or amortizer MSE vs final GenEval/FDr6). Because “condition-aware matching” is a headline claim of AMFD over FD-loss, a compact analysis tying amortizer fidelity to downstream gains would better support when AMFD-C should be preferred.
minor comments (5)
  1. [Figures 1–2] Figure 1/2 captions and the repeated figure blocks in the preprint are slightly redundant; a single clean comparison figure would improve readability.
  2. [§4.1, Eq. 12] Notation for residual affine operator R=−I+tA and time-centered features is introduced quickly in §4.1; a one-line reminder that injectivity of Σ↦R still holds under the residual reparameterization would help.
  3. [§6] Limitations (§6) correctly note post-training-only scope and two-moment restriction; briefly stating whether higher-degree polynomial amortizers are computationally blocked only by tensor size or also by optimization would align the theory section with the limitations.
  4. [Table 3] Table 3 PickScore for DMD2 is left blank with an explanation in the appendix; a footnote in the main table would avoid reader confusion.
  5. [Throughout / Appendix C] Minor typos: “alogrithm” (Appendix C.1), “Fr ´echet” spacing inconsistencies, and “V A-V AE” / “V AE” formatting.

Circularity Check

0 steps flagged

No significant circularity: moment hierarchy is standard L2 projection math; AMFD and empirical wins are not forced by definition or self-citation.

full rationale

The load-bearing theoretical claim (Theorem B.1 / 3.1 and Corollary B.2) is a self-contained Hilbert-space argument: the degree-n polynomial minimizer of the flow-matching residual is the orthogonal projection of v*, hence matches velocity–feature moment tensors G_k^t up to order n; the binomial expansion of the linear interpolant then yields a triangular recurrence that uniquely recovers data moments M_j up to order n+1. The affine case is ordinary population OLS, with an explicit invertible map Σ ↦ R_t. None of this defines moments in terms of the claimed result, fits a parameter and renames it a prediction, or imports a uniqueness theorem from the authors’ prior work. AMFD is an engineering instantiation (neural amortizer + matrix-free JVP + Jacobi pullback with ω_cov calibrated at isotropic Σ=I); that calibration is an approximation choice, not a circular derivation of the empirical metrics. Generator quality is measured on external benchmarks (FID, FDr6, GenEval, PickScore) against published baselines and teachers. Self-citations (e.g. Liu & Yue on ODE integration) are peripheral and not load-bearing. Score 0.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 3 invented entities

The central claim rests on standard flow-matching geometry, L2 polynomial projection, finite second moments, and the modeling choice that frozen vision-encoder features plus first-two-moment matching suffice for generation quality. Free parameters are mostly optimization/architecture knobs (t, ρ, ε, MLP width, ω_cov calibration point). Invented entities are the AMFD objective and the moment amortizer networks—method objects, not physical posits. No new particles or forces.

free parameters (5)
  • interpolation time t (default 0.25) = 0.25
    Fixed by ablation (Fig. 4); theory says any t∈(0,1) identifies moments, but finite-sample error depends on t.
  • multi-representation norm exponents ρ, ε = ρ=1, ε=0.01
    Copied from FD-loss settings to stabilize multi-encoder gradients (Eq. 20); not derived.
  • covariance weight ω_cov(t) = ½(t²+(1-t)²)⁴/(1-t)³
    Analytic form calibrated so generator force matches ½dΣ only at isotropic Σ=I (App. C.2); used globally.
  • amortizer MLP capacity c2048d16a4 and AdaLN grouping = c2048d16a4
    Chosen where MSE_μ plateaus (Fig. 5); architecture hyperparameter.
  • generator/amortizer learning rates and steps = per Table 6–7
    Task-specific (1e-5/1e-6 ImageNet; 5e-6 FLUX 1500 steps, etc.); standard fit-to-validation practice.
axioms (6)
  • domain assumption Linear interpolant flow matching: Xt=tX1+(1-t)X0, optimal v⋆=E[X1−X0|Xt,c]
    §2.2–3; standard in Lipman/Liu flow matching, used as the denoiser whose projections yield moments.
  • standard math Polynomial subspace is dense in L2(PZ|c) so projections converge to v⋆
    Theorem B.1 density assumption for the infinite hierarchy; not needed for finite affine AMFD.
  • standard math X1|c has finite second moment so affine OLS / Rt(c) is well-defined and Σ↦Rt injective for t∈(0,1)
    Corollary 3.2 / B.2; load-bearing for claiming covariance is identified.
  • domain assumption Frozen pretrained representation encoders ϕ make first-two-moment matching a sufficient generative signal
    Inherited from FD-loss / perceptual metric literature; native-space experiments (§5.3 Fig. 8) show this fails without strong semantics.
  • ad hoc to paper Alternating amortizer–generator optimization with stop-grad on fake branch reaches useful conditional moment estimates
    Algorithm 1 / §4.2; no convergence proof, justified empirically vs FD-loss dynamics (Figs. 6–7).
  • ad hoc to paper Jacobi pullback δa_J through A↦Σ at the generated branch is a better generator force than raw (Ag−Ar)u
    §4.2; authors state non-Jacobi variant is 'significantly less effective' without full ablation table in main text.
invented entities (3)
  • Amortized Fréchet Distance (AMFD) loss independent evidence
    purpose: Condition-aware, matrix-free second-order moment matching objective for generator post-training.
    Defined in §4 via neural μ and JVP-parameterized A; primary methodological object.
  • Neural moment amortizer ψ (affine denoiser form) independent evidence
    purpose: Predict conditional mean and covariance operator without materializing Σ or population batches.
    §4.1; trained by Eq. 14; evidence is estimation MSE tables and downstream gen metrics.
  • Denoising Projection Moment Hierarchy independent evidence
    purpose: Theoretical scaffold tying degree-n denoiser projections to moments through order n+1.
    Theorem 3.1/B.1; mathematical construction rather than physical entity; proved under stated assumptions.

pith-pipeline@v1.2.0-daily-grok45 · 33486 in / 4192 out tokens · 66644 ms · 2026-07-30T18:56:49.392920+00:00 · methodology

0 comments
read the original abstract

We propose amortized moment matching, utilizing neural networks to learn data moments as distributional training signals. By casting diffusion denoisers through polynomial projections, we establish a general framework for moment amortization, revealing that an $n$-th degree projection explicitly identifies data moments up to order $n+1$. Derived from the tractable affine case, we instantiate the Amortized Fr\'{e}chet Distance (AMFD) loss. Unlike FD-loss which relies on explicit marginal moment calculations, AMFD is able to dynamically learn conditional moments via an alternating, matrix-free optimization pipeline that effortlessly scales to high-dimensional data. When operating on global representation features, AMFD serves as a powerful post-training objective; empirically, its neural formulation yields more robust training dynamics than exact statistical matching, substantially surpassing the FD baseline on the FDr$^6$ metric and achieving superior one-step generation on ImageNet. Furthermore, it unlocks direct exploration within native generative spaces, suggesting that the first two moments can identify target distributions only in spaces with strong semantics. Finally, when scaled to text-to-image generation, the condition-aware nature of AMFD unlocks massive gains in instruction-following capabilities, enabling our one-step models to outperform their multi-step FLUX.2 [klein] 4B teachers on the GenEval benchmark while achieving on-par performance on PickScore. Code and checkpoints are available at https://github.com/poppuppy/amfd.

Figures

Figures reproduced from arXiv: 2607.26860 by Pengfei Wan, Wenze Liu, Xiangyu Yue, Xintao Wang.

Figure 1
Figure 1. Figure 1: Diffusion denoiser polynomial projection. The Maclaurin series expands e x into scalar polynomial terms; analo￾gously, expanding the optimal diffusion denoiser via tensor powers yields high￾dimensional coefficients. We show that these coefficients explicitly map to the un￾derlying distributional moments, where A0 corresponds to the mean, A1 to the covari￾ance, and An to the (n + 1)-th moment. Distribution … view at source ↗
Figure 2
Figure 2. Figure 2: Comparison between FD-loss and our AMFD loss. (a) FD-loss explicitly computes marginal distribution statistics over a large population, propagating gradients through mini-batches. (b) Our AMFD loss uses neural amortizers which can leverage neural generalization to naturally encode conditions. Through an alternating training scheme, these amortizers dynamically maintain the conditional mean and covariance s… view at source ↗
Figure 3
Figure 3. Figure 3: One-step text-to-image samples generated by our AMFD post-trained FLUX.2 [klein] 4B. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Estimation accuracy across amortizer MLP sizes. Variant MSEµ MSEΣ AMFD-C 5.75 2.18 AMFD-U 0.0106 0.114 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Impact of pretraining epochs on JiT-B/16 post-training [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: Uncurated samples from JiT-H post-trained with FD-loss (left), our AMFD-C (middle), [PITH_FULL_IMAGE:figures/full_fig_p028_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Uncurated samples from pMF-H post-trained with FD-loss (left), our AMFD-C (middle), [PITH_FULL_IMAGE:figures/full_fig_p029_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Additional text-to-image samples generated by the FLUX.2 [klein] 4B model post-trained [PITH_FULL_IMAGE:figures/full_fig_p030_11.png] 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

160 extracted references · 73 linked inside Pith

  1. [1]

    International conference on machine learning , pages=

    Deep unsupervised learning using nonequilibrium thermodynamics , author=. International conference on machine learning , pages=. 2015 , organization=

    2015

  2. [2]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  3. [3]

    arXiv preprint arXiv:2010.02502 , year=

    Denoising diffusion implicit models , author=. arXiv preprint arXiv:2010.02502 , year=

    Pith/arXiv arXiv 2010

  4. [4]

    International conference on machine learning , pages=

    Improved denoising diffusion probabilistic models , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  5. [5]

    Advances in neural information processing systems , volume=

    Generative modeling by estimating gradients of the data distribution , author=. Advances in neural information processing systems , volume=

  6. [6]

    arXiv preprint arXiv:2011.13456 , year=

    Score-based generative modeling through stochastic differential equations , author=. arXiv preprint arXiv:2011.13456 , year=

    Pith/arXiv arXiv 2011

  7. [7]

    arXiv preprint arXiv:2303.01469 , year=

    Consistency models , author=. arXiv preprint arXiv:2303.01469 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2310.14189 , year=

    Improved techniques for training consistency models , author=. arXiv preprint arXiv:2310.14189 , year=

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2310.02279 , year=

    Consistency trajectory models: Learning probability flow ode trajectory of diffusion , author=. arXiv preprint arXiv:2310.02279 , year=

    Pith/arXiv arXiv

  10. [10]

    arXiv preprint arXiv:2406.14548 , year=

    Consistency models made easy , author=. arXiv preprint arXiv:2406.14548 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2403.06807 , year=

    Multistep consistency models , author=. arXiv preprint arXiv:2403.06807 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2410.11081 , year=

    Simplifying, stabilizing and scaling continuous-time consistency models , author=. arXiv preprint arXiv:2410.11081 , year=

    Pith/arXiv arXiv

  13. [13]

    Advances in neural information processing systems , volume=

    Elucidating the design space of diffusion-based generative models , author=. Advances in neural information processing systems , volume=

  14. [14]

    Advances in neural information processing systems , volume=

    Diffusion models beat gans on image synthesis , author=. Advances in neural information processing systems , volume=

  15. [15]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Analyzing and improving the training dynamics of diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  16. [16]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  17. [17]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  18. [18]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Align your latents: High-resolution video synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  19. [19]

    Advances in Neural Information Processing Systems , volume=

    Video diffusion models , author=. Advances in Neural Information Processing Systems , volume=

  20. [20]

    arXiv preprint arXiv:2311.15127 , year=

    Stable video diffusion: Scaling latent video diffusion models to large datasets , author=. arXiv preprint arXiv:2311.15127 , year=

    Pith/arXiv arXiv

  21. [21]

    Advances in Neural Information Processing Systems , volume=

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps , author=. Advances in Neural Information Processing Systems , volume=

  22. [22]

    arXiv preprint arXiv:2211.01095 , year=

    Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models , author=. arXiv preprint arXiv:2211.01095 , year=

    Pith/arXiv arXiv

  23. [23]

    Advances in Neural Information Processing Systems , volume=

    Dpm-solver-v3: Improved diffusion ode solver with empirical model statistics , author=. Advances in Neural Information Processing Systems , volume=

  24. [24]

    arXiv preprint arXiv:2101.02388 , year=

    Knowledge distillation in iterative generative models for improved sampling speed , author=. arXiv preprint arXiv:2101.02388 , year=

    Pith/arXiv arXiv

  25. [25]

    International conference on machine learning , pages=

    Fast sampling of diffusion models via operator learning , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  26. [26]

    arXiv preprint arXiv:2202.00512 , year=

    Progressive distillation for fast sampling of diffusion models , author=. arXiv preprint arXiv:2202.00512 , year=

    Pith/arXiv arXiv

  27. [27]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    On distillation of guided diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  28. [28]

    European Conference on Computer Vision , pages=

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  29. [29]

    European Conference on Computer Vision , pages=

    Adversarial diffusion distillation , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  30. [30]

    SIGGRAPH Asia 2024 Conference Papers , pages=

    Fast high-resolution image synthesis with latent adversarial diffusion distillation , author=. SIGGRAPH Asia 2024 Conference Papers , pages=

    2024

  31. [31]

    arXiv preprint arXiv:2206.02262 , year=

    Diffusion-gan: Training gans with diffusion , author=. arXiv preprint arXiv:2206.02262 , year=

    Pith/arXiv arXiv

  32. [32]

    Advances in Neural Information Processing Systems , volume=

    Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation , author=. Advances in Neural Information Processing Systems , volume=

  33. [33]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    One-step diffusion with distribution matching distillation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  34. [34]

    Advances in Neural Information Processing Systems , volume=

    Improved distribution matching distillation for fast image synthesis , author=. Advances in Neural Information Processing Systems , volume=

  35. [35]

    Advances in Neural Information Processing Systems , volume=

    Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models , author=. Advances in Neural Information Processing Systems , volume=

  36. [36]

    arXiv preprint arXiv:2410.20898 , year=

    Diff-Instruct*: Towards Human-Preferred One-step Text-to-image Generative Models , author=. arXiv preprint arXiv:2410.20898 , year=

    Pith/arXiv arXiv

  37. [37]

    arXiv preprint arXiv:2405.16852 , year=

    Em distillation for one-step diffusion models , author=. arXiv preprint arXiv:2405.16852 , year=

    Pith/arXiv arXiv

  38. [38]

    Advances in Neural Information Processing Systems , volume=

    Multistep distillation of diffusion models via moment matching , author=. Advances in Neural Information Processing Systems , volume=

  39. [39]

    International Conference on Learning Representations , year=

    Learning fast samplers for diffusion models by differentiating through sample quality , author=. International Conference on Learning Representations , year=

  40. [40]

    arXiv preprint arXiv:2404.14507 , year=

    Align your steps: Optimizing sampling schedules in diffusion models , author=. arXiv preprint arXiv:2404.14507 , year=

    Pith/arXiv arXiv

  41. [41]

    arXiv preprint arXiv:2405.11326 , year=

    On the trajectory regularity of ode-based diffusion sampling , author=. arXiv preprint arXiv:2405.11326 , year=

    Pith/arXiv arXiv

  42. [42]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Accelerating diffusion sampling with optimized time steps , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  43. [43]

    arXiv preprint arXiv:2405.15506 , year=

    Learning to Discretize Denoising Diffusion ODEs , author=. arXiv preprint arXiv:2405.15506 , year=

    Pith/arXiv arXiv

  44. [44]

    arXiv preprint arXiv:2403.18035 , year=

    Bidirectional consistency models , author=. arXiv preprint arXiv:2403.18035 , year=

    Pith/arXiv arXiv

  45. [45]

    arXiv preprint arXiv:2406.07507 , year=

    Flow Map Matching , author=. arXiv preprint arXiv:2406.07507 , year=

    Pith/arXiv arXiv

  46. [46]

    arXiv preprint arXiv:2210.02747 , year=

    Flow matching for generative modeling , author=. arXiv preprint arXiv:2210.02747 , year=

    Pith/arXiv arXiv

  47. [47]

    arXiv preprint arXiv:2209.03003 , year=

    Flow straight and fast: Learning to generate and transfer data with rectified flow , author=. arXiv preprint arXiv:2209.03003 , year=

    Pith/arXiv arXiv

  48. [48]

    arXiv preprint arXiv:2303.08797 , year=

    Stochastic interpolants: A unifying framework for flows and diffusions , author=. arXiv preprint arXiv:2303.08797 , year=

    Pith/arXiv arXiv

  49. [49]

    arXiv preprint arXiv:2410.12557 , year=

    One step diffusion via shortcut models , author=. arXiv preprint arXiv:2410.12557 , year=

    Pith/arXiv arXiv

  50. [50]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Fast ode-based sampling for diffusion models in around 5 steps , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  51. [51]

    arXiv preprint arXiv:1503.02531 , year=

    Distilling the knowledge in a neural network , author=. arXiv preprint arXiv:1503.02531 , year=

    Pith/arXiv arXiv

  52. [52]

    The Twelfth International Conference on Learning Representations , year=

    Instaflow: One step is enough for high-quality diffusion-based text-to-image generation , author=. The Twelfth International Conference on Learning Representations , year=

  53. [53]

    Advances in neural information processing systems , volume=

    Generative adversarial nets , author=. Advances in neural information processing systems , volume=

  54. [54]

    Forty-first International Conference on Machine Learning , year=

    Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation , author=. Forty-first International Conference on Machine Learning , year=

  55. [55]

    arXiv preprint arXiv:2410.19310 , year=

    Flow generator matching , author=. arXiv preprint arXiv:2410.19310 , year=

    Pith/arXiv arXiv

  56. [56]

    arXiv preprint arXiv:2105.14080 , year=

    Gotta go fast when generating data with score-based models , author=. arXiv preprint arXiv:2105.14080 , year=

    Pith/arXiv arXiv

  57. [57]

    arXiv preprint arXiv:2106.00132 , year=

    On fast sampling of diffusion probabilistic models , author=. arXiv preprint arXiv:2106.00132 , year=

    Pith/arXiv arXiv

  58. [58]

    arXiv preprint arXiv:2201.06503 , year=

    Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models , author=. arXiv preprint arXiv:2201.06503 , year=

    Pith/arXiv arXiv

  59. [59]

    arXiv preprint arXiv:2206.05564 , year=

    gddim: Generalized denoising diffusion implicit models , author=. arXiv preprint arXiv:2206.05564 , year=

    Pith/arXiv arXiv

  60. [60]

    arXiv preprint arXiv:2204.13902 , year=

    Fast sampling of diffusion models with exponential integrator , author=. arXiv preprint arXiv:2204.13902 , year=

    Pith/arXiv arXiv

  61. [61]

    Advances in Neural Information Processing Systems , volume=

    Unipc: A unified predictor-corrector framework for fast sampling of diffusion models , author=. Advances in Neural Information Processing Systems , volume=

  62. [62]

    Advances in Neural Information Processing Systems , volume=

    Parallel sampling of diffusion models , author=. Advances in Neural Information Processing Systems , volume=

  63. [63]

    Forty-first International Conference on Machine Learning , year=

    Accelerating parallel sampling of diffusion models , author=. Forty-first International Conference on Machine Learning , year=

  64. [64]

    arXiv preprint arXiv:2206.07309 , year=

    Estimating the optimal covariance with imperfect mean in diffusion probabilistic models , author=. arXiv preprint arXiv:2206.07309 , year=

    Pith/arXiv arXiv

  65. [65]

    Advances in Neural Information Processing Systems , volume=

    Genie: Higher-order denoising diffusion solvers , author=. Advances in Neural Information Processing Systems , volume=

  66. [66]

    2009 , publisher=

    Learning multiple layers of features from tiny images , author=. 2009 , publisher=

    2009

  67. [67]

    2009 IEEE conference on computer vision and pattern recognition , pages=

    Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

    2009

  68. [68]

    Advances in neural information processing systems , volume=

    Gans trained by a two time-scale update rule converge to a local nash equilibrium , author=. Advances in neural information processing systems , volume=

  69. [69]

    arXiv preprint arXiv:2009.09761 , year=

    Diffwave: A versatile diffusion model for audio synthesis , author=. arXiv preprint arXiv:2009.09761 , year=

    Pith/arXiv arXiv 2009

  70. [70]

    arXiv preprint arXiv:2301.12503 , year=

    Audioldm: Text-to-audio generation with latent diffusion models , author=. arXiv preprint arXiv:2301.12503 , year=

    Pith/arXiv arXiv

  71. [71]

    International Conference on Machine Learning , pages=

    Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  72. [72]

    arXiv preprint arXiv:2207.12598 , year=

    Classifier-free diffusion guidance , author=. arXiv preprint arXiv:2207.12598 , year=

    Pith/arXiv arXiv

  73. [73]

    arXiv preprint arXiv:2303.04248 , year=

    Tract: Denoising diffusion models with transitive closure time-distillation , author=. arXiv preprint arXiv:2303.04248 , year=

    Pith/arXiv arXiv

  74. [74]

    arXiv preprint arXiv:2411.08378 , year=

    Physics informed distillation for diffusion models , author=. arXiv preprint arXiv:2411.08378 , year=

    Pith/arXiv arXiv

  75. [75]

    Advances in Neural Information Processing Systems , volume=

    One-step diffusion distillation through score implicit matching , author=. Advances in Neural Information Processing Systems , volume=

  76. [76]

    arXiv preprint arXiv:2503.07565 , year=

    Inductive Moment Matching , author=. arXiv preprint arXiv:2503.07565 , year=

    Pith/arXiv arXiv

  77. [77]

    arXiv preprint arXiv:2405.02730 , year=

    U-dits: Downsample tokens in u-shaped diffusion transformers , author=. arXiv preprint arXiv:2405.02730 , year=

    Pith/arXiv arXiv

  78. [78]

    Advances in neural information processing systems , volume=

    Visual autoregressive modeling: Scalable image generation via next-scale prediction , author=. Advances in neural information processing systems , volume=

  79. [79]

    Advances in Neural Information Processing Systems , volume=

    Autoregressive image generation without vector quantization , author=. Advances in Neural Information Processing Systems , volume=

  80. [80]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Maskgit: Masked generative image transformer , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

Showing first 80 references.