Pith. sign in

REVIEW 2 major objections 4 minor 61 references

Audio synthesizer inversion in symmetric parameter spaces with approximately equivariant flow matching

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read When a synthesizer has symmetries, inversion is best solved by learning an invariant density, not by regression.

desk verdict A useful, well-engineered paper on synthesizer inversion with a novel learned equivariance relaxation, let down mainly by an architecture confound in the key comparison and a prior assumption that the real-world data only approximately meets. read the letter →

arxiv 2506.07199 v1 pith:E5OOWF3B submitted 2025-06-08 cs.SD cs.LGeess.ASeess.SP

Ben Hayes , Charalampos Saitis , György Fazekas This is my paper
classification cs.SDcs.LGeess.ASeess.SP
keywords synthesizerinversionsoundmatchingpermutationsymmetryequivariantflowcontinuousnormalizingflowsgenerativemodelsSurgeXTaudioreconstruction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that the ill-posedness of synthesizer inversion — the same sound produced by many parameter settings — is largely caused by intrinsic symmetries, especially permutations among repeated components like oscillators or filters. It claims regression models that output one point estimate cannot handle this: they average over equivalent solutions and trigger the responsibility problem, a pathology where similar inputs demand very different outputs because responsibility for a target switches between set elements. The paper's solution is to treat inversion as conditional density estimation, using flow matching with a permutation-equivariant continuous normalizing flow so the learned parameter distribution is itself invariant to the synthesizer's symmetries. For real synthesizers with conditional or approximate symmetries, it introduces a learnable parameter-to-token assignment that discovers which symmetries to respect from data. On a synthetic oscillator task and on the Surge XT synthesizer, the equivariant generative model reconstructs audio more accurately than regression and symmetry-naive generative baselines.

What carries the argument

The load-bearing identity is the orbital factorization of the conditional parameter density, p(x|y)=p(O|y)p(g|O,y)η(O), which reduces inversion to orbit-level density estimation once the symmetry posterior is uniform. The mechanism that implements it is a permutation-equivariant continuous normalizing flow trained with flow matching, whose vector field is a Transformer without positional encoding so that each token sees the parameters of one repeated unit. To handle real synthesizers where symmetries are conditional or approximate, the paper adds Param2Tok, a learned sparse assignment from parameters to tokens, initialized near full permutation invariance and regularized with an L1 penalty; it can keep or break equivariance per parameter group as the data dictates.

What would settle it

Train the equivariant flow on a Surge XT dataset whose parameter prior is deliberately symmetry-breaking — for example, always set LFO 1's amplitude to zero while sampling LFOs 2 through 5 uniformly — and compare against a non-equivariant generative model of matched size; if the equivariant model no longer wins on audio reconstruction metrics, the uniform-symmetry-posterior assumption is the step that carries the result.

Watch

Extended reading notes

Core claim

The paper establishes that if a synthesizer has a symmetry group G acting on its parameter space, the posterior p(x|y) factorizes into an orbit posterior p(O|y), a symmetry posterior p(g|O,y), and a stabilizer factor; under a G-invariant prior and likelihood the symmetry posterior is uniform, so the whole problem reduces to learning a G-invariant density over orbits. It then shows how to do this with a continuous normalizing flow whose vector field is a permutation-equivariant Transformer, and introduces Param2Tok, a learned assignment matrix that maps parameters to tokens and can break equivariance where the data requires. The empirical claim is that this combination outperforms both point-estimate regression and non-equivariant generative baselines on audio reconstruction from the Surge XT synthesizer, and that the learned assignment matrices visibly organize parameters into the synthesizer's symmetric subsystems.

Load-bearing premise

The claim that the symmetry posterior p(g|O,y) is uniform, and hence that learning an invariant density is exactly right, rests on the assumption that the training data's parameter prior is invariant under the synthesizer's symmetries; the Surge XT datasets violate this by sampling some parameters from default-heavy or truncated distributions.

Editorial extensions

If this is right

  • Permutation symmetry in a synthesizer degrades point-estimate regression, even with permutation-invariant losses or sorting heuristics; the k-osc experiments show this across k=4, 8, 16, and 32 oscillators.
  • A conditional generative model that learns the parameter distribution rather than a point estimate substantially improves audio reconstruction under symmetry.
  • Making the generative model's density permutation-invariant — via an equivariant continuous normalizing flow — improves reconstruction further than a symmetry-naive generative model on in-domain audio.
  • Param2Tok discovers from data which symmetries to respect and which to break, so it matches the best model in both symmetric and asymmetric variants of the toy task.
  • On Surge XT, the equivariant flow beats the AST regression baseline and the VAE+RealNVP generative baseline across MSS, wMFCC, SOT, and RMS metrics, and the ordering persists on out-of-distribution audio.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If this reasoning extends beyond synthesizers, any neural network inverting an external system with structural symmetries — audio effects, physical simulators, or set-structured control tasks — should prefer generative density estimation over point prediction when symmetries are present.
  • The Param2Tok discovery mechanism suggests a broadly applicable inductive bias: initialize a model to be maximally symmetric and let training break equivariance only where the likelihood demands it, which may improve sample efficiency in other set-to-set prediction problems.
  • A testable next step the paper does not run is isolating quasi-symmetries in a k-osc-style task; the authors note this gap, and the outcome would clarify whether the Surge XT gains come from handling approximate symmetries or just from the generative formulation.
  • The paper's decision to evaluate in the audio domain rather than parameter space implies that published sound-matching comparisons relying only on parameter distances should be re-examined, because such metrics can penalize valid symmetric solutions while rewarding incorrect ones.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper addresses the ill-posedness of synthesizer inversion caused by symmetries, focusing on permutation invariance. It derives a factorization of the conditional parameter posterior into orbit and symmetry parts, shows that under a G-invariant prior the symmetry posterior is uniform, and advocates learning G-invariant densities via equivariant continuous normalizing flows. A relaxed equivariance mechanism, Param2Tok, is introduced to discover symmetries from data. Experiments on a synthetic k-oscillator task and on Surge XT show that generative models outperform regression baselines and that the proposed equivariant CNFs achieve the best audio reconstruction metrics.

Significance. If the equivariance benefit is established, the paper gives a principled framework for a practically important problem and connects synthesizer inversion to equivariant generative modeling. The synthetic k-osc task cleanly controls for symmetry, the Surge XT evaluation is large-scale and uses multiple audio metrics, and the authors provide code and audio examples. The theoretical derivation of the orbital factorization is explicit and self-contained. However, the central empirical claim that learning the invariant density, rather than simply using a more expressive architecture, drives the gains is currently not separated from an architecture confound.

major comments (2)
  1. [4.1.3 / Table 1 / Appendix B.1.2] The comparison between equivariant CNFs (CNF (Equivariant), CNF (Param2Tok)) and the non-equivariant CNF (MLP) simultaneously varies two factors: permutation equivariance and the presence of self-attention. CNF (MLP) is described as having the attention blocks removed, leaving a residual MLP, while the equivariant models are Transformers with self-attention. Thus the reported gains of the equivariant models over CNF (MLP) could be due to attention or token-based processing rather than to permutation equivariance. To support claim (ii) of the conclusion, the paper should include a non-equivariant DiT baseline of comparable capacity (e.g., a DiT with positional encodings that breaks permutation symmetry) and report the k-osc and Surge XT metrics for it.
  2. [3.2 / Appendix B.3] Param2Tok relies on several hand-chosen mechanisms: the near-invariant initialization of A, Z and Z prime, the L1 sparsity penalty on A, and the guidance scale at inference. The manuscript states these are necessary for good performance but provides no ablations or sensitivity analysis. Without ablation, it is unclear whether the improved results of CNF (Param2Tok) come from symmetry discovery, from the regularization, or from the initialization. A simple ablation on the k-osc task (e.g., varying the L1 weight or using a random initialization) would substantially strengthen the paper's central claims about relaxed equivariance.
minor comments (4)
  1. [4.2.3] The text says results are presented in Table 2, but the audio reconstruction results are in Table 1; Table 2 is the parameter table in Appendix E.
  2. [3 / Appendix E] The uniform posterior derivation in Section 3 and Appendix A.3 assumes a G-invariant prior, and the text says this is satisfied by random sampling. Appendix E, however, introduces default-biased and truncated sampling. It would help to state explicitly that these modifications are applied identically across permutable parameter groups, so the joint prior remains exchangeable on those groups despite the non-uniform marginals.
  3. [Throughout] The typographic rendering PA R A M2TO K is difficult to read; consider using Param2Tok consistently in the text and figures.
  4. [4.2.3] The VAE + RealNVP baseline collapsed to predicting average values. While the paper notes this and offers a plausible explanation, the collapse may indicate an implementation or tuning issue rather than a fundamental limitation of the method. Since the code is released, a brief verification or a reference to a reproduction script would increase confidence in the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the theoretical derivation is self-contained and the empirical claims are tested against external baselines and audio metrics.

full rationale

The paper's derivation chain is not circular. The orbital factorization of p(x|y) in Section 3 and Appendix A follows from standard probability and group-theoretic identities, with the uniformity of p(g|O,y) derived from explicit assumptions of G-invariance of the likelihood and prior; it does not assume the conclusion. The invariance property of equivariant continuous normalizing flows is imported from the external Köhler et al. result, not from the authors' own prior work. The only self-citation, the responsibility-problem work [9], motivates the regression pathology but is accompanied by an independent citation [10] and is independently demonstrated by the in-paper k-osc experiments; it is not load-bearing for the central generative-equivariance claim. The empirical comparisons use held-out audio reconstruction metrics against regression and generative baselines, so no prediction reduces by construction to a fitted input. Limitations do exist but are correctness risks, not circularity: the CNF (Equivariant) versus CNF (MLP) comparison confounds equivariance with Transformer attention and token-based processing; the Surge XT sampling scheme in Appendix E violates the G-invariant-prior assumption; and the conclusion acknowledges the lack of theoretical guarantees that PARAM2TOK discovers symmetries. None of these involve an equation or fitted parameter being renamed as a prediction, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the uniform-symmetry-posterior assumption, which in turn depends on the G-invariance of the training prior. The method itself introduces learned matrices A and Z that are not theoretically guaranteed to converge to the true symmetry. Several hyperparameters are hand-tuned, and the Surge XT data-generation procedure uses non-uniform distributions that violate a key assumption.

free parameters (4)
  • Classifier-free guidance scale = 2.0
    Selected via experimentation; used for all CNF models in both tasks.
  • Param2Tok sparse assignment L1 penalty weight = 0.01
    Chosen by hand; needed to encourage sparse token assignment and symmetry discovery.
  • Param2Tok initialization variances for A and Z = sigma^2_A = sigma^2_Z = 1e-4
    Hand-set to start near permutation-invariant; deemed necessary for reliable symmetry discovery.
  • Surge XT sampling probabilities for default values = e.g., 50% for LFO amplitude 0 and waveshaper type None
    These non-uniform sampling choices break exact G-invariance of the training prior, conflicting with an assumption in Section 3.
assumptions (4)
  • domain assumption G-invariance of the likelihood p(y|x) for the synthesizer's symmetry group G.
    Section 3: 'First, we assume G-invariance of the likelihood p(y|x) which arises naturally from the symmetry of our synthesizer.'
  • domain assumption G-invariance of the prior p(x), claimed to be satisfied by random sampling from G-invariant distributions.
    Section 3; Appendix E sampling details show non-uniform and truncated distributions, so this is not always satisfied exactly.
  • standard math Finite group G, product of permutation subgroups, with orbit decomposition valid except for low-probability non-trivial stabilizers.
    Used in Appendix A.2 derivation; the authors assume non-trivial stabilizers have very low probability.
  • ad hoc to paper Existence of a parameter-to-token assignment (Param2Tok) that can represent the relevant conditional symmetries.
    Only an existence proof is given in Appendix A.4, not a training guarantee. The paper states 'we currently lack theoretical guarantees that Param2Tok will discover symmetries.'

how reviews work

0 comments
Cite this review

Pith. "Pith review of Audio synthesizer inversion in symmetric parameter spaces with approximately equivariant flow matching." pith.science (2026). https://pith.science/paper/E5OOWF3B

@misc{pith2026250607199,
  author       = {Pith},
  title        = {Pith review of: Audio synthesizer inversion in symmetric parameter spaces with approximately equivariant flow matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E5OOWF3B}},
  note         = {Machine review of arXiv:2506.07199}
}
read the original abstract

Many audio synthesizers can produce the same signal given different parameter configurations, meaning the inversion from sound to parameters is an inherently ill-posed problem. We show that this is largely due to intrinsic symmetries of the synthesizer, and focus in particular on permutation invariance. First, we demonstrate on a synthetic task that regressing point estimates under permutation symmetry degrades performance, even when using a permutation-invariant loss function or symmetry-breaking heuristics. Then, viewing equivalent solutions as modes of a probability distribution, we show that a conditional generative model substantially improves performance. Further, acknowledging the invariance of the implicit parameter distribution, we find that performance is further improved by using a permutation equivariant continuous normalizing flow. To accommodate intricate symmetries in real synthesizers, we also propose a relaxed equivariance strategy that adaptively discovers relevant symmetries from data. Applying our method to Surge XT, a full-featured open source synthesizer used in real world audio production, we find our method outperforms regression and generative baselines across audio reconstruction metrics.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 46 canonical work pages

  1. [1]

    Analytic Signal Processing in Music Computation

    J. Justice. “Analytic Signal Processing in Music Computation”. In:IEEE Transactions on Acoustics, Speech, and Signal Processing27.6 (Dec. 1979), pp. 670–684.D O I: 10 . 1109 / TASSP . 1979 . 1163321

    work page 1979

  2. [2]

    responsibility

    BACKGROUND 2.1 Synthesizer inversion & sound matching Given an audio signal, the sound matching task aims to find a synthesizer parameter configuration that best approximates it [4, 5]. We focus in this paper onsynthesizer inversion, a sub-task of sound matching in which the audio signal we seek to approximate is knowna priorito have come from the synthes...

    work page

  3. [3]

    A synthesizer is a map between these spaces, f:P → S

    METHOD Let P ⊂Rk be the space of synthesizer parameters 1 and S ⊂Rn be the space of audio signals. A synthesizer is a map between these spaces, f:P → S. It is common that f is not injective. That is, there exist multiple sets of pa- rameters, e.g. x(1),x (2) ∈ P, that produce the same signal, i.e. f(x (1)) =f(x (2)). A trivial example is given when the sy...

    work page

  4. [4]

    par- tial

    EXPERIMENTS We train CNF models with the rectified flow probability path [33]. A minibatch approximation to the optimal trans- port coupling [34] is implemented using the Hungarian algorithm. Conditioning dropout is applied with a proba- bility of 10% and inference is performed with classifier- free guidance [41, 42] with a scale of 2.0. Sampling is perfo...

    work page

  5. [5]

    This extends beyond synthesizers, as audio effects also commonly exhibit permutation symmetries, as noted by Nercessian [22]

    CONCLUSION The implications of our findings are clear: if the synthesizer has a symmetry, it is better to (i) approach the problem generatively and (ii) learn the corresponding invariant den- sity. This extends beyond synthesizers, as audio effects also commonly exhibit permutation symmetries, as noted by Nercessian [22]. Beyond audio, these results are o...

    work page

  6. [6]

    In particular, our choice of VST synthesizer reflects a bias towards western popular music

    ETHICS STATEMENT Like any AI model, our work inherently encodes the biases and values of the authors. In particular, our choice of VST synthesizer reflects a bias towards western popular music. However, the more abstract nature of our synthetic exper- imentation does suggest that our results may reasonably be expected to generalize to tools that better re...

    work page

  7. [7]

    ACKNOWLEDGEMENTS B.H. would like to thank Christopher Mitcheltree, Marco Pasini, Chin-Yun Yu, Jack Loth, and Julien Guinot for their invaluable feedback on this manuscript in varying stages of completion, and Jordie Shier for the many inspiring and illuminating conversations on this topic. This research was supported by UK Research and Innova- tion [grant...

    doi:10.5281/zenodo.438045

  8. [8]

    Synthesizer Sound Matching Using Audio Spectrogram Transformers

    Fred Bruford, Frederik Blang, and Shahan Nerces- sian. “Synthesizer Sound Matching Using Audio Spectrogram Transformers”. In:Proceedings of the 27th International Conference on Digital Audio Ef- fects. DAFx24. Guildford, Surrey, Sept. 3–7, 2024

    work page 2024

Show all 61 references
  1. [9]

    Speech Analy- sis/Synthesis Based on a Sinusoidal Representation

    R. McAulay and T. Quatieri. “Speech Analy- sis/Synthesis Based on a Sinusoidal Representation”. In:IEEE Transactions on Acoustics, Speech, and Signal Processing34.4 (Aug. 1986), pp. 744–754. D O I:10.1109/TASSP.1986.1164910

    1986

  2. [10]

    Spectral Model- ing Synthesis: A Sound Analysis/Synthesis System Based on a Deterministic Plus Stochastic Decompo- sition

    Xavier Serra and Julius Smith. “Spectral Model- ing Synthesis: A Sound Analysis/Synthesis System Based on a Deterministic Plus Stochastic Decompo- sition”. In:Computer Music Journal14.4 (1990), pp. 12–24.D O I:10.2307/3680788

    1990 doi

  3. [11]

    A Comparison of Parametric Optimization Techniques for Musical Instrument Tone Matching

    Martin Roth and Matthew Yee-King. “A Comparison of Parametric Optimization Techniques for Musical Instrument Tone Matching”. In: Audio Engineering Society Convention 130. Audio Engineering Society, May 13, 2011

    2011

  4. [12]

    The Synthesizer Programming Prob- lem: Improving the Usability of Sound Synthesizers

    Jordie Shier. “The Synthesizer Programming Prob- lem: Improving the Usability of Sound Synthesizers”. MA thesis. University of Victoria, 2021

    2021

  5. [13]

    InverSynth: Deep Estimation of Synthesizer Parameter Configurations from Audio Signals

    Oren Barkan et al. “InverSynth: Deep Estimation of Synthesizer Parameter Configurations from Audio Signals”. In:IEEE/ACM Transactions on Audio, Speech, and Language Processing27.12 (Dec. 2019), pp. 2385–2396.D O I: 10.1109/TASLP.2019. 2944568. arXiv:1812.06349

    2019 arXiv

  6. [14]

    Improving Semi-Supervised Differentiable Synthesizer Sound Matching for Practical Applications

    Naotake Masuda and Daisuke Saito. “Improving Semi-Supervised Differentiable Synthesizer Sound Matching for Practical Applications”. In:IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing31 (2023), pp. 863–875.D O I: 10.1109/ TASLP.2023.3237161

    2023

  7. [15]

    DDSP: Differentiable Digital Signal Processing

    Jesse Engel et al. “DDSP: Differentiable Digital Signal Processing”. In:8th International Conference on Learning Representations. ICLR 2020. Addis Ababa, Ethiopia, Apr. 2020

    2020

  8. [16]

    The Responsibility Problem in Neural Networks with Unordered Targets

    Ben Hayes, Charalampos Saitis, and György Fazekas. “The Responsibility Problem in Neural Networks with Unordered Targets”. In:The First Tiny Papers Track at ICLR 2023. ICLR. Kigali, Rwanda, May 5, 2023

    2023

  9. [17]

    FSPool: Learning Set Representations with Featurewise Sort Pooling

    Yan Zhang, Jonathon Hare, and Adam Prügel- Bennett. “FSPool: Learning Set Representations with Featurewise Sort Pooling”. In: International Conference on Learning Representations. Sept. 23, 2019

    2019

  10. [18]

    Flow Synthesizer: Universal Audio Synthesizer Control with Normalizing Flows

    Philippe Esling et al. “Flow Synthesizer: Universal Audio Synthesizer Control with Normalizing Flows”. In:Applied Sciences10.1 (Dec. 2020), p. 302.D O I: 10.3390/app10010302

    2020 doi

  11. [19]

    Improving Synthesizer Programming From Variational Autoencoders Latent Space

    Gwendal Le Vaillant, Thierry Dutoit, and Sebastien Dekeyser. “Improving Synthesizer Programming From Variational Autoencoders Latent Space”. In: 2021 24th International Conference on Digital Au- dio Effects (DAFx). 2021 24th International Con- ference on Digital Audio Effects ...

    2021

  12. [20]

    Equivari- ant Flows: Exact Likelihood Generative Learning for Symmetric Densities

    Jonas Köhler, Leon Klein, and Frank Noe. “Equivari- ant Flows: Exact Likelihood Generative Learning for Symmetric Densities”. In:Proceedings of the 37th International Conference on Machine Learning. In- ternational Conference on Machine Learning. PMLR, Nov. 21, 2020, pp. 5361–5370

    2020

  13. [21]

    AST: Audio Spectrogram Transformer

    Yuan Gong, Yu-An Chung, and James Glass. “AST: Audio Spectrogram Transformer”. In: Proc. Inter- speech 2021. 2021, pp. 571–575.D O I: 10.21437/ Interspeech.2021-698

    2021

  14. [22]

    Neural Parametric Equalizer Matching Using Differentiable Biquads

    Shahan Nercessian. “Neural Parametric Equalizer Matching Using Differentiable Biquads”. In:Pro- ceedings of the 23rd International Conference on Digital Audio Effects. DAFx2020. Vienna, Austria, 2020, p. 8

    2020

  15. [23]

    A Review of Differentiable Digital Signal Processing for Music & Speech Synthesis

    Ben Hayes et al. “A Review of Differentiable Digital Signal Processing for Music & Speech Synthesis”. In:Frontiers in Signal Processing(2023)

    2023

  16. [24]

    Synthesizer Sound Matching with Differentiable DSP

    Naotake Masuda and Daisuke Saito. “Synthesizer Sound Matching with Differentiable DSP” (On- line). Nov. 7, 2021.D O I: 10 . 5281 / zenodo . 5624609

    2021

  17. [25]

    Noy Uzrad et al.DiffMoog: A Differentiable Modu- lar Synthesizer for Sound Matching. Jan. 23, 2024. D O I: 10.48550/arXiv.2401.12570 . arXiv: 2401.12570 [eess]. Pre-published

    2024 doi arXiv

  18. [26]

    White Box Search over Audio Synthesizer Parameters

    Yuting Yang et al. “White Box Search over Audio Synthesizer Parameters”. In:Proc. of the 24rd Int. Society for Music Information Retrieval Conf.ISMIR. Milan, Italy, 2023

    2023

  19. [27]

    Learning to Solve Inverse Problems for Perceptual Sound Matching

    Han Han, Vincent Lostanlen, and Mathieu Lagrange. “Learning to Solve Inverse Problems for Perceptual Sound Matching”. In:IEEE/ACM Transactions on Audio, Speech, and Language Processing32 (2024), pp. 2605–2615.D O I: 10.1109/TASLP.2024. 3393738

    2024 doi

  20. [28]

    Self-Supervised Pitch Detection by Inverse Audio Synthesis

    Jesse Engel, Rigel Swavely, and Adam Roberts. “Self-Supervised Pitch Detection by Inverse Audio Synthesis”. In:Proceedings of the International Con- ference on Machine Learning. ICML 2020. 2020, p. 9

    2020

  21. [29]

    Neural Ordinary Differential Equations

    Ricky T. Q. Chen et al. “Neural Ordinary Differential Equations”. In:Advances in Neural Information Processing Systems. V ol. 31. Curran Associates, Inc., 2018

    2018

  22. [30]

    Set Prediction without Imposing Structure as Conditional Density Estimation

    David W. Zhang, Gertjan J. Burghouts, and Cees G. M. Snoek. “Set Prediction without Imposing Structure as Conditional Density Estimation”. In: In- ternational Conference on Learning Representations. Jan. 12, 2021

    2021

  23. [31]

    Deep Set Prediction Networks

    Yan Zhang, Jonathon Hare, and Adam Prugel- Bennett. “Deep Set Prediction Networks”. In:Ad- vances in Neural Information Processing Systems. V ol. 32. Curran Associates, Inc., 2019

    2019

  24. [32]

    Conditional Set Generation with Trans- formers

    Adam R. Kosiorek, Hyunjik Kim, and Danilo J. Rezende. “Conditional Set Generation with Trans- formers”. In:Workshop on Object-Oriented Learn- ing. International Conference on Machine Learning. arXiv, July 1, 2020. arXiv:2006.16841 [cs]

    2020 arXiv

  25. [33]

    Jinwoo Kim et al.SetVAE: Learning Hierarchi- cal Composition for Generative Modeling of Set- Structured Data. Mar. 29, 2021.D O I:10.48550/ arXiv . 2103 . 15619. arXiv: 2103 . 15619 [cs]. Pre-published

    2021

  26. [34]

    Scalable Nor- malizing Flows for Permutation Invariant Densities

    Marin Biloš and Stephan Günnemann. “Scalable Nor- malizing Flows for Permutation Invariant Densities”. In:Proceedings of the 38th International Conference on Machine Learning. International Conference on Machine Learning. PMLR, July 1, 2021, pp. 957– 967

    2021

  27. [35]

    Fast Point Cloud Generation with Straight Flows

    Lemeng Wu et al. “Fast Point Cloud Generation with Straight Flows”. In:2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Vancouver, BC, Canada: IEEE, June 2023, pp. 9445–9454....

    2023

  28. [36]

    Equivariant Flow Matching for Molecular Conformer Generation

    Majdi Hassan et al. “Equivariant Flow Matching for Molecular Conformer Generation”. In: ICML’24 Workshop ML for Life and Material Science: From Theory to Industry Applications. July 17, 2024

    2024

  29. [37]

    FFJORD: Free-Form Contin- uous Dynamics for Scalable Reversible Generative Models

    Will Grathwohl et al. “FFJORD: Free-Form Contin- uous Dynamics for Scalable Reversible Generative Models”. In: International Conference on Learning Representations. Sept. 27, 2018

    2018

  30. [38]

    Improving and Generaliz- ing Flow-Based Generative Models with Minibatch Optimal Transport

    Alexander Tong et al. “Improving and Generaliz- ing Flow-Based Generative Models with Minibatch Optimal Transport”. In:Transactions on Machine Learning Research(2024)

    2024

  31. [39]

    Flow Matching for Generative Modeling

    Yaron Lipman et al. “Flow Matching for Generative Modeling”. In: The Eleventh International Confer- ence on Learning Representations. Feb. 1, 2023

    2023

  32. [40]

    In practice, we find it necessary to initialize Z,Z ′, and A such that PA R A M2TO Kis approximately invariant to any permutation of the parameter vector

    of the assignment matrix and another set of learned vectors Z′ ∈R k×d as follows: ˜x= Z′ ⊙ AT Xl 1d,(4) where 1d is simply a d-dimensional vector of ones and Xl is the output of the lth Transformer layer. In practice, we find it necessary to initialize Z,Z ′, and A such that P...

  33. [41]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. “Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow”. In: The Eleventh International Conference on Learning Representa- tions. Feb. 1, 2023

    2023

  34. [42]

    Multisample Flow Matching: Straightening Flows with Minibatch Cou- plings

    Aram-Alexandre Pooladian et al. “Multisample Flow Matching: Straightening Flows with Minibatch Cou- plings”. In:Proceedings of the 40th International Conference on Machine Learning. International Con- ference on Machine Learning. PMLR, July 3, 2023, pp. 28100–28127

    2023

  35. [43]

    Yuxuan Song et al.Equivariant Flow Matching with Hybrid Probability Transport. Dec. 12, 2023. D O I: 10.48550/arXiv.2312.07168 . arXiv: 2312.07168 [cs]. Pre-published

    2023 doi arXiv

  36. [44]

    Equiv- ariant Flow Matching

    Leon Klein, Andreas Krämer, and Frank Noe. “Equiv- ariant Flow Matching”. In:Advances in Neural In- formation Processing Systems36 (Dec. 15, 2023), pp. 59886–59910

    2023

  37. [45]

    Attention Is All You Need

    Ashish Vaswani et al. “Attention Is All You Need”. In:Advances in Neural Information Processing Sys- tems. V ol. 30. Curran Associates, Inc., 2017

    2017

  38. [46]

    Scalable Dif- fusion Models with Transformers

    William Peebles and Saining Xie. “Scalable Dif- fusion Models with Transformers”. In:2023 IEEE/CVF International Conference on Computer Vision (ICCV). 2023 IEEE/CVF International Confer- ence on Computer Vision (ICCV). Paris, France: IEEE, Oct. 1, 2023, pp. 4172–4182.D O I: 10...

    2023

  39. [47]

    Using the Output Em- bedding to Improve Language Models

    Ofir Press and Lior Wolf. “Using the Output Em- bedding to Improve Language Models”. In:Pro- ceedings of the 15th Conference of the European Chapter of the Association for Computational Lin- guistics: Volume 2, Short Papers. Proceedings of the 15th Conference of the European C...

    2017 doi

  40. [48]

    Classifier-Free Diffusion Guidance

    Jonathan Ho and Tim Salimans. “Classifier-Free Diffusion Guidance”. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Appli- cations. Dec. 8, 2021

    2021

  41. [49]

    Qinqing Zheng et al.Guided Flows for Genera- tive Modeling and Decision Making. Dec. 7, 2023. D O I: 10.48550/arXiv.2311.13443 . arXiv: 2311.13443 [cs]. Pre-published

    2023 doi arXiv

  42. [50]

    Per- ceptually Informed Synthesis of Bandlimited Classi- cal Waveforms Using Integrated Polynomial Inter- polation

    Vesa Välimäki, Jussi Pekonen, and Juhan Nam. “Per- ceptually Informed Synthesis of Bandlimited Classi- cal Waveforms Using Integrated Polynomial Inter- polation”. In:The Journal of the Acoustical Society of America131.1 (Jan. 1, 2012), pp. 974–986.D O I: 10.1121/1.3651227

    2012 doi

  43. [51]

    Parametric Correspondence and Chamfer Matching: Two New Techniques for Image Matching

    Harry G Barrow et al. “Parametric Correspondence and Chamfer Matching: Two New Techniques for Image Matching”. In:Proceedings: Image Under- standing Workshop. Science Applications, Inc, 1977, pp. 21–27

    1977

  44. [52]

    Version 0.7.3

    Peter Sobot.Pedalboard. Version 0.7.3. Zen- odo, Apr. 10, 2023.D O I: 10 . 5281 / ZENODO . 7817838

    2023

  45. [53]

    On Layer Normalization in the Transformer Architecture

    Ruibin Xiong et al. “On Layer Normalization in the Transformer Architecture”. In:Proceedings of the 37th International Conference on Machine Learn- ing. International Conference on Machine Learning. PMLR, Nov. 21, 2020, pp. 10524–10533

    2020

  46. [54]

    Ian Dunn and David Ryan Koes.Mixed Continuous and Categorical Flow Matching for 3D De Novo Molecule Generation. Apr. 30, 2024.D O I: 10 . 48550 / arXiv . 2404 . 19739. arXiv: 2404 . 19739 [q-bio]. Pre-published

    2024

  47. [55]

    I’m Sorry for Your Loss: Spectrally-Based Audio Distances Are Bad at Pitch

    Joseph Turian and Max Henry. “I’m Sorry for Your Loss: Spectrally-Based Audio Distances Are Bad at Pitch”. Dec. 9, 2020. arXiv: 2012.04572 [cs, eess]

    2020 arXiv

  48. [56]

    Unsupervised Harmonic Parameter Estimation Us- ing Differentiable DSP and Spectral Optimal Trans- port

    Bernardo Torres, Geoffroy Peeters, and Gaël Richard. “Unsupervised Harmonic Parameter Estimation Us- ing Differentiable DSP and Spectral Optimal Trans- port”. In:ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). ICASSP 2024...

    2024

  49. [57]

    Neural Audio Synthesis of Musical Notes with WaveNet Autoencoders

    Jesse Engel et al. “Neural Audio Synthesis of Musical Notes with WaveNet Autoencoders”. In:Proceedings of the 34th International Conference on Machine Learning - Volume 70. ICML’17. Sydney, Australia, Aug. 6, 2017, pp. 1068–1077

    2017

  50. [58]

    FSD50K: An Open Dataset of Human-Labeled Sound Events

    Eduardo Fonseca et al. “FSD50K: An Open Dataset of Human-Labeled Sound Events”. In:IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing30 (2022), pp. 829–852.D O I: 10.1109/ TASLP.2021.3133208

    2022

  51. [59]

    Density Estimation Using Real NVP

    Laurent Dinh, Jascha Sohl-Dickstein, and Samy Ben- gio. “Density Estimation Using Real NVP”. In: International Conference on Learning Representa- tions. Feb. 6, 2017

    2017

  52. [60]

    Fast and Flexible Neural Audio Synthesis

    Lamtharn Hantrakul et al. “Fast and Flexible Neural Audio Synthesis”. In:Proceedings of the 20th In- ternational Society for Music Information Retrieval Conference. ISMIR. Delft, The Netherlands, 2019, pp. 524–530

    2019

  53. [61]

    Crepe: A Convolutional Representation for Pitch Estimation

    Jong Wook Kim et al. “Crepe: A Convolutional Representation for Pitch Estimation”. In:2018 IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2018 - Proceedings. 2018 IEEE International Conference on Acoustics, Speech, and Signal Processing, ICAS...

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.