Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Music Boomerang: Reusing Diffusion Models for Data Augmentation and Audio Manipulation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Music Boomerang adapts image-domain local sampling to audio, turning Stable Audio Open into a label-preserving data augmenter and a text-controlled instrument-swap tool.

desk verdict First audio adaptation of Boomerang sampling with Stable Audio Open; solid limited-data augmentation results, but the forward noise equation looks inconsistent and needs checking against code before the numbers are fully trusted. read the letter →

arxiv 2507.04864 v1 pith:OY6SYUMG submitted 2025-07-07 cs.SD cs.LGeess.AS

Alexander Fichtinger , Jan Schlüter , Gerhard Widmer This is my paper · ORCID
classification cs.SDcs.LGeess.AS
keywords diffusionmodelsdataaugmentationbeattrackingaudiomanipulationBoomerangsamplinglatentmusicgenerationinstrumentreplacement
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

Music Boomerang adapts a diffusion-sampling trick from images to audio: start from a real recording, add a controlled amount of noise, and let a pretrained music diffusion model denoise it back to a slightly changed version. The paper argues this produces near-identical rhythmic structure with fresh timbral and micro-structural variation, making it useful as data augmentation for beat and downbeat tracking, and that text prompts can steer it to replace instruments on monophonic recordings. The evidence is strongest when training data is scarce: combining Boomerang variations with standard pitch, tempo, and mask augmentations lifts beat F1 from 79.8 to 82.5 and downbeat F1 from 64.1 to 67.8 on the reduced training set. A sympathetic reader would take away that an off-the-shelf latent diffusion model can act as a label-preserving augmentation source and a light-touch audio editor without retraining.

What carries the argument

The load-bearing mechanism is partial forward diffusion followed by partial reverse diffusion inside a pretrained latent diffusion model: an input waveform is encoded to latents, corrupted to a chosen noise level $n_{Boom}$ (mapped to a timestep $t_{Boom}$), then denoised back to a reconstruction $x'_0$ under text conditioning. For recordings longer than the model's 47-second limit, the pipeline processes overlapping windows and freezes the 25% overlap region during reverse diffusion to maintain temporal coherence. This machinery converts a global generator into a local sampler, and the noise level becomes the single dial that trades variation against preservation of rhythmic structure.

What would settle it

Take a set of recordings with human-verified beat annotations, apply Boomerang augmentation at 40% noise, and have human annotators or a high-accuracy tracker re-annotate the transformed audio; if the re-annotated beat positions differ from the original annotations by more than the typical augmentation variability, the claim that Boomerang is label-preserving for beat tracking fails.

Watch

Extended reading notes

Core claim

The central claim is that Boomerang sampling, implemented for Stable Audio Open, preserves the rhythmic skeleton of an input recording while generating local, uncontrolled variations, and that this behavior can be exploited in two ways. First, as a data augmentation tool: training a state-of-the-art beat tracker with Boomerang-generated variations of each training file, particularly when combined with conventional pitch shifting, time stretching, and masking, improves beat and downbeat F1 scores on the GTZAN test set, but only in the limited-data regime (the gain shrinks to marginal on the full data). Second, as a content manipulation tool: changing the text prompt that guides the reverse diffusion can replace a solo instrument's timbre (e.g., guitar to trumpet lead) while preserving harmonic structure, though polyphonic mixes are altered globally rather than per voice. The authors report that onset and beat preservation degrades as the noise level increases, with 40% noise striking the best balance for augmentation, and that rhythmic preservation is strongest for genres with clear rhythmic patterns such as reggae, disco, and pop.

Load-bearing premise

Rhythmic preservation is measured with an automatic onset and beat detector applied to the original and transformed audio, treating the original detections as ground truth, so the reported F1 scores partly reflect tracker stability rather than true musical rhythm preservation; the augmentation experiments then use Boomerang-augmented files with original beat annotations without verifying those annotations still match the changed audio.

Editorial extensions

If this is right

  • Combining Boomerang with standard augmentations raises beat F1 to 82.5 and downbeat F1 to 67.8 on the reduced dataset, versus 79.8 and 64.1 respectively for standard augmentations alone.
  • Boomerang sampling is cheaper than global sampling, costing roughly $t_{Boom}/T$ of the denoising steps, which makes it about five times faster at $t_{Boom}=0.2T$.
  • Text prompts can replace a solo instrument's timbre while keeping the harmonic framework, and the effect applies globally to polyphonic mixes rather than to individual voices.
  • The technique mainly helps in limited-data regimes; on the full dataset the improvement over the state of the art is marginal.
  • The authors publish code and audio examples so other tasks can try Boomerang augmentation and further applications.

Reading between the lines

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

  • Because Boomerang perturbs timbre and micro-structure while keeping rhythm, it may generalize as an augmentation engine for other rhythm- or structure-sensitive MIR tasks, such as chord estimation or structural segmentation, where label-preserving transformations are hard to engineer by hand.
  • The frozen-overlap windowing suggests a general recipe for stitching any fixed-length latent diffusion model to arbitrary-length inputs, which could enable diffusion-based editing of full-length recordings beyond the 47-second window.
  • An untested extension is prompt diversity: replacing the neutral 'Music' prompt with varied semantic descriptions could enlarge the augmentation manifold and possibly improve the limited-data gains further, as the authors note as future work.
  • The marginal full-data result hints that Boomerang's value is in regularizing small datasets, so its best use may be as a complement to, not a replacement for, conventional augmentations.
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

3 major / 5 minor

Summary. The paper adapts Boomerang sampling, a partial-diffusion local sampling technique from the image domain, to audio using the pretrained Stable Audio Open model. It proposes a pipeline that encodes audio into latent space, partially noises the latent, and partially denoises it, optionally with text prompting. The authors evaluate rhythmic preservation on GTZAN using librosa onset/beat detection as a proxy, use Boomerang-generated variations as data augmentation for beat and downbeat tracking on reduced and full training sets, and demonstrate qualitative text-based instrument replacement on monophonic inputs. The central claims are that rhythmic structure is mostly preserved, beat tracking improves under limited training data, and monophonic instrument replacement is feasible.

Significance. If the central claims hold, the paper demonstrates a practical way to reuse a pretrained audio diffusion model as a label-preserving data augmentation source and as a light-touch audio editor without retraining, which is of clear interest to the MIR and generative-audio communities. The paper ships a public implementation, selects hyperparameters on a validation split, and evaluates on a held-out test set with human beat annotations, which are concrete strengths. However, the validity of the quantitative claims depends on the correctness of the forward perturbation formula and on the reliability of the rhythmic-preservation proxy, so the significance is conditional on addressing those points.

major comments (3)
  1. [§4.2.2] The forward perturbation is written as z_tBoom = z0 + sigma_tBoom * epsilon, with no alpha_tBoom * z0 term. This is inconsistent with the v-objective forward process stated in §4.1.3, where x_t = alpha_t x0 + sigma_t epsilon. Unless alpha_t = 1 for the chosen schedule, which is not the case for Stable Audio Open's schedule, the resulting latent does not correspond to any diffusion timestep and lies off the model's training manifold. The noise-level sweeps in Tables 1 and 2 and the optimal nBoom = 40% are therefore not interpretable as partial-diffusion Boomerang sampling. Please correct the equation to include alpha_tBoom * z0, verify that the released code implements the corrected form, and rerun or re-interpret the affected experiments if the code matches the current text.
  2. [§5.1.1 and §5.2.1] The rhythmic-preservation evaluation treats onset/beat detections in the original audio as ground truth and detections in the Boomerang-transformed audio as predictions, using librosa for both. This measures tracker stability, not necessarily preservation of the true rhythmic structure; the authors themselves note that jazz and classical are difficult for automatic beat trackers. Because the augmentation experiments in §5.2.1 use Boomerang-augmented files with the original beat annotations without any direct check that the annotations still match the transformed audio, the central data-augmentation claim inherits this proxy. Please add at least a small human-annotated or manual verification of beat positions in transformed examples, or otherwise demonstrate that annotation mismatch is negligible.
  3. [Table 4 and §5.2.2] The full-dataset comparison borrows the 'pitch, tempo, masks' results from Foscarin et al. [29]. If the model, training schedule, data splits, and evaluation code are not all identical to those used for the other conditions, differences in implementation could account for the small observed gains. Please specify precisely which components are reused and which are re-run, or rerun the baseline under the same codebase to make the comparison apples-to-apples.
minor comments (5)
  1. [§3.1 and §4.2.2] The notation for the forward process is inconsistent: §3.1 uses sqrt(alpha_bar) x0 plus noise, while §4.2.2 omits the scaling; aligning these formulas would help readers see the intended connection to the diffusion schedule.
  2. [References] Reference [28] is listed as 'Swing ratio estimation' but is cited for the GTZAN genre dataset; please verify and correct the citation, as the standard GTZAN genre dataset reference is different.
  3. [§5.2.2] The paper reports means and standard deviations over three runs but does not report statistical significance tests; given the small differences in Table 4, a significance test or confidence intervals would make the 'marginal improvement' conclusion more robust.
  4. [§5.1.1] The paper states that a guidance scale of 1.0 and a negative prompt 'Low Quality' are used, but it does not specify how Stable Audio Open's classifier-free guidance and negative prompting are implemented; please add the exact inference settings for reproducibility.
  5. [§3.2] The computational cost claim that Boomerang sampling is approximately tBoom/T times cheaper assumes a linear relationship between timestep and solver steps; with DPM-Solver++ the step allocation and cost may differ, so please clarify the accounting.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the augmentation results are held-out empirical measurements, and the flagged equation mismatch in Sec. 4.2.2 is a correctness/consistency issue, not an input-to-output reduction.

full rationale

The paper's central claims are empirical rather than derivational. Boomerang-augmented training sets are evaluated against held-out human beat and downbeat annotations on GTZAN, with hyperparameters (noise level and number of variations) selected on a validation split in Table 2 before the test evaluation in Tables 3 and 4. No fitted parameter defines the quantity being predicted, and no result is obtained by substituting a definition into itself. The rhythm-preservation proxy in Sec. 5.1.1 compares librosa detections in original and transformed audio, and the authors explicitly acknowledge that this depends on the effectiveness of the detection algorithms; this is a validity limitation, not circularity. The citation of Foscarin et al. [29], which overlaps with the present authors, is used as a baseline and as a source of default augmentations, not to justify the Boomerang results, so it is not load-bearing. The one internal inconsistency is Eq. 4.2.2, which writes z_tBoom = z0 + sigma_tBoom epsilon while Sec. 4.1.3 defines the v-objective with x_t = alpha_t x0 + sigma_t epsilon and Sec. 3.1 gives the standard sqrt(alpha_bar_t) x0 scaling; if the released code matches the text, the perturbation is off-manifold. That is a correctness issue that could affect the noise-level sweep, but it does not make the results equivalent to their inputs by construction, which is the standard for circularity here.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central empirical claims rest on the pretrained Stable Audio Open model, on hyperparameters chosen by validation, and on two unverified proxies: detector-based rhythm preservation and annotation validity for augmented audio. No new theoretical entities are introduced.

free parameters (5)
  • Boomerang noise level nBoom = 40%
    Selected via validation search in Table 2; central to the augmentation results. At 20%, 60%, and 80% the method was less effective or worse than no augmentation.
  • Number of Boomerang variations per training file = 6
    Chosen from the validation search in Table 2 as the best setting, though most settings were close.
  • Text guidance scale = 1.0
    Set to favor audio input over prompt in the augmentation experiments; higher guidance is used qualitatively for instrument replacement.
  • Overlap percentage for long audio windows = 25%
    A design choice introduced in Sec. 4.2.4 to maintain boundary coherence; not systematically optimized.
  • Training epochs = 400
    Increased from 100 in Foscarin et al. to counter the reduced dataset size in the hyperparameter search.
assumptions (5)
  • domain assumption Stable Audio Open's autoencoder and diffusion model are faithful enough on musical audio that Boomerang resampling preserves content structure.
    Used throughout Sec. 4.2; if the autoencoder introduces large reconstruction artifacts, augmented files may not retain the original beat structure.
  • domain assumption Beat and downbeat annotations remain valid for Boomerang-augmented audio.
    Sec. 5.2 trains on augmented files with original annotations; the paper does not directly verify annotation validity on augmented samples.
  • domain assumption Librosa-based onset and beat detection is a valid proxy for rhythmic preservation.
    Sec. 5.1.1 treats original detections as ground truth; tracker failure on jazz and classical could distort the reported preservation scores.
  • domain assumption Foscarin et al.'s published baseline results are comparable to this paper's runs in the full-dataset setting.
    Table 4 borrows the 'Pitch, tempo, masks' row from Foscarin et al. rather than rerunning under this paper's exact training pipeline.
  • standard math The diffusion forward and reverse equations used in the pipeline are correctly implemented following Stable Audio Open's scheduler.
    Sec. 3 states the forward and reverse process formulas; the paper does not re-derive them but relies on them for the implementation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Music Boomerang: Reusing Diffusion Models for Data Augmentation and Audio Manipulation." pith.science (2026). https://pith.science/paper/OY6SYUMG

@misc{pith2026250704864,
  author       = {Pith},
  title        = {Pith review of: Music Boomerang: Reusing Diffusion Models for Data Augmentation and Audio Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OY6SYUMG}},
  note         = {Machine review of arXiv:2507.04864}
}
read the original abstract

Generative models of music audio are typically used to generate output based solely on a text prompt or melody. Boomerang sampling, recently proposed for the image domain, allows generating output close to an existing example, using any pretrained diffusion model. In this work, we explore its application in the audio domain as a tool for data augmentation or content manipulation. Specifically, implementing Boomerang sampling for Stable Audio Open, we augment training data for a state-of-the-art beat tracker, and attempt to replace musical instruments in recordings. Our results show that the rhythmic structure of existing examples is mostly preserved, that it improves performance of the beat tracker, but only in scenarios of limited training data, and that it can accomplish text-based instrument replacement on monophonic inputs. We publish our implementation to invite experiments on data augmentation in other tasks and explore further applications.

Figures

Figures reproduced from arXiv: 2507.04864 by the authors.

Figure 1
Figure 1. (a) Diffusion models can gradually transform [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) The first window of a longer piece is sampled [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 31 canonical work pages

  1. [29]

    DPM-solver++: Fast solver for guided sampling of diffusion probabilistic models,

    C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, “DPM-solver++: Fast solver for guided sampling of diffusion probabilistic models,” 2023. [Online]. Available: https://openreview.net/forum?id= 4vGwQqviud5

    work page 2023

  2. [1]

    In the field of music generation, these models have to capture complex distributions that define elements such as melody, rhythm and harmony

    INTRODUCTION Generative models are designed to learn the underlying probability distribution of data, then generate new samples that match the characteristics of the training dataset. In the field of music generation, these models have to capture complex distributions that define elements such as melody, rhythm and harmony. Diffusion models [1] in particu...

    work page 2025

  3. [2]

    RELA TED WORK For context and background, we will give a brief overview on modern strands of generative audio models and on con- trolled sampling from diffusion models. 2.1 Audio Generation using Deep Learning Deep learning has significantly advanced audio genera- tion, with models evolving from early autoregressive ap- proaches to modern diffusion-based ...

    work page Pith review arXiv 2025

  4. [3]

    The underlying principle transfers natu- rally to the audio domain, and in this work we explore its application to music audio recordings

    BOOMERANG SAMPLING Boomerang sampling [5] is a local sampling technique originally introduced for image generation with pretrained diffusion models. The underlying principle transfers natu- rally to the audio domain, and in this work we explore its application to music audio recordings. The core idea is to apply a suitable noise level 𝑛Boom for perturbing...

    work page

  5. [4]

    This process must be carried out step by step, exactly as in global sampling of diffusion models, and this is where most of the computational effort occurs

    The reverse pro- cess follows standard stochastic diffusion-based sampling, where each step is modelled as a conditional Gaussian: 𝑝𝜃(x𝑡−1| x𝑡) =𝒩 (𝜇𝜃(x𝑡, 𝑡), 𝜎2 𝑡 I) where 𝜇𝜃 represents the learned mean function, and 𝜎𝑡 de- fines the noise schedule of the reverse process. This process must be carried out step by step, exactly as in global sampling of dif...

    work page

  6. [5]

    MUSIC BOOMERANG To apply Boomerang sampling to audio generation, we in- tegrate it with the publicly available Stable Audio Open model [14]. In the following, we first provide a brief overview of the model, explaining its key components and functionality, then explain in detail how this model is used in the Boomerang sampling pipeline. 4.1 Stable Audio Op...

    work page

  7. [6]

    Music”) and negative prompt (“Low Quality

    EXPERIMENTS We conduct three experiments with Boomerang sampling on music recordings via Stable Audio Open, investigating whether it 1) preserves the rhythmic structure of a record- ing, 2) is suitable as data augmentation for beat tracking, and 3) enables prompt-based content manipulation. 5.1 Analysis of Rhythmic Structure Preservation Our first goal is...

    work page

  8. [7]

    Our results show that it is suitable for data augmentation in training a beat and downbeat detector, but more so if only limited train- ing data is available

    CONCLUSIONS In this work, we adapted Boomerang sampling [5] to the audio domain by implementing it for a pretrained diffu- sion model, Stable Audio Open [14]. Our results show that it is suitable for data augmentation in training a beat and downbeat detector, but more so if only limited train- ing data is available. We also show that by guiding the sampli...

    work page 2020

Show all 37 references
  1. [8]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Advances in Neural Information Processing Systems , vol. 33, 2020, pp. 6840–6851. [Online]. Available: https://proceedings.neurips.cc/paper_files/paper/2020/ file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf

    2020

  2. [9]

    Long-form music generation with latent diffusion,

    Z. Evans, J. D. Parker, C. Carr, Z. Zukowski, J. Taylor, and J. Pons, “Long-form music generation with latent diffusion,” arXiv preprint arXiv:2404.10301, 2024

    2024 arXiv

  3. [10]

    Grad-TTS: A diffusion probabilistic model for text-to-speech,

    V . Popov, I. V ovk, V . Gogoryan, T. Sadekova, and M. Kudinov, “Grad-TTS: A diffusion probabilistic model for text-to-speech,” in Proceedings of the 38th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, vol. 139, 2021, pp. 8599–860...

    2021

  4. [11]

    AudioLDM: Text- to-audio generation with latent diffusion models,

    H. Liu, Z. Chen, Y . Yuan, X. Mei, X. Liu, D. Mandic, W. Wang, and M. D. Plumbley, “AudioLDM: Text- to-audio generation with latent diffusion models,” in Proceedings of the 40th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, vol. ...

    2023

  5. [12]

    Boomerang: Local sampling on image manifolds using dif- fusion models,

    L. Luzi, P. M. Mayer, J. Casco-Rodriguez, A. Siahkoohi, and R. Baraniuk, “Boomerang: Local sampling on image manifolds using dif- fusion models,” Transactions on Machine Learn- ing Research , 2024. [Online]. Available: https: //openreview.net/forum?id=NYdThkjNW1

    2024

  6. [13]

    WaveNet: A generative model for raw audio,

    A. v. d. Oord, “WaveNet: A generative model for raw audio,” arXiv preprint arXiv:1609.03499, 2016

    2016 arXiv

  7. [14]

    Jukebox: A generative model for music,

    P. Dhariwal, H. Jun, C. Payne, J. W. Kim, A. Radford, and I. Sutskever, “Jukebox: A generative model for music,” arXiv preprint arXiv:2005.00341, 2020

    2005 arXiv

  8. [15]

    Musika! Fast infinite waveform music generation,

    M. Pasini and J. Schlüter, “Musika! Fast infinite waveform music generation,” in Proceedings of the 23rd International Society for Music Information Retrieval Conference , 2022, pp. 543–550. [Online]. Available: https://doi.org/10.5281/zenodo.7316720

    2022 doi

  9. [16]

    Masked audio generation using a single non-autoregressive transformer,

    A. Ziv, I. Gat, G. L. Lan, T. Remez, F. Kreuk, J. Copet, A. Défossez, G. Synnaeve, and Y . Adi, “Masked audio generation using a single non-autoregressive transformer,” in International Conference on Learning Representations, 2024. [Online]. Available: https: //openreview.net/...

    2024

  10. [17]

    High fidelity neural audio compression,

    A. Défossez, J. Copet, G. Synnaeve, and Y . Adi, “High fidelity neural audio compression,” Transactions on Machine Learning Research , 2023. [Online]. Avail- able: https://openreview.net/forum?id=ivCd8z8zR2

    2023

  11. [18]

    Fast timing-conditioned latent audio diffusion,

    Z. Evans, C. Carr, J. Taylor, S. H. Hawley, and J. Pons, “Fast timing-conditioned latent audio diffusion,” in Proceedings of the 41st International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, vol. 235. PMLR, 2024, pp. 12 652–12 665. [Online]...

    2024

  12. [19]

    U-net: Con- volutional networks for biomedical image segmenta- tion,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Con- volutional networks for biomedical image segmenta- tion,” in 18th International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015, pp. 234–241

    2015

  13. [20]

    Moûsai: Efficient text-to-music diffusion models,

    F. Schneider, O. Kamal, Z. Jin, and B. Schölkopf, “Moûsai: Efficient text-to-music diffusion models,” in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics , 2024, pp. 8050–8068. [Online]. Available: https://aclanthology. org/2024.acl-long.437

    2024

  14. [21]

    Stable audio open,

    Z. Evans, J. D. Parker, C. Carr, Z. Zukowski, J. Tay- lor, and J. Pons, “Stable audio open,” arXiv preprint arXiv:2407.14358, 2024

    2024 arXiv

  15. [22]

    SDEdit: Guided image synthesis and editing with stochastic differential equations,

    C. Meng, Y . He, Y . Song, J. Song, J. Wu, J.- Y . Zhu, and S. Ermon, “SDEdit: Guided image synthesis and editing with stochastic differential equations,” in International Conference on Learning Representations, 2022. [Online]. Available: https: //openreview.net/forum?id=aBsCjcPu_tE

    2022

  16. [23]

    Come-closer-diffuse- faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction,

    H. Chung, B. Sim, and J.-C. Ye, “Come-closer-diffuse- faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction,” IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 12 403–12 412,

  17. [24]

    RoFormer: Enhanced transformer with rotary position embedding,

    J. Su, M. Ahmed, Y . Lu, S. Pan, W. Bo, and Y . Liu, “RoFormer: Enhanced transformer with rotary position embedding,” Neurocomput- ing, vol. 568, p. 127063, 2024. [Online]. Avail- able: https://www.sciencedirect.com/science/article/ pii/S0925231223011864

    2024

  18. [25]

    Truncated diffusion probabilistic models and diffusion-based ad- versarial auto-encoders,

    H. Zheng, P. He, W. Chen, and M. Zhou, “Truncated diffusion probabilistic models and diffusion-based ad- versarial auto-encoders,” in International Conference on Learning Representations , 2023. [Online]. Avail- able: https://openreview.net/forum?id=HDxgaKk956l

    2023

  19. [26]

    Consistency models,

    Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency models,” in Proceedings of the 40th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, vol. 202. PMLR, 2023, pp. 32 211–32 252. [Online]. Available: https://proceedings.ml...

    2023

  20. [27]

    Score-based generative modeling through stochastic differen- tial equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Ku- mar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differen- tial equations,” in International Conference on Learning Representations, 2021. [Online]. Available: https://openreview.net/forum?id=PxTIG12RRHS

    2021

  21. [28]

    DPM-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps,

    L. Cheng, Z. Yuhao, B. Fan, C. Jianfei, L. Chongxuan, and Z. Jun, “DPM-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps,” in Advances in Neural Information Processing Systems , vol. 35, 2022. [Online]. Available: https://proceedings.neuri...

    2022

  22. [30]

    Exploring the limits of transfer learning with a unified text-to-text transformer,

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,” J. Mach. Learn. Res., vol. 21, no. 1, Jan. 2020. [Online]. Available: https://jmlr.org/papers/v...

    2020

  23. [31]

    Neural networks fail to learn periodic functions and how to fix it,

    L. Ziyin, T. Hartwig, and M. Ueda, “Neural networks fail to learn periodic functions and how to fix it,” in Advances in Neural Information Processing Systems , vol. 33, 2020, pp. 1583–1594. [Online]. Available: https://proceedings.neurips.cc/paper_files/paper/2020/ file/116045...

    2020

  24. [33]

    Progressive distil- lation for fast sampling of diffusion mod- els,

    T. Salimans and J. Ho, “Progressive distil- lation for fast sampling of diffusion mod- els,” in International Conference on Learn- ing Representations , 2022. [Online]. Available: https://openreview.net/forum?id=TIdIXIpzhoI

    2022

  25. [34]

    librosa: Audio and music signal analysis in python,

    B. McFee, C. Raffel, D. Liang, D. P. Ellis, M. McVicar, E. Battenberg, and O. Nieto, “librosa: Audio and music signal analysis in python,” in Proceedings of the 14th Python in Science Conference, vol. 8, 2015

    2015

  26. [35]

    mir_eval: A transparent implementation of common MIR metrics,

    C. Raffel, B. McFee, E. J. Humphrey, J. Salamon, O. Nieto, D. Liang, D. P. Ellis, and C. C. Raffel, “mir_eval: A transparent implementation of common MIR metrics,” inProceedings of the 15th International Society of Music Information Retrieval Conference , 2014

    2014

  27. [36]

    Swing ratio estimation,

    U. Marchand and G. Peeters, “Swing ratio estimation,” in Proceedings of the 18th International Conference on Digital Audio Effects, 2015

    2015

  28. [37]

    Beat this! Accurate beat tracking without DBN postprocessing,

    F. Foscarin, J. Schlüter, and G. Widmer, “Beat this! Accurate beat tracking without DBN postprocessing,” in Proceedings of the 25th International Society for Music Information Retrieval Conference , Nov. 2024. [Online]. Available: https://doi.org/10.5281/zenodo. 14877491

    2024 doi

  29. [2022]

    Available: https://doi.org/10.1109/ CVPR52688.2022.01209

    [Online]. Available: https://doi.org/10.1109/ CVPR52688.2022.01209

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.