Pith. sign in

REVIEW 4 major objections 4 minor 63 references

Zero-shot editing of real music recordings is made practical by reusing inversion velocities to seed a high-order ODE solver.

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 →

T0 review · deepseek-v4-flash

2026-08-01 17:42 UTC pith:UIN7LREA

load-bearing objection The AB3 claim doesn't survive contact with Eq. (14), but the paper is a useful empirical study of a possibly-effective heuristic—worth refereeing if the authors can reposition it honestly. the 4 major comments →

arxiv 2607.17526 v1 pith:UIN7LREA submitted 2026-07-20 cs.SD cs.LGcs.NAeess.ASeess.SPmath.NA

FlowSonic: Stable Zero-Shot Music Editing via High-Order Trajectory Integration

classification cs.SD cs.LGcs.NAeess.ASeess.SPmath.NA
keywords music editingrectified flowzero-shotODE solverAdams-Bashforthdynamic history cachingcross-attention injection
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.

The paper sets out to show that a pretrained rectified-flow text-to-music model can edit real-world recordings — changing timbre or genre — without any training, fine-tuning, or per-sample optimization. The route is a two-stage process: deterministic inversion of the audio into the model's latent space, with cross-attention features and velocity evaluations cached along the way; then generation under a new prompt that reuses the cached attention features to preserve structure and seeds a third-order Adams-Bashforth solver with the cached inversion velocities via Dynamic History Caching. The paper argues that this seeding removes a numerical discontinuity at the inversion-generation boundary, allowing high-order integration to start at the very first step. Reported results show consistent gains over existing zero-shot editing methods across semantic alignment, harmonic preservation, structural consistency, and perceptual quality, and the whole recipe requires no model modification or extra optimization.

Core claim

The central claim is that the failure mode of inversion-based rectified-flow editing is largely a startup problem in numerical integration. Multi-step solvers like third-order Adams-Bashforth need derivative history that does not exist at the beginning of generation, so conventional pipelines fall back to lower-order warm-start steps; this introduces a 'Multi-Step Startup Asymmetry' whose errors propagate through the entire trajectory. FlowSonic's Dynamic History Caching closes that gap by taking the final two velocity evaluations from the deterministic inversion and feeding them directly into the AB3 formula for the first generation step. Combined with reuse of cached key-value attention te

What carries the argument

The two load-bearing mechanisms are the third-order Adams-Bashforth multi-step ODE integrator (AB3), which computes the next latent state from the current velocity plus two previous velocities, and Dynamic History Caching (DHC), which supplies those previous velocities from the inversion stage instead of a lower-order warm-start. DHC makes the first generation step third-order accurate and removes the startup asymmetry between inversion and generation. Structural preservation is carried separately by reusing cached key-value cross-attention tensors during generation, with full key-value replacement giving the best balance of fidelity and editability.

Load-bearing premise

Velocities cached under the source-prompt inversion are treated as valid Adams-Bashforth history for a generation trajectory that runs under a different prompt and in the reversed time direction.

What would settle it

Run the editing pipeline with the cached history taken from a different recording's inversion, or with the two cached velocities exchanged in time order; if edit quality and structural preservation do not degrade, the claimed trajectory-specific benefit of Dynamic History Caching is not the operative mechanism.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A pretrained rectified-flow model can be converted into an editor of arbitrary recordings by adding a caching pass, with no retraining or per-sample optimization.
  • The seeding strategy should transfer to other multi-step solvers and to future rectified-flow models, as the paper states.
  • Most of the perceptual improvement comes from the numerical seeding alone, with attention injection adding a smaller complementary gain.
  • The gains are obtained at the same computational cost as the original pipeline, since DHC reuses velocity evaluations that inversion already computed.

Where Pith is reading between the lines

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

  • Whether the cached velocities need to come from the same recording is testable: replacing them with velocities from another song's inversion (or reversing their order) and measuring whether edit quality drops would isolate the trajectory-specific vs generic-smoothing role of DHC.
  • The seeding crosses a direction reversal and a prompt change, so the classic AB3 error bounds cited in the paper do not directly apply; a numerical check of the local truncation error at the first seeded step on the target trajectory would clarify how far the heuristic departs from the theory.
  • The same inversion-history seeding could be carried over to image and video editing with rectified flow, where the same warm-start problem appears.
  • Comparing against a fourth-order single-step solver (e.g., classical Runge-Kutta) at matched cost would separate the benefit of higher order from the benefit of reusing inversion-specific history.

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

4 major / 4 minor

Summary. The paper presents FlowSonic, a zero-shot text-guided music editing framework built on a pretrained rectified-flow transformer (FluxMusic). The method deterministically inverts a real-world recording into the latent space, preserves musical structure by reusing cached cross-attention key/value representations during generation, and introduces a 'Seeded-AB3' solver with Dynamic History Caching (DHC). DHC initializes the third-order Adams–Bashforth integrator using velocity evaluations obtained during inversion, allegedly avoiding lower-order warm-start steps and improving numerical stability. The paper reports experiments on timbre-transfer and genre-transfer tasks, comparing against AudioLDM2, MusicGen, ZETA, and a FluxMusic editing baseline, and claims consistent improvements in semantic alignment, structural preservation, and perceptual quality. Code is released publicly.

Significance. If the numerical claims were sound, DHC would be an attractive, training-free way to initialize high-order multi-step solvers in inversion-based editing, with no extra network evaluations. The paper also contributes a systematic comparison of integration schemes and a reproducible codebase. However, the central theoretical claim—that Eq. (14) is a third-order AB3 step with O(h^4) local error—is invalid because the cached velocities come from a different vector field and a reversed integration path. This flaw undermines the main novel contribution, leaving the attention-injection mechanism as a known adaptation from image editing. The empirical support is further weakened by a per-solver CFG equalization protocol, a small subjective study, and an overclaim contradicted by the paper's own FAD results.

major comments (4)
  1. [§III-C, Eq. (14)] The update z1 = z0 + Δt[(23/12)v0 − (16/12)v_K_inv + (5/12)v_{K−1}_inv] is not a third-order Adams–Bashforth step. AB3 requires the three latest derivatives evaluated at preceding points on the same solution curve (t0, t0−h, t0−2h) of the same vector field. The cached v_K_inv and v_{K−1}_inv are computed on the source-prompt inversion path in the reverse direction with source CFG=1, not on the target-prompt forward trajectory. The local truncation error O(h^4) therefore does not follow, and the claim that DHC 'preserves third-order accuracy' is unsupported. Since the solver is the main novel contribution, this is load-bearing. The authors should either provide a correct multi-step error analysis for the proposed update or remove the order guarantees and reframe DHC as a heuristic.
  2. [§IV / §V-B, Equitonal Transfer] The Equitonal Transfer protocol adjusts the CFG scale per solver so that all configurations achieve 'comparable' CLAP similarity. This makes the CLAP differences in Tables I and II uninformative by construction and introduces a per-solver free parameter that is not reported. Without the adjusted CFG values, the comparison is not reproducible, and the claim that Seeded AB3 achieves the 'highest CLAP similarity' (Table I, 0.238 vs 0.236) is not meaningful evidence. Please report the CFG values used for each solver and analyze sensitivity, or use a protocol that does not equalize the headline semantic metric.
  3. [Abstract / §V-C, Table III] The claim that FlowSonic 'consistently outperforms existing music editing methods across ... perceptual audio quality' is contradicted by the paper's own FAD results: AudioLDM2 achieves the lowest FAD (3.623) on the timbre-transfer task, beating FlowSonic KV (3.938) and V (3.887). The text acknowledges AudioLDM2's low FAD, but the abstract and conclusion still assert consistent superiority. Please qualify the claims and discuss why the proposed method trails on this metric.
  4. [§V-D and Appendix A] The subjective evaluation uses 20 participants, each rating one timbre and one genre example per condition, and reports only mean scores without variance or significance testing. Differences such as MOS-T 4.00 vs 4.10 are within sampling noise. The claims of 'dramatic improvement' and 'consistently' are not statistically supported. Please add more participants/items, report score distributions, and use inferential statistics (e.g., paired tests with multiple-comparison correction).
minor comments (4)
  1. [§III-C, Eq. (12)] The notation t_{−1}, t_{−2} and the integration grid are undefined. Please specify the grid spacing and the exact time indices of the cached inversion velocities used in Eq. (14).
  2. [Figures 2 and 11] The PCA components explain only 8.1%/3.7% and 11.5%/4.4% of the variance, respectively. Please report the total explained variance and note that two-dimensional projections may not faithfully represent trajectory stability.
  3. [§V-A, Figure 5] The caption describes a comparison of numerical integration strategies, but the text says the figure examines cross-attention feature injection. Please align the caption with the actual content.
  4. [References] References [20] and [34] are cited for supervised editing and MusicGen, but [20] is M2UGen and [34] is MusicGen-stem. Please ensure the citations match the statements they support.

Circularity Check

1 steps flagged

No load-bearing circularity; one self-referential evaluation step: the Equitonal Transfer protocol calibrates CFG to CLAP and then reports CLAP as a headline metric.

specific steps
  1. fitted input called prediction [Section III-C (Equitonal Transfer protocol), reported in Section V-B, Tables I-II]
    "To fairly compare different numerical integration strategies, we further introduce the Equitonal Transfer protocol. ... We therefore adjust the CFG scale slightly for each solver so that all methods achieve a comparable level of semantic transformation, measured using CLAP similarity."

    The CFG scale is adjusted per solver to equalize CLAP similarity, and the same CLAP metric is then reported as evidence that Seeded AB3 'achieves the highest CLAP similarity (0.238)'. The reported CLAP differences are therefore constrained by the calibration target rather than being an independent measurement of semantic editing strength. This makes the CLAP column in Tables I-II partly an output of the normalization protocol rather than an independent prediction. The effect is mild because the adjustment is described as slight and the main contributions are also supported by Chroma, CQT-PCC, FAD, and MOS.

full rationale

The derivation chain of FlowSonic is otherwise self-contained: deterministic inversion integrates the pretrained rectified-flow field, attention reuse follows cited image-editing methods, and the claimed gains are empirical comparisons on external audio metrics. The Eq. (14) Dynamic History Caching step is a real numerical-consistency concern, but it is not circularity: reusing inversion velocities as AB3 history is an unsupported identification of a different vector field (source-prompt, CFG=1, reverse-time) with the generation trajectory, not a prediction that equals its own input by construction. The PCA trajectory visualization is illustrative post hoc evidence, not a fitted parameter. No load-bearing self-citation chain appears; references to the pretrained model and to prior feature-injection work are external. The only self-referential element is the Equitonal Transfer protocol, which uses CLAP to calibrate CFG and then reports CLAP among the headline metrics; this is a mild evaluation circularity and does not invalidate the structural-preservation or perceptual results. Score 2 reflects that minor self-referential protocol, not a circular derivation.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central method rests on a small set of hand-set hyperparameters (K, CFG, injection schedule, IB, M, n) that are tuned on the evaluation tasks, plus several domain assumptions about ODE regularity, inversion accuracy, and metric validity. The most consequential item is the ad hoc assumption that source-prompt inversion velocities constitute valid multistep history for the target-prompt generation ODE — this is the load-bearing premise of DHC and is not justified in the paper.

free parameters (6)
  • number of diffusion timesteps K = 25
    Chosen empirically in Section IV-E as a 'favorable compromise' versus the original FluxMusic's 50 steps.
  • target CFG scale = 5-15 per task/solver
    Selected empirically; Equitonal Transfer further adjusts CFG per solver to equalize CLAP similarity.
  • source CFG scale = 1
    Fixed to 1 'yields stable inversion' (Section IV-E).
  • injection steps = not specified
    Number of timesteps at which cached KV features are injected is tuned (Figure 3) but the exact schedule used for final results is not reported.
  • injection block count (IB) = not specified
    IB count is tuned as a transferability-fidelity trade-off knob (Figure 3); the final value is not stated.
  • cached blocks M and cached timesteps n = not specified
    Section III-F says 'last M single-stream blocks over the final n inversion timesteps' are cached, but M and n are never given.
axioms (5)
  • ad hoc to paper Cached inversion velocities are valid Adams-Bashforth history for the target-conditioned generation ODE.
    Eq. (14) uses v_K_inv and v_{K-1}_inv from the source-prompt inversion as f_{-1}, f_{-2}. AB3 consistency requires history from the same vector field on the same forward trajectory; no proof is provided that this holds when the prompt and integration direction change.
  • domain assumption The rectified-flow learned velocity field is smooth and straight enough for AB3's local truncation error analysis to apply at 25 steps.
    The paper asserts O(h^4) local error without verifying regularity of v_theta or step-size-dependent behavior in the mel-latent space.
  • domain assumption Deterministic Euler inversion of 10-second real recordings yields a latent z0 sufficiently accurate for downstream editing.
    The inversion section acknowledges Euler inversion has accumulating errors, yet the editing pipeline assumes the resulting z0 is a valid starting point for the target-prompt forward pass.
  • domain assumption CLAP similarity, chroma similarity, CQT-PCC, and FAD jointly and reliably measure music-edit quality.
    Metrics are treated as monotone proxies for semantic alignment, harmonic preservation, structural consistency, and perceptual realism without validation against human judgments in the objective tables.
  • domain assumption Injecting cached source KV tensors preserves musical structure without destroying editability.
    The method is built on this empirical assumption; the paper studies K vs V vs KV but does not derive why injected source features should transfer across a changed text condition.

pith-pipeline@v1.3.0-alltime-deepseek · 21619 in / 15196 out tokens · 134493 ms · 2026-08-01T17:42:56.638061+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of FlowSonic: Stable Zero-Shot Music Editing via High-Order Trajectory Integration." pith.science (2026). https://pith.science/paper/UIN7LREA

@misc{pith2026260717526,
  author       = {Pith},
  title        = {Pith review of: FlowSonic: Stable Zero-Shot Music Editing via High-Order Trajectory Integration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UIN7LREA}},
  note         = {Machine review of arXiv:2607.17526}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Zero-shot text-guided editing of real-world music recordings requires balancing semantic modification with faithful preservation of the original musical structure. Although recent diffusion transformers trained with rectified flow have achieved remarkable success in text-to-music generation, extending them to edit existing recordings remains challenging because editing requires accurate deterministic inversion, reliable structural preservation, and numerically stable integration throughout the inversion and generation processes. We present FlowSonic, a zero-shot music editing framework built upon a pretrained diffusion transformer trained with rectified flow. FlowSonic first deterministically inverts a real-world recording into the latent space and preserves its musical structure during editing by reusing cross-attention representations extracted during inversion. To improve the numerical reliability of inversion-based editing, we introduce a high-order ODE solver and systematically investigate how different numerical integration schemes influence trajectory stability, structural preservation, and semantic controllability. Comprehensive experiments on timbre-transfer and genre-modification tasks demonstrate that FlowSonic consistently outperforms existing music editing methods across semantic alignment, harmonic preservation, structural consistency, and perceptual audio quality. We further provide geometric and empirical analyses showing how the proposed numerical integration strategy improves latent trajectory stability and leads to more reliable music editing.

Figures

Figures reproduced from arXiv: 2607.17526 by Ali Boudaghi, Hadi Zare.

Figure 1
Figure 1. Figure 1: Overview of the proposed FlowSonic framework for zero-shot text-guided music editing. During deterministic inversion [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Global latent-space trajectory visualization for the timbre-transfer task projected using Principal Component Analysis [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Influence of injection steps and injection block (IB) count on the transferability–fidelity trade-off for the timbre transfer [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of Mel spectrograms generated using different numerical integration strategies within the [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of Mel spectrograms generated using different numerical integration strategies within the [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 9
Figure 9. Figure 9: Results of injecting the key (K) components of the [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Results of injecting both key and value (K + V) [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 8
Figure 8. Figure 8: Results of injecting the value (V) components of the [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 11
Figure 11. Figure 11: Global latent-space trajectory visualization for the genre-transfer task projected onto the first two principal components [PITH_FULL_IMAGE:figures/full_fig_p017_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

63 extracted references · 39 linked inside Pith

  1. [1]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” 2020. [Online]. Available: https://arxiv.org/abs/2006.11239

  2. [2]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” 2021. [Online]. Available: https://arxiv.org/abs/ 2011.13456

  3. [3]

    Elucidating the design space of diffusion-based generative models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,” 2022. [Online]. Available: https://arxiv.org/abs/2206.00364

  4. [4]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” 2023. [Online]. Available: https://arxiv.org/abs/2210.02747

  5. [5]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” 2022. [Online]. Available: https://arxiv.org/abs/2112.10752

  6. [6]

    Photorealistic text-to-image diffusion models with deep language understanding,

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. Denton, S. K. S. Ghasemipour, B. K. Ayan, S. S. Mahdavi, R. G. Lopes, T. Salimans, J. Ho, D. J. Fleet, and M. Norouzi, “Photorealistic text-to-image diffusion models with deep language understanding,” 2022. [Online]. Available: https://arxiv.org/abs/2205.11487

  7. [7]

    AudioLDM 2: Learning holis- tic audio generation with self-supervised pretraining,

    H. Liu, Q. Tian, Y . Yuan, X. Liu, X. Mei, Q. Kong, Y . Wang, W. Wang, Y . Wang, and M. D. Plumbley, “AudioLDM 2: Learning holis- tic audio generation with self-supervised pretraining,”arXiv preprint arXiv:2308.05734, 2023

  8. [8]

    Riffusion: Stable diffusion for real-time music generation,

    S. Forsgren and H. Martiros, “Riffusion: Stable diffusion for real-time music generation,” 2022. [Online]. Available: https://riffusion.com

  9. [9]

    Noise2music: Text-conditioned music generation with diffusion models,

    Q. Huang, D. S. Park, T. Wang, T. I. Denk, A. Ly, N. Chen, Z. Zhang, Z. Zhang, J. Yu, C. Frank, J. Engel, Q. V . Le, W. Chan, Z. Chen, and W. Han, “Noise2music: Text-conditioned music generation with diffusion models,”arXiv preprint arXiv:2302.03917, 2023

  10. [10]

    Auffusion: Leveraging the power of diffusion and large language models for text-to-audio generation,

    J. Xue, Y . Deng, Y . Gao, and Y . Li, “Auffusion: Leveraging the power of diffusion and large language models for text-to-audio generation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 4700–4712, 2024

  11. [11]

    MusicLDM: Enhancing novelty in text-to-music generation using beat-synchronous mixup strategies,

    K. Chen, Y . Wu, H. Liu, M. Nezhurina, T. Berg-Kirkpatrick, and S. Dub- nov, “MusicLDM: Enhancing novelty in text-to-music generation using beat-synchronous mixup strategies,”arXiv preprint arXiv:2308.01546, 2023

  12. [12]

    Flux that plays music,

    Z. Fei, M. Fan, C. Yu, and J. Huang, “Flux that plays music,” 2024. [Online]. Available: https://arxiv.org/abs/2409.00587

  13. [13]

    AudioLDM: Text-to-audio generation with latent dif- fusion 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 dif- fusion models,” inProceedings of International Conference on Machine Learning (ICML), 2023

  14. [14]

    Accomontage: Accompaniment arrangement via phrase selection and style transfer,

    J. Zhao and G. Xia, “Accomontage: Accompaniment arrangement via phrase selection and style transfer,” inProceedings of the 22nd International Society for Music Information Retrieval Conference, ISMIR 2021, Online, November 7-12, 2021, J. H. Lee, A. Lerch, Z. Duan, J. Nam, P. Rao, P. van Kranenburg, and A. Srinivasamurthy, Eds., 2021, pp. 833–840. [Online]...

  15. [16]

    MuseCoco: Generating symbolic music from text,

    P. Lu, X. Xu, C. Kang, B. Yu, C. Xing, X. Tan, and J. Bian, “MuseCoco: Generating symbolic music from text,”arXiv preprint arXiv:2306.00110, 2023

  16. [17]

    Simple and controllable music generation,

    J. Copet, F. Kreuk, I. Gat, T. Remez, D. Kant, G. Synnaeve, Y . Adi, and A. D ´efossez, “Simple and controllable music generation,”Advances in Neural Information Processing Systems (NeurIPS), 2024

  17. [18]

    Mustango: Toward controllable text-to-music generation,

    J. Melechovsky, Z. Guo, D. Ghosal, N. Majumder, D. Herremans, and S. Poria, “Mustango: Toward controllable text-to-music generation,” inProceedings of Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2024

  18. [19]

    Diffusion based text-to-music generation with global and local text based conditioning,

    J. Zhang, P. P. Parada, M. Asif Jalal, and K. Saravanan, “Diffusion based text-to-music generation with global and local text based conditioning,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

  19. [20]

    M 2UGen: Multi-modal music understanding and generation with the power of large language models,

    A. S. Hussain, S. Liu, C. Sun, and Y . Shan, “M 2UGen: Multi-modal music understanding and generation with the power of large language models,”arXiv preprint arXiv:2311.11255, 2023

  20. [21]

    InstructME: An instruction guided music edit and remix frame- work with latent diffusion models,

    B. Han, J. Dai, X. Song, W. Hao, X. He, D. Guo, J. Chen, Y . Wang, and Y . Qian, “InstructME: An instruction guided music edit and remix frame- work with latent diffusion models,”arXiv preprint arXiv:2308.14360, 2023

  21. [22]

    AUDIT: Audio editing by following instructions with latent diffusion models,

    Y . Wang, Z. Ju, X. Tan, L. He, Z. Wu, J. Bian, and S. Zhao, “AUDIT: Audio editing by following instructions with latent diffusion models,” arXiv preprint arXiv:2304.00830, 2023

  22. [23]

    Musicmagus: Zero-shot text-to-music editing via diffusion models,

    Y . Zhang, Y . Ikemiya, G. Xia, N. Murata, M. A. Mart ´ınez-Ram´ırez, W.-H. Liao, Y . Mitsufuji, and S. Dixon, “Musicmagus: Zero-shot text-to-music editing via diffusion models,” 2024. [Online]. Available: https://arxiv.org/abs/2402.06178

  23. [24]

    Zero-shot unsupervised and text-based audio editing using ddpm inversion,

    H. Manor and T. Michaeli, “Zero-shot unsupervised and text-based audio editing using ddpm inversion,” 2024. [Online]. Available: https://arxiv.org/abs/2402.10009

  24. [25]

    Transplayer: Timbre style transfer with flexible timbre control,

    Y . Wu, Y . He, X. Liu, Y . Wang, and R. B. Dannenberg, “Transplayer: Timbre style transfer with flexible timbre control,” in2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023

  25. [26]

    A diffusion- based generative equalizer for music restoration,

    E. Moliner, M. Turunen, F. Elvander, and V . V ¨alim¨aki, “A diffusion- based generative equalizer for music restoration,” inProceedings of the 27th International Conference on Digital Audio Effects (DAFx24), 2024, p. –, arXiv preprint arXiv:2403.18636, 2024. [Online]. Available: https://arxiv.org/abs/2403.18636

  26. [27]

    Improving the training of rectified flows,

    S. Lee, Z. Lin, and G. Fanti, “Improving the training of rectified flows,” 2024. [Online]. Available: https://arxiv.org/abs/2405.20320

  27. [28]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” 2022. [Online]. Available: https://arxiv.org/abs/2209.03003

  28. [29]

    Hart: Efficient visual generation with hybrid autoregressive transformer,

    H. Tang, Y . Wu, S. Yang, E. Xie, J. Chen, J. Chen, Z. Zhang, H. Cai, Y . Lu, and S. Han, “Hart: Efficient visual generation with hybrid autoregressive transformer,”arXiv preprint arXiv:2410.10812, 2024

  29. [30]

    Sana: Efficient high-resolution image synthesis with linear diffusion transformers,

    E. Xie, J. Chen, J. Chen, H. Cai, Y . Lin, Z. Zhang, M. Li, Y . Lu, and S. Han, “Sana: Efficient high-resolution image synthesis with linear diffusion transformers,”arXiv preprint arXiv:2410.10629, 2024

  30. [31]

    Cogvideox: Text-to-video diffusion models with an expert transformer,

    Z. Yang, J. Teng, W. Zheng, M. Ding, S. Huang, J. Xu, Y . Yang, W. Hong, X. Zhang, G. Fenget al., “Cogvideox: Text-to-video diffusion models with an expert transformer,”arXiv preprint arXiv:2408.06072, 2024

  31. [32]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” inInternational Conference on Computer Vision, 2023, pp. 4195–4205

  32. [33]

    MusicLM: Generating music from text,

    A. Agostinelli, T. I. Denk, Z. Borsos, J. Engel, M. Verzetti, A. Caillon, Q. Huang, A. Jansen, A. Roberts, M. Tagliasacchiet al., “MusicLM: Generating music from text,”arXiv preprint arXiv:2301.11325, 2023

  33. [34]

    Musicgen-stem: Multi- stem music generation and edition through autoregressive modeling,

    S. Rouard, R. S. Roman, Y . Adi, and A. Roebel, “Musicgen-stem: Multi- stem music generation and edition through autoregressive modeling,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

  34. [35]

    Inspiremusic: Integrating super resolution and large language model for high-fidelity long-form music generation,

    C. Zhang, Y . Ma, Q. Chen, W. Wang, S. Zhao, Z. Pan, H. Wang, C. Ni, T. H. Nguyen, K. Zhou, Y . Jiang, C. Tan, Z. Gao, Z. Du, and B. Ma, “Inspiremusic: Integrating super resolution and large language model for high-fidelity long-form music generation,” 2025. [Online]. Available: https://arxiv.org/abs/2503.00084

  35. [36]

    Diffrhythm: Blazingly fast and embarrassingly simple end-to-end full-length song generation with latent diffusion,

    Z. Ning, H. Chen, Y . Jiang, C. Hao, G. Ma, S. Wang, J. Yao, and L. Xie, “Diffrhythm: Blazingly fast and embarrassingly simple end-to-end full-length song generation with latent diffusion,” 2025. [Online]. Available: https://arxiv.org/abs/2503.01183

  36. [37]

    M ¨ousai: Text-to-music generation with long-context latent diffusion,

    F. Schneider, Z. Jin, and B. Sch ¨olkopf, “M ¨ousai: Text-to-music generation with long-context latent diffusion,”arXiv preprint arXiv:2301.11757, 2023

  37. [38]

    Text-to-audio generation using instruction-tuned LLM and latent diffusion model,

    D. Ghosal, N. Majumder, A. Mehrish, and S. Poria, “Text-to-audio generation using instruction-tuned LLM and latent diffusion model,” arXiv preprint arXiv:2304.13731, 2023

  38. [39]

    Efficient parallel audio generation using group masked language modeling,

    M. Jeong, M. Kim, J. Y . Lee, and N. S. Kim, “Efficient parallel audio generation using group masked language modeling,”arXiv preprint arXiv:2401.01099, 2024

  39. [40]

    Ditto: Diffusion inference-time t-optimization for music generation,

    Z. Novack, J. McAuley, T. Berg-Kirkpatrick, and N. J. Bryan, “Ditto: Diffusion inference-time t-optimization for music generation,” 2024

  40. [41]

    Music Con- trolNet: Multiple time-varying controls for music generation,

    S.-L. Wu, C. Donahue, S. Watanabe, and N. J. Bryan, “Music Con- trolNet: Multiple time-varying controls for music generation,”arXiv preprint arXiv:2311.07069, 2023

  41. [42]

    Content-based controls for music large language modeling,

    L. Lin, G. Xia, J. Jiang, and Y . Zhang, “Content-based controls for music large language modeling,”arXiv preprint arXiv:2310.17162, 2023

  42. [43]

    Musflow: Multimodal music generation via conditional flow matching,

    J. Song and Y . Wang, “Musflow: Multimodal music generation via conditional flow matching,” 2025. [Online]. Available: https: //arxiv.org/abs/2504.13535

  43. [44]

    High fidelity text-guided music editing via single-stage flow matching,

    G. L. Lan, B. Shi, Z. Ni, S. Srinivasan, A. Kumar, B. Ellis, D. Kant, V . Nagaraja, E. Chang, W.-N. Hsu, Y . Shi, and V . Chandra, “High fidelity text-guided music editing via single-stage flow matching,” 2024. [Online]. Available: https://arxiv.org/abs/2407.03648 15

  44. [45]

    Audio prompt adapter: Unleashing music editing abilities for text-to-music with lightweight finetuning,

    F.-D. Tsai, S.-L. Wu, H. Kim, B.-Y . Chen, H.-C. Cheng, and Y .-H. Yang, “Audio prompt adapter: Unleashing music editing abilities for text-to-music with lightweight finetuning,” 2024. [Online]. Available: https://arxiv.org/abs/2407.16564

  45. [46]

    Score-based generative modeling through stochastic differential equations,

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

  46. [47]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inInternational Conference on Learning Representations, 2021. [Online]. Available: https://openreview.net/forum?id=St1giarCHLP

  47. [48]

    Null- text inversion for editing real images using guided diffusion models,

    R. Mokady, A. Hertz, K. Aberman, Y . Pritch, and D. Cohen-Or, “Null- text inversion for editing real images using guided diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 6038–6047

  48. [49]

    Beyond first-order tweedie: Solving inverse problems using latent diffusion,

    L. Rout, Y . Chen, A. Kumar, C. Caramanis, S. Shakkottai, and W.- S. Chu, “Beyond first-order tweedie: Solving inverse problems using latent diffusion,” in2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

  49. [50]

    Direct inversion: Optimization-free text-driven real image editing with diffusion models,

    A. Elarabawy, H. Kamath, and S. Denton, “Direct inversion: Optimization-free text-driven real image editing with diffusion models,” arXiv preprint arXiv:2211.07825, 2022

  50. [51]

    Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models,

    D. Miyake, A. Iohara, Y . Saito, and T. Tanaka, “Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models,”arXiv preprint arXiv:2305.16807, 2023

  51. [52]

    Text-to-image rectified flow as plug-and-play priors,

    X. Yang, C. Chen, X. Yang, F. Liu, and G. Lin, “Text-to-image rectified flow as plug-and-play priors,”arXiv preprint arXiv:2406.03293, 2024

  52. [53]

    Semantic image inversion and editing using rectified stochastic differ- ential equations,

    L. Rout, Y . Chen, N. Ruiz, C. Caramanis, S. Shakkottai, and W. Chu, “Semantic image inversion and editing using rectified stochastic differ- ential equations,” 2024

  53. [54]

    Scaling instruction-finetuned language models,

    H. W. Chung, L. Hou, S. Longpre, B. Zoph, Y . Tay, W. Fedus, Y . Li, X. Wang, M. Dehghani, S. Brahmaet al., “Scaling instruction-finetuned language models,”Journal of Machine Learning Research (JMLR), 2024

  54. [55]

    Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,

    Y . Wu, K. Chen, T. Zhang, Y . Hui, T. Berg-Kirkpatrick, and S. Dubnov, “Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,” in2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023

  55. [56]

    Prompt-to-prompt image editing with cross attention control,

    A. Hertz, R. Mokady, J. Tenenbaum, K. Aberman, Y . Pritch, and D. Cohen-Or, “Prompt-to-prompt image editing with cross attention control,” 2022. [Online]. Available: https://arxiv.org/abs/2208.01626

  56. [57]

    Plug-and-play diffusion features for text-driven image-to-image translation,

    N. Tumanyan, M. Geyer, S. Bagon, and T. Dekel, “Plug-and-play diffusion features for text-driven image-to-image translation,” 2022. [Online]. Available: https://arxiv.org/abs/2211.12572

  57. [58]

    A manifold representation of the key in vision transformers,

    L. Meng, M. Goodwin, A. Yazidi, and P. Engelstad, “A manifold representation of the key in vision transformers,” 2024. [Online]. Available: https://arxiv.org/abs/2402.00534

  58. [59]

    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,” inInternational Conference on Learning Representations (ICLR), 2022

  59. [60]

    Steermusic: Enhanced musical consistency for zero-shot text- guided and personalized music editing,

    X. Niu, K. W. Cheuk, J. Zhang, N. Murata, C.-H. Lai, M. Mancusi, W. Choi, G. Fabbro, W.-H. Liao, C. P. Martin, and Y . Mitsufuji, “Steermusic: Enhanced musical consistency for zero-shot text- guided and personalized music editing,” 2025. [Online]. Available: https://arxiv.org/abs/2504.10826

  60. [61]

    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 SciPy, 2015

  61. [62]

    Calculation of a constant q spectral transform,

    J. C. Brown, “Calculation of a constant q spectral transform,”The Journal of the Acoustical Society of America, 1991

  62. [63]

    Subjective video quality assessment methods for multimedia applications,

    ITU-T Recommendation, “Subjective video quality assessment methods for multimedia applications,”International Telecommunication Union, 1999

  63. [64]

    Methods for subjective determination of transmission quality,

    ——, “Methods for subjective determination of transmission quality,” International Telecommunication Union, 1996. 16 APPENDIX A. Additional Results on Attention Injection Variants Fig. 6: Results of injecting the key (K) components of the attention mechanism during timbre transfer task. InjectingK leads to moderate improvements in transferability but sligh...