Pith. sign in

REVIEW 3 major objections 4 minor 38 references

Beyond Reconstruction: Full-Context Generative DiT for Music Generation

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read This paper claims that rendering music from a codec plan should be treated as full-context generation from an imperfect plan, and that error-matched training corruptions make a diffusion renderer recover clean audio from flawed tokens.

desk verdict A solid, well-ablated system paper for hybrid music generation; the EMDC mechanism is genuinely new and the leaderboard result is real, but the headline robustness gain is partly self-confirming because the synthetic corruption comes from the same tables used to train the model. read the letter →

arxiv 2608.08787 v2 pith:ZY66KRRI submitted 2026-08-09 eess.AS

Yunjia Li , Menglin Wu , Junyu Dai , Xinyue Fan , Xiangang Li , Haoxu Wang , Jianwei Yu , Huaicheng Zhang
show 8 more authors
Han Zhao Weiqin Li Yufei Shi Cheng Wen Sitong Zhao Qixi Zheng Haina Zhu Wei Li
This is my paper · ORCID
classification eess.AS
keywords musicgenerationdiffusiontransformerresidualvectorquantizationcodec-interfaceexposurebiaserror-matcheddistractorconditioningclassifier-freeguidanceflowmatchingaudiorendering
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

Hybrid music generators split the task: a language model plans tokens, and a renderer turns them into audio. The renderer is trained on clean tokens extracted from real audio but deployed on imperfect predictions, so the interface is systematically cleaner in training than in use. This paper argues that the fix is to treat rendering as full-context generation from a possibly wrong plan, and introduces FullDiT plus a training corruption scheme called Error-Matched Distractor Conditioning (EMDC). EMDC corrupts each of the eight token streams at the rate the upstream model actually makes mistakes, replacing tokens with near-miss alternatives from embedding neighborhoods while keeping the acoustic target unchanged. Under matched ablations the scheme raises ViSQOL from 2.4342 to 3.2036 on synthetic corruption and wins 69.7% of non-tied blind comparisons on language-model-generated plans.

What carries the argument

FullDiT is a non-causal flow-matching Transformer that fuses eight frame-aligned RVQ token streams with separately encoded captions and lyrics, and attends over the entire acoustic latent sequence. The load-bearing training device is EMDC, which for each frame and codebook samples a Bernoulli mask with probability $1 - \mathrm{Acc@1}$ (the teacher-forced top-1 error rate) and, when a replacement is drawn, samples a near-miss token from the cosine top-$K$ neighborhood of the clean token. Because the acoustic target latent is never changed, EMDC teaches the renderer to repair bad plans rather than to model corrupted audio.

What would settle it

Collect the real token predictions of the upstream language model on a held-out set, align them frame-by-frame with teacher-forced ground-truth tokens, and compare the per-codebook replacement rates and replacement-token choices with EMDC's calibration tables. If the real error rates diverge substantially from $1 - \mathrm{Acc@1}$, or if a renderer trained with real-error distractors fails to beat the cosine-KNN version, the mechanism is not what the paper claims.

Watch

Extended reading notes

Core claim

The central claim is that codec-interface exposure bias is why hybrid renderers sound worse than they should, and that it can be removed without changing architecture or data budget. The discovery is that a renderer conditioned on a corrupted eight-stream residual-vector-quantization (RVQ) plan whose per-codebook replacement rates equal teacher-forced top-1 error rates, and whose replacements are drawn from cosine-KNN neighborhoods of the true token, learns to complete the plan from text and full-song context rather than reconstruct it blindly. This makes the renderer robust at the deployed interface while preserving clean-condition quality.

Load-bearing premise

The central bet is that teacher-forced top-1 error rates measured on clean plans approximate how often and where the deployed autoregressive language model will actually be wrong, and that the nearest-neighbor tokens it is likely to confuse live near the true token in the codec embedding space.

Editorial extensions

If this is right

  • Renderers in hybrid pipelines can be made robust to upstream prediction errors by matching corruption to the planner's measured per-codebook error profile, without retraining the language model.
  • Because deeper residual codebooks are replaced at higher rates, the renderer learns to rely more on text and context at high codebook levels, shifting where information is recovered.
  • Full-song non-causal context is itself a major factor: a local 30-second window drops production quality from 8.213 to 6.011, so long-range relationships in the plan matter for acoustic realization.
  • Four-way classifier-free guidance allows separate control of how much the renderer trusts the codec plan versus captions versus lyrics, and blind listening preferred the tuple $(1,2,1)$.

Reading between the lines

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

  • If the error-rate calibration transfers across languages and genres, EMDC could be applied as a plug-in training recipe to other codec-conditioned diffusion and flow renderers, not just music.
  • A natural testable extension is to replace the cosine-KNN near-miss sampler with tokens the upstream language model actually confuses, sampled from its softmax; if those outperform geometric near-misses, the geometry is a proxy rather than the mechanism itself.
  • The same interface-bias argument should apply to other generative pipelines where an autoregressive model produces discrete tokens for a diffusion renderer, such as speech or audiovisual synthesis.
  • Because the synthetic-corruption test is built from the same tables used in training, the externally generated language-model plan test is the more decisive evidence for EMDC's real-world benefit.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. Hybrid music generators combine an autoregressive language model that plans codec tokens with a diffusion/flow renderer, and the paper identifies codec-interface exposure bias: renderers trained on clean, target-derived codec tokens are deployed on imperfect LM predictions. It introduces FullDiT, an 8B non-causal DiT flow renderer conditioned on eight frame-aligned RVQ streams, caption and lyric embeddings, with four-way classifier-free guidance. During training, Error-Matched Distractor Conditioning (EMDC) replaces tokens per codebook at rates equal to teacher-forced top-1 error rates, sampling near-miss tokens from cosine-KNN neighborhoods while leaving the acoustic target unchanged. The paper reports matched 1.5B ablations showing that full-song context, renderer-side text conditioning, and EMDC each help, with EMDC improving ViSQOL from 2.4342 to 3.2036 under synthetic corruption and PQ from 8.122 to 8.213 with fixed LM-generated tokens, plus 69.7% non-tied blind preference. The complete system is reported to outperform five commercial systems on 15 of 18 automatic metrics and to rank in the top three on the Artificial Analysis Music with Vocals Leaderboard.

Significance. The paper addresses a real and practically important interface mismatch in hybrid music generation, and the controlled ablation design is a genuine strength: M1-M3 hold architecture, training budget, full-song context, and text conditioning fixed while isolating EMDC, and the LM-generated codec cache is fixed across models. The EMDC idea is simple and potentially general, and the external leaderboard result provides some independent support for the complete system. The main limitation is that the synthetic-corruption evaluation is generated from the same teacher-forced error rates and M1's KNN tables used to train EMDC, so the headline ViSQOL gain is partially self-confirming; the LM-generated condition, which is the non-circular evidence, is statistically thin (100 songs, five judges, no significance testing). If the requested validation against real LM error statistics is supplied, the central claim would be substantially stronger.

major comments (3)
  1. [§3.3, Eq. (6)–(7); §4.3] The EMDC calibration is never validated against the actual error distribution of the deployed LM. The replacement rates p_k = 1 - Acc@1_k are teacher-forced top-1 error rates measured with ground-truth context, and the replacement tokens are drawn from cosine-KNN neighborhoods of the clean token; at inference the LM's own tokens form the context, so error propagation, sampling temperature, and decoding choices can change both the per-codebook error rates and the confusion sets. More importantly, the synthetic-corruption evaluation in §4.3 uses the same teacher-forced error rates and M1's step-150k KNN tables that generated the training corruption, so the 0.77 ViSQOL gain (2.4342 to 3.2036) largely measures robustness to the exact corruption distribution used in training. Please validate p_k and the KNN neighborhoods against the free-running LM's outputs on a paired held-out set, and report the synthetic-corruption results under a distribution built from those measured statistics.
  2. [Table 2, §5.1] The LM-generated codec condition is the only external grounding for EMDC, but it is reported without any measure of uncertainty. The PQ improvement from 8.122 to 8.213 is computed on 200 songs, and the blind preference of 69.7% is computed on 100 songs with five judges; no confidence intervals, significance tests, or inter-annotator agreement are given. Please report these, or at least a binomial confidence interval for the 69.7% preference and the number of non-tied songs, so the reader can judge whether the EMDC effect under the deployed interface is reliable.
  3. [Table 3, §5.3] The complete-system advantage over commercial baselines is based on point estimates without uncertainty, and several winning margins are extremely small (e.g., Audiobox Production Complexity 6.8774 vs. Mureka 6.8768; CMI-RM Musicality 2.7611 vs. 2.7590). Given that the metrics are learned or heuristic evaluators, please report significance or effect sizes, or at least the variance across the 500 evaluation songs, before claiming to "outperform" five commercial systems on 15 of 18 metrics.
minor comments (4)
  1. [§3.3] The phrase "linear ramp of each replacement rate from zero topk over 30k–40k" appears to contain a typo; it should presumably read "from zero to p_k."
  2. [Table 2, §5.1] For the M2a vs. M1 and M2b vs. M1 comparisons, reporting "0.0%" non-tied preference without the number of non-tied songs is uninformative; please provide the denominator.
  3. [Abstract, §5.3] The claim of "15 of 18 automatic metrics" is consistent with the best point estimates in Table 3, but the abstract could state explicitly that this is a point-estimate comparison.
  4. [Figure 2] The caption contains a typo: "submitt" should be "submit."

Circularity Check

1 steps flagged · score 2.0 of 10

EMDC's synthetic-corruption test uses the same teacher-forced error rates and M1's own cosine-KNN tables that define its training corruption, making the headline 0.77 ViSQOL gain partly self-confirming; the LM-generated condition and external leaderboard provide independent support.

  1. fitted input called prediction [Section 3.3 Eq. (6)-(7); Section 4.3 (Evaluation Protocols); Table 2]
    "Mi,k ∼ Bernoulli(p k), p k = 1−Acc@1 k ... Every controlled model receives exactly the same synthetic-corrupted token sequences. Their replacement rates come from teacher-forced top-1 errors, while replacement identities come from M1’s step-150k cosine-KNN tables. ... Under synthetic corruption, M1 raises ViSQOL from 2.4342 to 3.2036, reduces log-mel L1 from 2.2372 to 1.1842, and reduces MR-STFT from 2.3346 to 1.7943."

    By construction, the synthetic-corruption test distribution is the same distribution used to train EMDC: replacement rates are the teacher-forced top-1 error rates p_k of Eq. (6), and replacement tokens are sampled from M1's own step-150k cosine-KNN tables, the same tables used during training. Thus the headline ViSQOL gain (2.4342 to 3.2036) is a matched-distribution check: it shows FullDiT with EMDC tolerates the exact corruption model it was trained on. It does not independently validate that p_k or the KNN choices match the upstream LM's free-running error distribution. The LM-generated codec condition (PQ 8.122 to 8.213, 69.7% non-tied preference) is a separate and more external result, which is why this is only partial circularity and does not collapse the paper's central claim.

full rationale

Most of the paper is an empirical systems study, not a derivation chain. The matched ablations (M1 vs M2a, M2b, M3) and the external leaderboard evaluation are independent of each other. The only load-bearing step that reduces by construction is the synthetic-corruption evaluation of EMDC: Section 3.3 defines replacement rates p_k = 1 - Acc@1_k and samples replacement tokens from cosine-KNN neighborhoods; Section 4.3 states that the synthetic-corrupted test sequences use the same teacher-forced top-1 error rates and M1's step-150k cosine-KNN tables. Consequently, the 0.77 ViSQOL gain under synthetic corruption is a robustness check against the model's own training-time corruption distribution, not an independent calibration of the error model against the deployed LM's actual free-running errors. This is a real but partial self-confirmation. It does not rise to derivational circularity because (i) the model must still learn to recover the clean target from corrupted inputs, and (ii) the LM-generated codec condition provides an external test with PQ rising from 8.122 to 8.213 and a 69.7% non-tied preference for M1 over M3, which does not use the synthetic tables. The mismatch between teacher-forced and free-running LM errors is an unvalidated assumption, not a circularity. Self-citations (e.g., LeVo, SongBench, HeartMuLa) are used as baselines or components rather than as load-bearing justification. Overall, the central claims retain independent empirical content, so the circularity score is low.

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

The central claim rests on EMDC calibration values p_k and K_k that are measured or hand-chosen on unreleased internal data and an unreleased upstream LM, so the key numbers cannot be independently reproduced. The method also assumes that teacher-forced errors and embedding-space neighborhoods characterize real LM confusions. Standard flow-matching background is the only standard-math input. No invented entities are postulated.

free parameters (5)
  • Per-codebook replacement rate p_k = 0.375, 0.379, 0.340, 0.302, 0.230, 0.189, 0.141, 0.108 for codebooks 0-7
    Set equal to 1 minus teacher-forced top-1 accuracy (Eq. 6, Table 1). Measured on internal data with the upstream LM, so values are dataset- and model-specific and cannot be transferred or verified without release.
  • KNN neighborhood widths K_k = 20, 20, 20, 30, 50, 80, 125, 180
    Chosen by the heuristic K_k = round(2 exp(loss_k)); described as an engineering bandwidth heuristic, not a derived quantity (Section 3.3, Table 1).
  • KNN sampling temperature tau = 1.0
    Fixed at 1.0 with no sensitivity study; controls how sharply near-miss tokens resemble the query token (Eq. 7).
  • 4-CFG guidance tuple (scodec, slyr, scap) = (1, 2, 1)
    Selected by automatic PQ screening on N=200 followed by a blind three-way listening test (Section 5.2); used in all controlled comparisons, so the operating point is tuned on the same evaluation family.
  • EMDC curriculum boundaries and embedding freeze = clean 0-30k; ramp 30k-40k; target rates 40k-150k; embeddings frozen first 5k
    Hand-set schedule; no ablation or sensitivity analysis presented (Section 3.3).
assumptions (5)
  • domain assumption Teacher-forced top-1 error rates approximate the true error distribution of the autoregressive LM's codec predictions at inference.
    EMDC replacement rates are set to p_k = 1 - Acc@1_k (Eq. 6). Teacher-forcing is optimistic relative to autoregressive sampling, and no calibration against actual LM-generated error statistics is provided.
  • domain assumption Cosine-KNN neighborhoods in the codec embedding space contain near-miss tokens that mimic plausible LM errors.
    Replacement identities are drawn only from top-K cosine neighbors (Eq. 7). No evidence is given that the LM's errors are localized in embedding space.
  • domain assumption Automatic quality metrics (SongBench, SongEval, Audiobox-Aesthetics, CMI-RM) are valid proxies for human preference in the complete-system comparison.
    Table 3's 15/18 wins use these learned metrics on an internal 500-song set; the external leaderboard is human, but the common-set comparison is not.
  • domain assumption The internal held-out sets (N=200, N=100, N=500) are representative of general vocal music across Chinese, English, Japanese, Korean, and Spanish.
    All controlled and common-set evaluations use internal data curated by the authors; no external corpus is used for component claims.
  • standard math Standard flow-matching and UniPC ODE-solver assumptions hold for the conditional full-song flow model.
    The method relies on Lipman et al. flow matching and Zhao et al. UniPC without modification; these are standard prior results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Reconstruction: Full-Context Generative DiT for Music Generation." pith.science (2026). https://pith.science/paper/ZY66KRRI

@misc{pith2026260808787,
  author       = {Pith},
  title        = {Pith review of: Beyond Reconstruction: Full-Context Generative DiT for Music Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZY66KRRI}},
  note         = {Machine review of arXiv:2608.08787}
}
read the original abstract

Hybrid music generators combine the long-range planning of an autoregressive language model with the fidelity of a diffusion- or flow-based acoustic renderer. Yet renderers are trained with clean, target-derived codec tokens but deployed with imperfect language-model predictions, creating codecinterface exposure bias. Rather than treating rendering as a simple reconstruction task,we formulate it as full-context generation from an imperfect discrete plan. We introduce FullDiT, a conditional DiT that fuses eight frame-aligned RVQ streams with independently encoded captions and lyrics and uses non-causal self-attention over the complete acoustic latent sequence. During training, Error-Matched Distractor Conditioning (EMDC) matches per-codebook replacement rates to teacher-forced top-1 error rates and samples near-miss tokens from cosine-KNN neighborhoods without changing the acoustic target. At inference, four-way classifier-free guidance (4-CFG) independently scales codec, lyric, and caption guidance increments. Matched ablations show that EMDC improves ViSQOL by 0.77 under synthetic corruption and is clearly preferred in non-tied comparisons with fixed languagemodel tokens. Further ablations show gains from full-song context and renderer-side text conditioning. The complete system outperforms five commercial systems on 15 of 18 automatic metrics and ranks among the top three on the Artificial Analysis Music with Vocals Leaderboard. The demo page is available at https://selinacloudl.github.io/fulldit-demo/.

Figures

Figures reproduced from arXiv: 2608.08787 by the authors.

Figure 1
Figure 1. FullDiT acoustic renderer. Independently encoded lyrics and captions provide textual conditions, while frame-aligned [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Four-way CFG analysis: (a) Audiobox PQ screening of seven guidance tuples; (b) song-weighted selection shares from [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Archived Artificial Analysis Music with Vocals Leaderboard snapshot used in this paper. The highlighted Lucky [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 23 canonical work pages

  1. [1]

    I.; Borsos, Z.; Engel, J.; Verzetti, M.; Caillon, A.; Huang, Q.; Jansen, A.; Roberts, A.; Tagliasacchi, M.; Sharifi, M.; Zeghidour, N.; and Frank, C

    Agostinelli, A.; Denk, T. I.; Borsos, Z.; Engel, J.; Verzetti, M.; Caillon, A.; Huang, Q.; Jansen, A.; Roberts, A.; Tagliasacchi, M.; Sharifi, M.; Zeghidour, N.; and Frank, C. 2023. MusicLM : Generating Music From Text. arXiv preprint arXiv:2301.11325

    arXiv 2023

  2. [2]

    Artificial Analysis . 2026. Music with Vocals Leaderboard. https://artificialanalysis.ai/music/leaderboard/vocals. Accessed 2026-07-20; the manuscript reports a dated archived snapshot

    work page 2026

  3. [3]

    Bai, Y.; Chen, H.; Chen, J.; Chen, Z.; Deng, Y.; Dong, X.; Hantrakul, L.; Hao, W.; Huang, Q.; Huang, Z.; Jia, D.; La, F.; Le, D.; Li, B.; Li, C.; Li, H.; Li, X.; Liu, S.; Lu, W.-T.; Lu, Y.; Shaw, A.; Spijkervet, J.; Sun, Y.; Wang, B.; Wang, J.-C.; Wang, Y.; Wang, Y.; Xu, L.; Yang, Y.; Yao, C.; Zhang, S.; Zhang, Y.; Zhang, Y.; Zhao, H.; Zhao, Z.; Zhong, D....

    arXiv 2024

  4. [4]

    Bengio, S.; Vinyals, O.; Jaitly, N.; and Shazeer, N. 2015. Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks. In Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc

    work page 2015

  5. [5]

    Chinen, M.; Lim, F. S. C.; Skoglund, J.; Gureev, N.; O'Gorman, F.; and Hines, A. 2020. ViSQOL v3 : An Open Source Production Ready Objective Speech and Audio Metric. In 2020 Twelfth International Conference on Quality of Multimedia Experience, 1--6

    work page 2020

  6. [6]

    Chiu, C.-C.; Qin, J.; Zhang, Y.; Yu, J.; and Wu, Y. 2022. Self-Supervised Learning with Random-Projection Quantizer for Speech Recognition. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, 3915--3924. PMLR

    work page 2022

  7. [7]

    Copet, J.; Kreuk, F.; Gat, I.; Remez, T.; Kant, D.; Synnaeve, G.; Adi, Y.; and D \'e fossez, A. 2023. Simple and Controllable Music Generation. In Advances in Neural Information Processing Systems, volume 36, 47704--47720. Curran Associates, Inc

    work page 2023

  8. [8]

    D \'e fossez, A.; Copet, J.; Synnaeve, G.; and Adi, Y. 2023. High Fidelity Neural Audio Compression. Transactions on Machine Learning Research

    work page 2023

Show all 38 references
  1. [9]

    W.; Radford, A.; and Sutskever, I

    Dhariwal, P.; Jun, H.; Payne, C.; Kim, J. W.; Radford, A.; and Sutskever, I. 2020. Jukebox : A Generative Model for Music. arXiv preprint arXiv:2005.00341

    2020 arXiv

  2. [10]

    H.; and Pons, J

    Evans, Z.; Carr, C.; Taylor, J.; Hawley, S. H.; and Pons, J. 2024. Fast Timing-Conditioned Latent Audio Diffusion. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, 12652--12665. PMLR

    2024

  3. [11]

    Gong, J.; Zhao, S.; Wang, S.; Xu, S.; and Guo, J. 2025. ACE-Step : A Step Towards Music Generation Foundation Model. arXiv preprint arXiv:2506.00045

    2025 arXiv

  4. [12]

    Google . 2026. Lyria 3 Expands to More Google Products and Adds More Features. https://blog.google/innovation-and-ai/technology/ai/lyria-3-pro/. Product release, accessed 2026-07-20

    2026

  5. [13]

    Gulati, A.; Qin, J.; Chiu, C.-C.; Parmar, N.; Zhang, Y.; Yu, J.; Han, W.; Wang, S.; Zhang, Z.; Wu, Y.; and Pang, R. 2020. Conformer: Convolution-Augmented Transformer for Speech Recognition. In Interspeech 2020, 5036--5040

    2020

  6. [14]

    J.; Norouzi, M.; and Salimans, T

    Ho, J.; Saharia, C.; Chan, W.; Fleet, D. J.; Norouzi, M.; and Salimans, T. 2022. Cascaded Diffusion Models for High Fidelity Image Generation. Journal of Machine Learning Research, 23(47): 1--33

    2022

  7. [15]

    Ho, J.; and Salimans, T. 2022. Classifier-Free Diffusion Guidance. arXiv preprint arXiv:2207.12598

    2022 arXiv

  8. [16]

    S.; Wang, T.; Denk, T

    Huang, Q.; Park, D. S.; Wang, T.; Denk, T. I.; Ly, A.; Chen, N.; Zhang, Z.; Zhang, Z.; Yu, J.; Frank, C.; Engel, J.; Le, Q. V.; Chan, W.; Chen, Z.; and Han, W. 2023. Noise2Music : Text-Conditioned Music Generation with Diffusion Models. arXiv preprint arXiv:2302.03917

    2023 arXiv

  9. [17]

    Lam, M. W. Y.; Tian, Q.; Li, T.; Yin, Z.; Feng, S.; Tu, M.; Ji, Y.; Xia, R.; Ma, M.; Song, X.; Chen, J.; Wang, Y.; and Wang, Y. 2023. Efficient Neural Music Generation. In Advances in Neural Information Processing Systems, volume 36, 17450--17463

    2023

  10. [18]

    Lei, S.; Xu, Y.; Lin, Z.; Zhang, H.; Tan, W.; Chen, H.; Zhang, Y.; Yang, C.; Zhu, H.; Wang, S.; Wu, Z.; and Yu, D. 2025. LeVo : High-Quality Song Generation with Multi-Preference Alignment. In Advances in Neural Information Processing Systems, volume 38

    2025

  11. [19]

    Lipman, Y.; Chen, R. T. Q.; Ben-Hamu, H.; Nickel, M.; and Le, M. 2023. Flow Matching for Generative Modeling. In The Eleventh International Conference on Learning Representations

    2023

  12. [20]

    H.; Wang, Q.; Gong, Y.; and Glass, J

    Liu, A. H.; Wang, Q.; Gong, Y.; and Glass, J. 2024. A Closer Look at Neural Codec Resynthesis: Bridging the Gap between Codec and Waveform Generation. arXiv preprint arXiv:2410.22448. NeurIPS 2024 Audio Imagination Workshop

    2024 arXiv

  13. [21]

    Ma, Y.; Xia, H.; Gao, H.; Chen, W.; Ye, Y.; Yang, Y.; Chang, S.; Ding, M.; Li, Y.; Yuan, R.; Dixon, S.; and Benetos, E. 2026. CMI-RewardBench : Evaluating Music Reward Models with Compositional Multimodal Instruction. arXiv preprint arXiv:2603.00610. Accepted by ICML 2026

    2026 arXiv

  14. [22]

    MiniMax . 2026. MiniMax Music 2.6 : Four Stories We Want to Tell. https://www.minimax.io/news/music-26. Product release, accessed 2026-07-20

    2026

  15. [23]

    Mureka . 2026. Mureka API Platform Changelog. https://platform.mureka.ai/docs/en/changelog.html. Documents the enhanced mureka-8 release; accessed 2026-07-20

    2026

  16. [24]

    Ning, Z.; Chen, H.; Jiang, Y.; Hao, C.; Ma, G.; Wang, S.; Yao, J.; and Xie, L. 2025. DiffRhythm : Blazingly Fast and Embarrassingly Simple End-to-End Full-Length Song Generation with Latent Diffusion. arXiv preprint arXiv:2503.01183

    2025 arXiv

  17. [25]

    Peebles, W.; and Xie, S. 2023. Scalable Diffusion Models with Transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 4195--4205

    2023

  18. [26]

    San Roman, R.; Adi, Y.; Deleforge, A.; Serizel, R.; Synnaeve, G.; and D \'e fossez, A. 2023. From Discrete Tokens to High-Fidelity Audio Using Multi-Band Diffusion. In Advances in Neural Information Processing Systems, volume 36

    2023

  19. [27]

    Shulman, M. 2026. Suno v5.5 : More Expressive. More You. https://suno.com/blog/v5-5. Product release, accessed 2026-07-20

    2026

  20. [28]

    Suno . 2025. Introducing v5. https://suno.com/release-notes/introducing-v5-the-world-s-best-music-model. Product release note, accessed 2026-07-20

    2025

  21. [29]

    Tjandra, A.; Wu, Y.-C.; Guo, B.; Hoffman, J.; Ellis, B.; Vyas, A.; Shi, B.; Chen, S.; Le, M.; Zacharov, N.; Wood, C.; Lee, A.; and Hsu, W.-N. 2025. Meta Audiobox Aesthetics : Unified Automatic Quality Assessment for Speech, Music, and Sound. arXiv preprint arXiv:2502.05139

    2025 arXiv

  22. [30]

    Wu, D.; Lei, S.; Tan, W.; Li, G.; Wang, Y.; Zhang, H.; Zuo, L.; and Wu, Z. 2026. SongBench : A Fine-Grained Multi-Aspect Benchmark for Song Quality Assessment. arXiv preprint arXiv:2604.25937

    2026 arXiv

  23. [31]

    Yamamoto, R.; Song, E.; and Kim, J.-M. 2020. Parallel WaveGAN : A Fast Waveform Generation Model Based on Generative Adversarial Networks with Multi-Resolution Spectrogram. In 2020 IEEE International Conference on Acoustics, Speech and Signal Processing, 6199--6203

    2020

  24. [32]

    Yang, C.; Wang, S.; Chen, H.; Tan, W.; Yu, J.; and Li, H. 2025. SongBloom : Coherent Song Generation via Interleaved Autoregressive Sketching and Diffusion Refinement. In Advances in Neural Information Processing Systems, volume 38

    2025

  25. [33]

    Yang, D.; Xie, Y.; Yin, Y.; Wang, Z.; Yi, X.; Zhu, G.; Weng, X.; Xiong, Z.; Ma, Y.; Cong, D.; Liu, J.; Huang, Z.; Ru, J.; Huang, R.; Wan, H.; Wang, P.; Yu, K.; Wang, H.; Liang, L.; Zhuang, X.; Wang, Y.; Wang, D.; Guo, H.; Cao, J.; Ju, Z.; Liu, S.; Cao, Y.; Weng, H.; and Zou, Y...

    2026 arXiv

  26. [34]

    Yao, J.; Ma, G.; Xue, H.; Chen, H.; Hao, C.; Jiang, Y.; Liu, H.; Yuan, R.; Xu, J.; Xue, W.; Liu, H.; and Xie, L. 2025. SongEval : A Benchmark Dataset for Song Aesthetics Evaluation. arXiv preprint arXiv:2505.10793

    2025 arXiv

  27. [35]

    Zeghidour, N.; Luebs, A.; Omran, A.; Skoglund, J.; and Tagliasacchi, M. 2022. SoundStream : An End-to-End Neural Audio Codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30: 495--507

    2022

  28. [36]

    H.; Zhou, K.; Jiang, Y.; Tan, C.; Gao, Z.; Du, Z.; and Ma, B

    Zhang, C.; Ma, Y.; Chen, Q.; Wang, W.; Zhao, S.; Pan, Z.; Wang, H.; Ni, C.; Nguyen, T. H.; Zhou, K.; Jiang, Y.; Tan, C.; Gao, Z.; Du, Z.; and Ma, B. 2025 a . InspireMusic : Integrating Super Resolution and Large Language Model for High-Fidelity Long-Form Music Generation. arXi...

    2025 arXiv

  29. [37]

    Zhang, Y.; Li, M.; Long, D.; Zhang, X.; Lin, H.; Yang, B.; Xie, P.; Yang, A.; Liu, D.; Lin, J.; Huang, F.; and Zhou, J. 2025 b . Qwen3 Embedding : Advancing Text Embedding and Reranking Through Foundation Models. arXiv preprint arXiv:2506.05176

    2025 arXiv

  30. [38]

    Zhao, W.; Bai, L.; Rao, Y.; Zhou, J.; and Lu, J. 2023. UniPC : A Unified Predictor-Corrector Framework for Fast Sampling of Diffusion Models. In Advances in Neural Information Processing Systems, volume 36, 49842--49869. Curran Associates, Inc

    2023

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.