Pith. sign in

REVIEW 4 major objections 6 minor 3 cited by

AudioMorphix: Training-free audio editing with diffusion probabilistic models

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A training-free method edits audio by morphing spectrogram latents between source and reference sounds.

desk verdict The benchmark and code are real contributions, but the evaluation never supplies the masks the method needs, so the headline results don't test the actual contribution. read the letter →

arxiv 2505.16076 v1 pith:PIVOGWKQ submitted 2025-05-21 eess.AS

Jinhua Liang , Yuanzhe Chen , Yi Yuan , Dongya Jia , Xiaobin Zhuang , Zhuo Chen , Yuping Wang , Yuxuan Wang This is my paper · ORCID
classification eess.AS
keywords audioeditingdiffusionmodelstraining-freeguidancespectrogrammaskslatentmorphingenergyfunctionsself-attentioncachingbenchmark
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that precise audio editing can be done without any training by treating sound as a point on a latent manifold and editing as a morphing cycle: mixing is morphing two sounds together, and separating is demorphing. The proposed system, AudioMorphix, runs a frozen text-to-audio diffusion model, replaces the noisy latent of the target region by a spherical interpolation of the source and reference latents, and then guides each denoising step with energy functions that enforce content consistency in the masked region and contrast with the removed sound. It also caches self-attention keys and values from the reference pass to keep fine detail. On the new AudioSet-E benchmark, the paper reports lower FAD and KL scores than DDIM inversion, DDPM inversion, and AUDIT for addition, removal, and replacement, plus qualitative results for time shifting, stretching, and pitch shifting. If the latent-geometry premise holds, this gives a general recipe for region-specific, reference-guided audio editing with no per-task training.

What carries the argument

The central object is the morphing cycle on the diffusion latent manifold. AudioMorphix first inverts raw and reference spectrograms into their noised latents at the final diffusion step, then combines them with spherical linear interpolation (SLERP) to obtain the starting latent for the edited output; for removal, it optimizes the two component latents with SGD so that their SLERP matches the mixture latent while the components stay far apart. Each denoising step is then rectified by energy functions built from masked cosine similarities at selected decoder self-attention layers, and the key/value pairs from the reference trajectory are substituted into those layers. This converts editing into trajectory manipulation rather than retraining.

What would settle it

Encode a known two-sound mixture and its isolated components, then decode the SLERP latent at the correct mixing ratio; if the decoded audio does not match the true mixture (measured by FAD), the morphing premise is wrong. Likewise, remove one component via the paper's latent optimization and check whether the residual audio still contains the removed sound.

Watch

Extended reading notes

Core claim

AudioMorphix claims that six audio editing tasks—addition, removal, replacement, time shifting, time stretching, and pitch shifting—reduce to one operation in the latent space of a pretrained diffusion model: move the latent of the raw audio along a great-circle arc toward the latent of the reference audio, then denoise under stepwise energy guidance. The guiding energies measure cosine similarity between intermediate decoder features in the masked time-frequency region (consistency) and between the generated and reference audio (contrast), and the self-attention key/value cache transfers reference detail into the output. The result, according to the paper, is an editor that changes only the masked region and keeps the rest of the recording intact, and that reports better fidelity and distribution matching than the comparison methods on the new AudioSet-E benchmark.

Load-bearing premise

The latent of a mixed audio is a simple blend of the latents of its parts, so adding a sound to a latent or subtracting it by optimization yields the intended audible change.

Editorial extensions

If this is right

  • A user can edit a specific time-frequency patch simply by drawing a mask and supplying a reference sound; unmasked regions remain untouched.
  • The same frozen diffusion model, with no fine-tuning, handles addition, removal, replacement, time shifting/stretching, and pitch shifting.
  • Because guidance is defined as energy functions on decoder features, the method transfers across prediction objectives and across pretrained backbones such as AudioLDM and Tango.
  • The AudioSet-E benchmark, with roughly 4,700 curated samples, gives the community a way to compare audio editors on addition, removal, and replacement using both objective and subjective scores.

Reading between the lines

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

  • If the morphing premise generalizes, the same mask-and-reference recipe could apply to music stems, sound effects, and speech wherever a frozen diffusion model provides a meaningful latent manifold.
  • The energy functions could be replaced by explicit hard constraints in latent space, which might yield faster or more stable edits than gradient-based guidance.
  • The removal task's reliance on a second reference track to disambiguate the optimization suggests that a single reference is insufficient; a natural test is whether a contrastive loss can remove that extra track.
  • The paper's claim that AudioMorphix can surpass ground truth in consistency stems from avoiding clipping and concatenation artifacts; this is testable on longer, multi-event mixtures where such artifacts are more visible.
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

4 major / 6 minor

Summary. The paper proposes AudioMorphix, a training-free audio editing method that manipulates the latents of a pretrained latent diffusion model to perform localized time-frequency edits. The method casts audio addition, removal, and replacement as a morphing cycle in latent space, uses energy-based guidance during sampling with self-attention key/value substitution for detail preservation, and introduces a new benchmark, AudioSet-E, with FAD/KL and subjective evaluations. The paper claims state-of-the-art results on addition, removal, and replacement, alongside qualitative demonstrations of time shifting, time stretching, and pitch shifting.

Significance. If the claimed local time-frequency editing capability were properly verified, AudioMorphix would be a useful contribution: it is training-free, compatible with at least two pretrained text-to-audio models (AudioLDM and Tango), and ships with a new benchmark and evaluation kit. The AudioSet-E dataset and code release are valuable assets for the community. However, the current quantitative evaluation does not substantiate the central claim, because the method is evaluated without the masking information that its own energy functions require.

major comments (4)
  1. [Section 4.2 and Appendix B] The method as described depends on binary masks m_c and m_r: Eq. (10) uses masked indexing F^c_{t,l}[m_c], and Eqs. (11)-(14) build the guidance terms from these masked features. However, Appendix B states, 'For a fair comparison, our AudioMorphix was provided with no masking information same as the other editing methods.' This makes the masked indexing undefined and means that Tables 1 and 2, which are the paper's main quantitative evidence, do not exercise the local-editing mechanism that is the central claim. The authors should either derive and supply masks for the AudioSet-E benchmark and re-run the evaluation, or clearly separate the unmasked variant as a baseline and report local-editing results with masks, including a description of how masks are obtained for each task.
  2. [Section 5.2 and Table 2] The text states that 'AudioMorphix surpasses the ground truth in consistency', but Table 2 reports ground-truth consistency of 60.25 versus AudioMorphix at 56.73. This claim is numerically contradicted by the paper's own table and should be corrected. The subsequent explanation that the end-to-end generative approach avoids clipping and concatenation artifacts is not supported by the reported numbers.
  3. [Section 3.2 and Table 3] The configuration used in the main comparison is ambiguous. The 'Our method (w/ AudioLDM)' row in Table 1 reports FAD 5.58 and KL 0.83 for addition, exactly matching the ablation row in Table 3 with text description but without tangent-space projection. Yet Section 3.2 states that the default setting enables tangent-space projection g_tan. Table 3 also shows that adding tangent-space projection under text guidance degrades all three tasks (addition FAD 6.10 vs 5.58, removal 3.38 vs 2.83, replacement 5.91 vs 2.67). The authors should clarify which configuration produced Tables 1 and 2, and reconcile the ablation result with the description of Algorithm 1 and the claimed benefits of manifold preservation.
  4. [Section 3.2, Eq. (8) and the zm ∝ zr + zc premise] The morphing-based addition and removal mechanisms rest on the premise that the latent of a mixture is captured by spherical interpolation of component latents and that demorphing can be achieved by optimizing this latent. This premise is asserted as an empirical observation but is not quantitatively validated anywhere in the paper. Given that the benchmark evaluation does not isolate this component, the validity of the morphing mechanism is not established. Please provide a direct validation, for example by comparing cosine similarity between SLERP(z_c, z_r) and the encoded mixture latent, and by showing that the demorphing optimization in Algorithm 1 produces latents that are closer to the true components than random or identity initialization.
minor comments (6)
  1. [Section 3.1] The heading text contains 'AudioMorhix', which should be 'AudioMorphix'.
  2. [Section 2.1] There are typos: 'd-dimentional' should be 'd-dimensional' and 'rewroten' should be 'rewritten'.
  3. [Section 5.2] The sentence 'This suggests that AudioMorphix provides more accurate and realistic image edits compared to...' refers to audio, so 'image edits' should be 'audio edits'.
  4. [Appendix E, Figure 7 caption] The caption contains 'AudioMophix', which should be 'AudioMorphix'.
  5. [Eq. (10) and Eq. (11)] The notation for masked indexing is imprecise: it should be stated explicitly how the binary masks m_c and m_r are broadcast onto the intermediate features F^c_{t,l} and F^r_{t,l}, and the subscript/superscript usage in Eq. (11) should be made consistent.
  6. [Section 5.1] The informal phrase 'We didn't implement DreamBooth...' should be replaced with a more formal 'We did not implement...'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the editing mechanism is a training-free energy-guided procedure evaluated against external benchmarks; the unvalidated latent-geometry premise and the mask-free evaluation are correctness risks, not circularity.

full rationale

The paper's derivation chain is not circular. AudioMorphix takes a frozen text-to-audio diffusion model and defines edits through (i) SLERP interpolation/demorphing of latents (Eq. 9, Algorithm 1), (ii) energy guidance on intermediate features (Eqs. 10-14), and (iii) key/value substitution in self-attention (Sec. 4.4). No parameter is fitted to the AudioSet-E targets: FAD and KL are computed with external embeddings, and subjective ratings are provided by independent raters, so the quantitative claims are not forced by construction. The latent-geometry premise (Eq. 8; 'zm ∝ zr + zc' in Sec. 3.2) is an empirical assumption with no proof, and the morphing prior may be wrong, but an unverified assumption is not a reduction to the output. Self-citations (AudioLDM as frozen backbone; Yuan et al. for metric protocol) are normal use of external artifacts, not load-bearing self-citation. One protocol issue is flagged for completeness: Appendix B says 'our AudioMorphix was provided with no masking information same as the other editing methods,' while Eqs. 10-14 are defined over binary masks m_c and m_r; this means the main benchmark does not exercise the mask-based local-editing mechanism. That is an internal-validity/correctness concern, not circularity, and it does not change the circularity score.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The central claim relies on several empirical assumptions about the latent space of the pretrained diffusion model, plus several unreported hyperparameters (energy weights, penalty term). No new physical entities are introduced; the 'morphing cycle' is a conceptual metaphor.

free parameters (9)
  • w_content
    Weight for the content-consistency energy in Eqs (13)-(14); value not reported in the paper, set by hand.
  • w_edit
    Weight for the edit-consistency/contrast energy in Eqs (13)-(14); value not reported.
  • niters = 100
    Number of latent optimization iterations in Algorithm 1; chosen empirically in Section 3.2.
  • lr = 1e-4
    Learning rate for SGD in Algorithm 1; chosen empirically.
  • CFG guidance scale = 1 (AudioLDM), 1.2 (Tango)
    Classifier-free guidance scale in Appendix B, set per backbone.
  • KV substitution layers = 2 and 3
    Self-attention layers in the decoder used for key/value substitution, Section 4.4.
  • Similarity scaling in Eq (10) = 0.5
    Hand-chosen constant to map cosine similarity to [0,1].
  • Energy normalization in Eq (11) = 4
    Hand-chosen constant in the denominator of S_consist.
  • Penalty P in Algorithm 1
    The penalty term used in latent optimization is not defined in the paper; a free design choice.
assumptions (5)
  • domain assumption Latent Spatial Consistency (Eq 8): sim(zi,zj) ∝ sim(xi,xj)
    Stated as a premise in Section 3.1, citing Yang et al. 2024; the paper provides no quantitative evidence for audio latents.
  • domain assumption Mixture latent additivity: zm ∝ zr + zc
    Stated as 'According to our observation' in Section 3.2; unproven and central to the morphing cycle.
  • domain assumption SLERP on latents produces meaningful audio mixtures
    Assumed in Eq (9) following image morphing works (He et al. 2024, Yang et al. 2024); no audio-specific validation.
  • domain assumption Self-attention features from U-Net decoder are valid signals for consistency and contrast
    Used in Eqs (10)-(14) without justification beyond analogy to image editing.
  • ad hoc to paper Hand-crafted energy functions (Eqs 11-14) provide correct guidance direction
    The energy forms are invented for this paper and their sign/direction for removal is ambiguous (S_contrast is defined as similarity).

how reviews work

0 comments
Cite this review

Pith. "Pith review of AudioMorphix: Training-free audio editing with diffusion probabilistic models." pith.science (2026). https://pith.science/paper/PIVOGWKQ

@misc{pith2026250516076,
  author       = {Pith},
  title        = {Pith review of: AudioMorphix: Training-free audio editing with diffusion probabilistic models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PIVOGWKQ}},
  note         = {Machine review of arXiv:2505.16076}
}
read the original abstract

Editing sound with precision is a crucial yet underexplored challenge in audio content creation. While existing works can manipulate sounds by text instructions or audio exemplar pairs, they often struggled to modify audio content precisely while preserving fidelity to the original recording. In this work, we introduce a novel editing approach that enables localized modifications to specific time-frequency regions while keeping the remaining of the audio intact by operating on spectrograms directly. To achieve this, we propose AudioMorphix, a training-free audio editor that manipulates a target region on the spectrogram by referring to another recording. Inspired by morphing theory, we conceptualize audio mixing as a process where different sounds blend seamlessly through morphing and can be decomposed back into individual components via demorphing. Our AudioMorphix optimizes the noised latent conditioned on raw input and reference audio while rectifying the guided diffusion process through a series of energy functions. Additionally, we enhance self-attention layers with a cache mechanism to preserve detailed characteristics from the original recordings. To advance audio editing research, we devise a new evaluation benchmark, which includes a curated dataset with a variety of editing instructions. Extensive experiments demonstrate that AudioMorphix yields promising performance on various audio editing tasks, including addition, removal, time shifting and stretching, and pitch shifting, achieving high fidelity and precision. Demo and code are available at this url.

Figures

Figures reproduced from arXiv: 2505.16076 by the authors.

Figure 1
Figure 1. Audio editing tasks of which our AudioMorphix is capable with no training cost. We use [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed AudioMorphix. The AudioMorphix generates the clean latent [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. A schematic overview of our AudioMorphix in comparison with DDIM inversion [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Illustration of adapting self-attention layers to preserve [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Examples of three audio manipulations: time stretching, time shifting, and pitch shifting. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Qualitative evaluation between our AudioMorphix and other audio editing methods. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Ablation study on the impact of SLERP in the audio addition task. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Subjective evaluation across different tasks, with “top tie” denoting the highest value among AUDIT, DDIM inversion, and DDPM [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. RIME: Enabling Large-Scale Agentic Music Post-Production

    cs.SD 2026-07 conditional novelty 6.0 of 10

    RIME generates 3,000 synthetic music post-production edit triples and shows that current multimodal LLM agents can recover edit structure but often fail to set effect parameters correctly.

  2. RFM-Editing 2: Text-Guided Audio Editing with Rectified Flow Matching and Coarse-to-Fine Diffusion Transformers

    cs.SD 2026-06 unverdicted novelty 6.0 of 10

    Hybrid two-stage diffusion transformer architecture for instruction-guided audio editing via rectified flow that performs joint attention at low resolution then alternates joint and cross-attention at high resolution ...

  3. SemanticAudio: Audio Generation and Editing in Semantic Space

    eess.AS 2026-01 conditional novelty 6.0 of 10

    SemanticAudio improves text-to-audio alignment by generating a compact semantic plan first with a Flow Matching planner and then rendering acoustic latents from that plan, and it performs training-free audio editing b...

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages · cited by 3 Pith papers

  1. [1]

    Audio texture manipulation by exemplar- based analogy

    [Chenget al., 2025 ] Kan Jen Cheng, Tingle Li, and Gopala Anumanchipalli. Audio texture manipulation by exemplar- based analogy. In2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5,

    work page 2025

  2. [6]

    Gemmeke, Daniel P

    [Gemmekeet al., 2017 ] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Chan- ning Moore, Manoj Plakal, and Marvin Ritter. Audio Set: An ontology and human-labeled dataset for audio events. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 776–780, New Or- leans, LA, March

    work page 2017

  3. [8]

    Adapting frechet audio distance for generative music evaluation

    [Guiet al., 2024 ] Azalea Gui, Hannes Gamper, Sebastian Braun, and Dimitra Emmanouilidou. Adapting frechet audio distance for generative music evaluation. In2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1331–1335,

    work page 2024

  4. [9]

    InstructME: an instruction guided music edit framework with latent diffusion models

    [Hanet al., 2024 ] Bing Han, Junyu Dai, Weituo Hao, Xinyan He, Dong Guo, Jitong Chen, Yuxuan Wang, Yanmin Qian, and Xuchen Song. InstructME: an instruction guided music edit framework with latent diffusion models. In2024 Pro- ceedings of the Thirty-Third International Joint Conference on Artificial Intelligence,

    work page 2024

  5. [10]

    Manifold preserving guided diffusion

    [Heet al., 2024 ] Yutong He, Naoki Murata, Chieh-Hsin Lai, Yuhta Takida, Toshimitsu Uesaka, Dongjun Kim, Wei- Hsiang Liao, Yuki Mitsufuji, J Zico Kolter, Ruslan Salakhutdinov, and Stefano Ermon. Manifold preserving guided diffusion. InThe Twelfth International Conference on Learning Representations,

    work page 2024

  6. [11]

    Prompt-to-prompt image editing with cross-attention con- trol

    [Hertzet al., 2023 ] Amir Hertz, Ron Mokady, Jay Tenen- baum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention con- trol. InThe Eleventh International Conference on Learning Representations,

    work page 2023

  7. [12]

    Denoising diffusion probabilistic models

    [Hoet al., 2020 ] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InProceedings of the 34th International Conference on Neural Information Processing Systems,

    work page 2020

  8. [14]

    An edit friendly ddpm noise space: Inversion and manipula- tions

    [Huberman-Spiegelglaset al., 2024 ] Inbar Huberman- Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipula- tions. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12469–12478,

    work page 2024

Show all 34 references
  1. [15]

    Plumbley, Huy Phan, and Emmanouil Benetos

    [Lianget al., 2024 ] Jinhua Liang, Huan Zhang, Haohe Liu, Yin Cao, Qiuqiang Kong, Xubo Liu, Wenwu Wang, Mark D. Plumbley, Huy Phan, and Emmanouil Benetos. WavCraft: Audio editing and generation with large language models. InICLR 2024 Workshop on LLM Agents,

    2024

  2. [16]

    Plumbley, Huy Phan, and Emmanouil Benetos

    [Lianget al., 2025 ] Jinhua Liang, Xubo Liu, Wenwu Wang, Mark D. Plumbley, Huy Phan, and Emmanouil Benetos. Acoustic prompt tuning: Empowering large language mod- els with audition capabilities.IEEE Transactions on Audio, Speech and Language Processing, 33:949–961,

    2025

  3. [17]

    Arrange, inpaint, and refine: steerable long-term music audio generation and editing via content-based con- trols

    [Linet al., 2024 ] Liwei Lin, Gus Xia, Yixiao Zhang, and Jun- yan Jiang. Arrange, inpaint, and refine: steerable long-term music audio generation and editing via content-based con- trols. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence,

    2024

  4. [18]

    AudioLDM: Text-to-Audio Generation with La- tent Diffusion Models

    [Liuet al., 2023 ] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. AudioLDM: Text-to-Audio Generation with La- tent Diffusion Models. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and J...

    2023

  5. [19]

    Plumbley, and Wenwu Wang

    [Liuet al., 2025 ] Xubo Liu, Zhongkai Zhu, Haohe Liu, Yi Yuan, Meng Cui, Qiushi Huang, Jinhua Liang, Yin Cao, Qiuqiang Kong, Mark D. Plumbley, and Wenwu Wang. WavJourney: Compositional audio creation with large lan- guage models.IEEE Transactions on Audio, Speech and Language ...

    2025

  6. [20]

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

    [Luet al., 2023 ] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-solver++: Fast solver for guided sampling of diffusion probabilistic mod- els,

    2023

  7. [21]

    Tango 2: Aligning diffusion-based text- to-audio generative models through direct preference opti- mization

    [Majumderet al., 2024 ] Navonil Majumder, Chia-Yu Hung, Deepanway Ghosal, Wei-Ning Hsu, Rada Mihalcea, and Soujanya Poria. Tango 2: Aligning diffusion-based text- to-audio generative models through direct preference opti- mization. InACM Multimedia 2024,

    2024

  8. [22]

    Zero-shot unsupervised and text-based audio editing using DDPM inversion

    [Manor and Michaeli, 2024] Hila Manor and Tomer Michaeli. Zero-shot unsupervised and text-based audio editing using DDPM inversion. InProceedings of the 41st International Conference on Machine Learning, volume 235 ofProceedings of Machine Learning Research, pages 34603–34629,

    2024

  9. [23]

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

    [Mokadyet al., 2023 ] Ron Mokady, Amir Hertz, Kfir Aber- man, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In 2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 6038–6047,

    2023

  10. [24]

    DragonDiffusion: Enabling Drag-style Manipulation on Diffusion Models

    [Mouet al., 2024 ] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. DragonDiffusion: Enabling Drag-style Manipulation on Diffusion Models. InThe Twelfth International Conference on Learning Representa- tions,

    2024

  11. [25]

    Investigating Personalization Methods in Text to Music Generation, September

    [Plitsiset al., 2023 ] Manos Plitsis, Theodoros Kouzelis, Georgios Paraskevopoulos, Vassilis Katsouros, and Yannis Panagakis. Investigating Personalization Methods in Text to Music Generation, September

    2023

  12. [26]

    [Ruizet al., 2023 ] Nataniel Ruiz, Yuanzhen Li, Varun Jam- pani, Yael Pritch, Michael Rubinstein, and Kfir Aberman

    arXiv:2309.11140 [cs, eess]. [Ruizet al., 2023 ] Nataniel Ruiz, Yuanzhen Li, Varun Jam- pani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In2023 IEEE/CVF Confer- ence on Computer Visio...

    2023 arXiv

  13. [27]

    Progressive distillation for fast sampling of diffusion mod- els

    [Salimans and Ho, 2022] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion mod- els. InInternational Conference on Learning Representa- tions,

    2022

  14. [28]

    Audiobox: Unified audio generation with natural language prompts,

    [Vyaset al., 2023 ] Apoorv Vyas, Bowen Shi, Matthew Le, Andros Tjandra, Yi-Chiao Wu, Baishan Guo, Jiemin Zhang, Xinyue Zhang, Robert Adkins, William Ngan, Jeff Wang, Ivan Cruz, Bapi Akula, Akinniyi Akinyemi, Brian Ellis, Rashel Moritz, Yael Yungster, Alice Rakotoarison, Liang ...

    2023

  15. [29]

    AUDIT: Audio Editing by Following Instructions with Latent Dif- fusion Models

    [Wanget al., 2023 ] Yuancheng Wang, Zeqian Ju, Xu Tan, Lei He, Zhizheng Wu, Jiang Bian, and sheng zhao. AUDIT: Audio Editing by Following Instructions with Latent Dif- fusion Models. In A. Oh, T. Neumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Ne...

    2023

  16. [30]

    SV4D: Dynamic 3d content generation with multi-frame and multi-view con- sistency

    [Xieet al., 2025 ] Yiming Xie, Chun-Han Yao, Vikram V oleti, Huaizu Jiang, and Varun Jampani. SV4D: Dynamic 3d content generation with multi-frame and multi-view con- sistency. InThe Thirteenth International Conference on Learning Representations,

    2025

  17. [31]

    IMPUS: Image morphing with perceptually-uniform sampling using diffusion models

    [Yanget al., 2024 ] Zhaoyuan Yang, Zhengyang Yu, Zhiwei Xu, Jaskirat Singh, Jing Zhang, Dylan Campbell, Peter Tu, and Richard Hartley. IMPUS: Image morphing with perceptually-uniform sampling using diffusion models. In The Twelfth International Conference on Learning Repre- se...

    2024

  18. [32]

    Plumbley, and Wenwu Wang

    [Yuanet al., 2023 ] Yi Yuan, Haohe Liu, Jinhua Liang, Xubo Liu, Mark D. Plumbley, and Wenwu Wang. Leveraging pre-trained audioldm for sound generation: A benchmark study. In2023 31st European Signal Processing Conference (EUSIPCO), pages 765–769,

    2023

  19. [33]

    Inversion-based Style Transfer with Diffusion Models

    [Zhanget al., 2023 ] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Chang- sheng Xu. Inversion-based Style Transfer with Diffusion Models. InProceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR)...

    2023

  20. [34]

    Utilizing the times- tamps of sound events in AudioSet-SL, we mixed 2-3 audio tracks together with or without the selected sound events

    Appendix A Dataset curation We curated a new dataset to evaluate various audio editing tasks, including addition, removal, and replacement, based on the temporally strong subset of the AudioSet dataset (AudioSet-SL) [Gemmekeet al., 2017 ]. Utilizing the times- tamps of sound e...

    2017

  21. [2017]

    [Ghosalet al., 2023 ] Deepanway Ghosal, Navonil Majumder, Ambuj Mehrish, and Soujanya Poria

    IEEE. [Ghosalet al., 2023 ] Deepanway Ghosal, Navonil Majumder, Ambuj Mehrish, and Soujanya Poria. Text-to-audio gener- ation using instruction guided latent diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, MM ’23, page 3590–3598,

    2023

  22. [2020]

    Make-an- audio: Text-to-audio generation with prompt-enhanced diffusion models

    [Huanget al., 2023 ] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an- audio: Text-to-audio generation with prompt-enhanced diffusion models. InProceedings of the 40th International Conference...

    2023

  23. [2021]

    An image is worth one word: Personal- izing text-to-image generation using textual inversion

    [Galet al., 2023 ] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personal- izing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Repre...

    2023

  24. [2023]

    Diffusion models beat GANs on image synthe- sis.Advances in Neural Information Processing Systems, 11:8780–8794,

    [Dhariwal and Nichol, 2021] Prafulla Dhariwal and Alex Nichol. Diffusion models beat GANs on image synthe- sis.Advances in Neural Information Processing Systems, 11:8780–8794,

    2021

  25. [2024]

    Simple and controllable music gener- ation

    [Copetet al., 2023 ] Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi, and Alexandre D´efossez. Simple and controllable music gener- ation. InThirty-seventh Conference on Neural Information Processing Systems,

    2023

  26. [2025]

    Style injection in diffusion: A training-free ap- proach for adapting large-scale diffusion models for style transfer

    [Chunget al., 2024 ] Jiwoo Chung, Sangeek Hyun, and Jae- Pil Heo. Style injection in diffusion: A training-free ap- proach for adapting large-scale diffusion models for style transfer. In2024 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 8795–8805,

    2024

Pith tools

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