acceptodds
Under review as a conference paper at ICLR 2027

Don’t Describe, Demonstrate: Learning Transferable Music Edits from Before–After Examples

Abstract

Existing music editing methods mainly rely on text instructions, which are often insufficient for fine-grained or difficult-to-describe transformations. We introduce editing by demonstration, where a before–after music pair specifies an edit to be transferred to unseen inputs. This setting poses two key challenges: (1) translating a demonstrated transformation into an executable editing condition, and (2) isolating the transferable edit from reference-specific musical content. To address them, we propose Edit Condition Synthesis, which recovers a soft edit condition in the native conditioning space of a frozen flow-based editor. We further introduce a transition-aware semantic constraint that captures source-to-target changes and improves the transferability of the synthesized condition. Experiments on instrument and style editing show improved edit transfer and source preservation over existing text-guided and demonstration-based methods. An online demo with audio examples is available here.

Primary area:
generative models
Keywords:
Music Editing, Editing by Demonstration, Conditional Generation, Music Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_D6viOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces editing by demonstration for music: instead of a text instruction, a before–after audio pair specifies an edit, and the goal is to transfer that edit to unseen music. The proposed method, Edit Condition Synthesis, works with a frozen flow-matching music editor (MelodyFlow). Given a demonstration pair, the authors encode source and target, build two flow states sharing the same noise draw, and use their relative target velocity as the transition signal — the noise term cancels exactly, leaving a deterministic source-to-target direction (Eq. 11). A continuous condition is then optimized in the editor's native conditioning space so that the condition-induced relative dynamics match this demonstrated transition (Eq. 13), which makes the condition executable without retraining the generator. To keep the condition transferable rather than tied to the reference track, they adapt MuQ into a transition encoder with a contrastive objective over edit types and add a transition-consistency term 1 − cos between the demonstrated transition and the transition induced by the synthesized edit (Eqs. 16–18). Two paired benchmarks are constructed: instrument editing from Slakh2100-redux (7,500 pairs) and style editing from MusicCaps/MeLBench (8,400 pairs), each with ~10% no-op samples and piece-level splits. The method outperforms eight baselines (text-guided flow editors, adapted demonstration-based methods VISII/EditCLIP, and audio analogy) on automatic metrics, with ablations for each component and a 16-person user study.

Critical strengths

For a new task, the experimental protocol is careful and honest: two paired benchmarks with no-op controls and splits by musical piece, demonstration-based baselines (VISII, EditCLIP) re-implemented on the same editor rather than quoted from other domains, per-component ablations including the depth of MuQ adaptation, a λ sensitivity sweep, and a small subjective study. The AI-use statement is appreciated.

Critical weaknesses
  1. Transfer is demonstrated only among edit types the transition encoder was trained on (instrument add/remove/replace and seven styles). Since the contrastive objective defines the edit space, generalization to a genuinely novel transformation — arguably the main selling point of demonstration-based editing — is untested, and the paper offers no mechanism to detect when a demonstration falls outside the trained edit space.

  2. After listening to the audio examples on the authors' demo page, I am less convinced that the synthesized condition reliably captures the demonstrated edit. In several of the examples the edited output does not audibly reflect the transformation shown in the before–after pair: the change in the demonstrated direction is weak or hard to notice. I know qualitative listening is not a substitute for metrics, and the demo examples may not be representative, but for a paper whose central claim is that the recovered condition encodes the transferable transformation, what I heard on the demo page suggests a gap between the automatic numbers and perceived edit strength. The paper contains no failure analysis of partial transfers, and none of the reported metrics (FAD, CLAP, EIA) directly measures whether the demonstrated change is audible to a listener.

Questions for authors
  1. How long does condition synthesis take per demonstration (optimization steps, wall-clock time, memory), and how does output quality vary with the optimization budget? A convergence curve would substantially strengthen the practical case.

  2. Are the examples on the demo page representative of typical behavior? In several of them the edit is barely audible relative to the demonstrated pair. Could you report a failure analysis over the test set quantifying how often the transferred edit is only partial or effectively a no-op, and how often listeners can identify the demonstrated transformation in a blind A/B setting?

Reviewer_Pp4kOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a music editing method by providing before-after examples as controls instead of text instructions. The method is based on a pretrained music editing model, MelodyFlow, and optimizes a target editing condition from the provided example, which is then used to edit new music. Experiments show improvements in instrument and style editing tasks.

Critical strengths
  1. The paper explores an interesting and meaningful question. Using an example as a control is indeed more intuitive in many music editing scenarios. Problems like this are, in general, harder to define and evaluate.
Critical weaknesses
  1. The method does not sufficiently support the motivation. Although the system allows extracting editing controls from example pairs, it is unclear how fine-grained these controls can be. In training, the overall pipeline still relies on a text-aligned conditioning space. As illustrated by the authors in Figure 1, examples can sometimes be more expressive and more detailed than text prompts. There is no explicit mechanism for this. It is not shown in the experiments whether more flexible and fine-grained control can emerge or generalize.
  2. The experiments do not sufficiently support the motivation. The experiments contain predefined instrument and style editing tasks, which can be described as well by text instructions. Providing examples can also be ambiguous in the demonstrated case, because, e.g., examples of "instrument removal" can be ambiguous, because it can mean either removing the instrument from the example or maintaining the instruments in the example.
  3. The generation quality on the demo page is not convincing enough. The authors should include more music editing cases for at least a qualitative demonstration.
Questions for authors

I would appreciate it if the authors could address the concerns raised in the weakness section.

Reviewer_2YryOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes a music editing scheme by following a given demonstration pair. First, the work established a fine-tuned MuQ encoder that aims to capture the transition demonstrated by the given pair in its representation without being affected by the music content. Then, the work search for a condition representation that can guide the underlying frozen flow-matching based editor model (Melodyflow) to edit the source music following the transition demonstrated by the given pair. It is an optimization process governed by two consistency metrics: the flow transition consistency ensures the velocity induced by the condition matches the velocity between demonstration pairs at various timesteps; the edit semantic consistency regulates the optimized condition to represent the same transition as the one demonstrated by the given pair under the fine-tuned MuQ representation. The quantitative evaluation shows the performance of proposed scheme and ablation study justified the necessity of each component.

Critical strengths
  • Once the condition representation is learnt, it can be reused for the same editing operation without extra effort.
Critical weaknesses
  • The conditioning representation is optimized on single demonstration pair. It may not be generalizable to other (x^s_d, x^t_d) pair for the same editing operation. This could be linked to the observation that, in the demonstration page, removing instruments worked well, but the results of adding instrument and style transfer are not so convincing.
  • The methodology is capped by the conditioning alignment of underlying frozen editor.
Questions for authors
  • In Figure2, human annotated prompt is fed into Melodyflow, but it is an editor that receives audio and condition then output edited audio. Does it actually mean the prompt is fed into the text encoder(T5) within the melody flow? Is there any other conversion before the alignment is evaluated with the difference vector resulted from fine-tuned MuQ?
  • In Figure 2, the part about edit semantic consistency, are the "before" and "after" samples are actually the demonstration pair?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal Updates and Discussion

Dear Program Chairs, Senior Area Chairs, Area Chair, and Reviewers,

Thank you for the thoughtful reviews and for taking the time to consider our responses and listen to the updated demos. As the discussion period draws to a close, we would like to summarize the main updates and outcomes.

  • Method clarification and demo correction (Reviewer 2Yry). We revised Figure 2 to clarify how the editing condition is learned and reused, and explained how transfer depends on the frozen editor. We also corrected Example 09 to show the demonstration-guided and text-conditioned outputs side by side. The reviewer appreciated the revisions, acknowledged the gains over the baselines, and maintained their positive rating.

  • Demonstration-based control and evaluation (Reviewer Pp4k). We clarified how audio demonstrations refine the text-initialized condition and guide instrument removal while preserving unrelated content in the query. We cited the comparison with text prompts (main paper Table 4) and our higher edit audibility ratings (rebuttal Table 14). Source preservation varies across examples, and New 3’s listener ratings reflect the same loss of content the reviewer heard. For New 6, we clarified that the two demonstration clips need not have similar musical content and added Example A to illustrate a more focused style edit.

  • Additional evidence and remaining limitations (Reviewer D6vi). The reviewer found that the additional experiments addressed their main concerns and raised their rating from 2 to 3. We clarified that generalization to genuinely new edit operations remains unverified and that the current method cannot detect demonstrations outside its supported edit space.

We appreciate the constructive discussion and your consideration of the revised manuscript.

Best regards,
The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.