acceptodds
Under review as a conference paper at ICLR 2027

Bridge Adapters: Learning Restoration Bridges for Music Restoration

Abstract

Pretrained music generators learn powerful priors for clean, perceptually realistic audio, but restoration requires more than generation: the output must remain faithful to a degraded recording. Existing approaches either train degraded-to-clean restoration flows from scratch, or reuse pretrained generators by conditioning their original noise-to-clean trajectory on the degraded input. We propose Bridge Adapters, a parameter-efficient alternative that changes the transport itself. Starting from a frozen foundation music generator, we train small degradation-specific adapters with a degraded-to-clean flow-matching objective, converting the pretrained generative vector field into a restoration bridge while leaving the backbone unchanged. On the SonicMaster benchmark, spanning nineteen music degradations, Bridge Adapters outperform the prompted SonicMaster all-in-one restoration model across paired reconstruction metrics and Fr'echet Audio Distance, while requiring only a few thousands training steps per degradation. Compared with a ControlNet-style baseline over the same pretrained backbone, adapting the transport yields better restoration quality with substantially fewer sampling steps, reaching strong performance in fewer than five steps. We further show that the adapted bridges inherit the strength of the pretrained prior: stronger music backbones improve both reconstruction and perceptual quality. These results suggest that restoration bridges need not necessarily be learned from scratch; for music restoration, many are only a small adapter away from a pretrained generative model.

Primary area:
generative models
Keywords:
Generative Models, Diffusion Models, Flow Matching, Restoration, Music Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UNQzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Bridge Adapters, which convert a frozen pretrained music generator from a Gaussian-to-clean transport into a degraded-to-clean restoration transport using LoRA modules and a paired flow-matching objective. The main targeted system trains one rank-128 adapter per degradation on ACE-Step 1.5 XL, while a shared blind adapter and several alternative backbones are also evaluated. On 19 independently applied synthetic SonicMaster degradations, Bridge Adapters outperform prompted SonicMaster in average SNR, SI-SDR, and FMA-Pop FAD, and improve paired fidelity over same-backbone ControlNet and capacity-matched Conditional LoRA baselines. The evaluation also covers adapter capacity, backbone choice, solver steps, and parameter cost, but includes no formal listening study or real/mixed-degradation test.

Critical strengths

1.Convincing controlled comparison of transport design. The capacity-matched Conditional LoRA uses the same backbone, adapter capacity, data, training budget, and solver. Its comparison with Bridge Adapters provides credible evidence that starting from the degraded latent improves paired fidelity over using it only as conditioning.

2.Detailed diagnostic evaluation. The paper reports per-degradation results, uncertainty, solver sweeps, multiple backbones and adapter ranks, codec effects, and both active and stored parameter counts. The limitations of the specialist setting are stated clearly.

Critical weaknesses

1.Net perceptual restoration quality is not established. The unprocessed input has higher average SNR and SI-SDR than every restoration system, while the conditional baselines have better FMA-Pop FAD than Bridge Adapters. A codec round trip alone improves FMA-Pop FAD while substantially reducing paired fidelity. Without a formal listening study, it is unclear whether Bridge Adapters produce a net audible improvement over the input or competing systems.

2.The strongest setting is narrow. The main system assumes a known single synthetic degradation and stores 19 specialist adapters totaling 3.19B parameters. The blind comparison changes backbone, capacity, and training budget and lacks a matched blind Conditional LoRA. Mixed, unknown real-world, and compound degradations are not evaluated.

Questions for authors

1.In a loudness-matched listening study, do listeners prefer Bridge Adapters over the degraded input, SonicMaster, and Conditional LoRA?

2.For each degradation, and after excluding the pure-volume case, how often does the method improve over the degraded input in SNR and SI-SDR?

3.Does the source-endpoint advantage remain for a blind Conditional LoRA matched to the blind Bridge Adapter in backbone, capacity, data, and training budget?

2 comments

We thank the reviewer for their comments and feedback. We hope our responses increase their confidence in the practical benefits of our approach.

  • Question: In a loudness-matched listening study, do listeners prefer Bridge Adapters over the degraded input, SonicMaster, and Conditional LoRA?

  • Answer: Yes. As shown in Table 1, Bridge Adapters are consistently preferred over the degraded input (65.1% vs. 12.4%) and SonicMaster (68.2% vs. 4.5%). On the subset where the input degradation is audible, preference for Bridge Adapters rises to 73.9%, while the degraded input is preferred only 2.1% of the time. Consistent with our perceptual metrics, Bridge Adapters and Blind Conditional LoRA perform closest to one another, with Bridge Adapters still holding a clear edge (30.7% vs. 25.3%).

**Table 1: Human restoration preferences **

Method AMethod BPrefer ANo PreferencePrefer B
BridgeDegraded input65.1%22.6%12.4%
BridgeSonicMaster68.2%27.3%4.5%
BridgeBlind Conditional LoRA30.7%44.0%25.3%
Blind Conditional LoRASonicMaster76.5%11.8%11.8%
Prespecified subset: input with audible degradation
BridgeDegraded input73.9%24.0%2.1%

  • Question: For each degradation, and after excluding the pure-volume case, how often does the method improve over the degraded input in SNR and SI-SDR?

  • Answer: Excluding the pure-volume degradation, Bridge Adapters outperform the degraded input 42.3% of the time in SI-SDR and 42.6% of the time in SNR (40.1% and 45.0% across all 19 degradation families, respectively). As detailed in Table 2, win rates are highest on severe additive, reverberant, and spectral degradations (e.g., real, mud, vocal, mic, and boom), whereas subtle stereo/EQ perturbations (e.g., airy, bright, stereo) naturally incur waveform-level phase shifts from codec re-synthesis despite improving perceptual quality.

Table 2: Win rates of Bridge restoration over the degraded input across all 19 degradations

DegradationClipsSI-SDR better (%)SNR better (%)
airy9960.0%0.0%
big99648.9%24.2%
boom99665.2%55.4%
bright9960.0%0.1%
clarity99626.7%18.1%
clip9966.9%99.4%
comp99636.7%69.7%
dark9962.8%27.8%
mic99679.1%67.0%
mix99663.4%30.6%
mud99687.3%79.5%
punch9969.4%10.1%
real99690.9%95.3%
small99647.5%25.5%
stereo9960.0%4.5%
vocal99683.4%66.4%
volume9960.0%90.0%
warm99657.0%48.4%
xband99655.9%44.0%
All 1918,92440.1%45.0%
Without volume17,92842.3%42.6%

  • Question: Does the source-endpoint advantage remain for a blind Conditional LoRA matched to the blind Bridge Adapter in backbone, capacity, data, and training budget?

  • Answer: Yes. When matching backbone, adapter capacity, data, and training budget in the blind setting, the exact same pattern holds as in the prompt-conditioned case: the blind Bridge Adapter (Ours_blind) consistently outperforms Blind Conditional LoRA across all fidelity metrics (Mel distance, SNR, and SI-SDR) while maintaining comparable perceptual quality (FAD_FMA-Pop), as shown in Table 3.

Table 3: Comparison of Blind Conditional LoRA and Ours_blind matched in backbone, capacity, data, and training budget.

MethodMel ↓\downarrowSNR ↑\uparrowSI-SDR ↑\uparrowFADFMA-Pop\text{FAD}_{\text{FMA-Pop}} ↓\downarrow
Blind Conditional LoRA0.4952.140.310.206
Oursblind\text{Ours}_{\text{blind}}0.4902.790.730.213

  • Question: The strongest setting is narrow.

  • Answer: Evaluating on well-defined, paired degradations is the standard controlled benchmark in generative bridge and inverse-problem literature to cleanly isolate transport behavior. While evaluating arbitrary compound and unseen real-world degradations is beyond the scope of this initial study (and would prematurely conflate transport design with data-curation effects when single-operator restoration still has room to improve), our formulation is explicitly designed as a stepping stone toward that goal. By preserving the pretrained foundation model's generative prior and isolating task-specific transport in lightweight adapters, our approach enables modular, incremental Mixture-of-Experts routing and compositional restoration, avoiding the need to retrain a monolithic black-box model whenever new degradation types are introduced.

Thanks for additional experiments and response. However, the advantage of the proposed method over Blind Conditional LoRA is limited, so I keep my original score.

Reviewer_aT31OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a way to reuse a pretrained music generator for restoration. They freeze the generator and train a LoRA adapter with a flow-matching loss for one adapter per degradation type, and also one shared “blind” adapter.

Critical strengths
  1. The conditional LoRA control is a fair test of the main idea. Both models use the same backbone, the same rank-128 adapter, the same data, the same steps and the same solver. Only position of the degraded audio differs. The Bridge Adapter win is not just from putting LoRA on a pretrained model. This is the one result that tests the paper’s central claim.
  2. The method is cheap. Each adapter trains in a few hours.
Critical weaknesses
  1. The gain over sonicmaster is rather small (0.81 dB SI-SDR, 0.48 dB SNR).
  2. The claim that bridges “need not be learned from scratch” has no matching control. There is no run of the same ACE-Step architecture trained from random initialization, and none with full fine-tuning. So we do not know how much of the gain comes from the pretrained prior and how much from the data and the architecture.
  3. There is no listening test. Since the input beats all methods on SNR and SI-SDR, a listening test is the clearest way to show the restoration is better.
Questions for authors
  1. Was SonicMaster retrained on your single-degradation training pairs, or used as released?
  2. See weaknesses.
1 comment

We thank the reviewer for their comments and detailed feedback, as well as for their patience while we conducted these additional experiments. We hope our responses increase their confidence in the practical benefits of our approach and its promise as a foundation for future work.


  • Question: The gain over SonicMaster is rather small (0.81 dB SI-SDR, 0.48 dB SNR).

  • Answer: While the gains over SonicMaster on waveform fidelity metrics (SI-SDR and SNR) are modest, this aligns directly with our design goals. As noted in our response to other reviewers, degraded inputs often retain decent waveform fidelity while suffering from poor perceptual quality. Therefore, the real challenge is to substantially improve perceptual quality while preserving or slightly improving fidelity. Because our approach leverages the generative priors of a pretrained foundation model, it excels precisely where it matters most: across both objective perceptual metrics (11.8% lower FADFMA\text{FAD}_{\text{FMA}} than SonicMasterprompt\text{SonicMaster}_{\text{prompt}}, and 15.5% lower in the blind setting) and human preference evaluations (68.2% vs. 4.5% preference over SonicMaster in Table 2), our method substantially outperforms SonicMaster.

Furthermore, unlike SonicMaster, which requires training a model from scratch, our approach is a lightweight, post-training LoRA adaptation. As a result, it can seamlessly inherit improvements as stronger audio foundation models emerge in the future.

**Table 1: Human restoration preferences.

Method AMethod BPrefer ANo PreferencePrefer B
BridgeDegraded input65.1%22.6%12.4%
BridgeSonicMaster68.2%27.3%4.5%
BridgeConditional LoRA30.7%44.0%25.3%
Conditional LoRASonicMaster76.5%11.8%11.8%
Prespecified subset: input with audible degradation
BridgeDegraded input73.9%24.0%2.1%

  • Question: The claim that bridges “need not be learned from scratch” has no matching control. There is no run of the same ACE-Step architecture trained from random initialization, and none with full fine-tuning.

  • Answer: Thank you for raising this important point. Our statement that bridges “need not be learned from scratch” does not imply that training from scratch would yield inferior results; rather, it highlights a practical alternative—adapting an existing transport map that already shares the target endpoint. Indeed, if one had access to the full pretraining dataset of the foundation model (which is often unavailable), training a bridge from scratch could potentially achieve even stronger performance. However, that paradigm is orthogonal to our post-training modular framework, which is specifically designed to be lightweight, data-efficient, easy to extend with new restoration “experts,” and naturally scalable as stronger pretrained backbones emerge.

Conversely, we agree that full fine-tuning (FT) is a natural extension of our LoRA adapters. We therefore ran controlled experiments comparing full fine-tuning of the conditional baseline (Conditional FT) against full fine-tuning of our bridge formulation (Bridge FT), reported in Table 2. We train both variants for 50k training steps with a batchsize of 64. Both full-FT variants, as expected, improve over LoRA adapters while updating all model parameters, but the core conclusion from the paper holds: Bridge FT achieves noticeably higher signal fidelity (up to 1.79 dB SI-SDR and 3.78 dB SNR vs. 1.22 dB and 3.36 dB for Conditional FT) while maintaining comparable perceptual quality.

Table 2: Restoration results on all 19 degradations.

MethodMel ↓\downarrowSNR ↑\uparrowSI-SDR ↑\uparrowFAD ↓\downarrow
Conditional FT0.48793.361.220.2035
Bridge FT0.48423.781.790.2082

  • Question: There is no listening test.

  • Answer: To address this, we conducted a formal blinded human listening evaluation comparing Bridge Adapters against the unprocessed input and main baselines (reported in Table 1). We have also uploaded side-by-side audio examples across representative mild and severe degradations to our demo page: https://bridgeadapter.github.io/audio-examples/.


  • Question: Was SonicMaster retrained on your single-degradation training pairs, or used as released?

  • Answer: We used the official SonicMaster checkpoint as released, without any retraining or fine-tuning on our single-degradation pairs.

Reviewer_wj76OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Bridge Adapters, which adapt a frozen pre-trained music generation model into a restoration model mapping degraded audio to clean audio, via LoRA and flow matching. Experiments on 19 categories of synthetic degradations show that the method achieves better paired fidelity than a conditional-adapter baseline and prompted SonicMaster, but its performance is inconsistent across perceptual and reconstruction metrics.

Critical strengths
  • The comparison against a Conditional LoRA of equal capacity directly tests the role of the transport-path design: under five-step sampling, SI-SDR improves by 2.00 dB.

  • Compared with prompted SonicMaster, the method improves on most degradation categories, and the paper provides per-category results, uncertainty estimates, and an ablation over the number of sampling steps.

Critical weaknesses
  • The practical restoration benefit remains insufficiently validated. Even after excluding gain-only degradation categories, the unprocessed input still has a higher average SI-SDR than Bridge Adapters (2.32 dB vs 0.99 dB), and SonicMaster achieves a better Mel distance. These observations may partly stem from codec loss and waveform re-synthesis, but for that very reason a direct perceptual evaluation is all the more necessary. The authors’ informal listening is not sufficient evidence that the method improves perceived quality while preserving musical content.

  • The main perceptual evaluation is under-specified for reproduction, and its configurations are inconsistent. The embedding model, reference set, and preprocessing details used for FAD are not clearly specified. Moreover, the FAD of the main method is computed on float32 decoded outputs, while the conditional baseline uses bfloat16 decoded outputs, and the paper does not quantify the impact of this discrepancy. This weakens the credibility of the perceptual-quality comparison.

Questions for authors
  • Could the authors provide a small-scale blinded listening comparison on representative mild and severe degradations, comparing the method against the unprocessed input and the main baselines, assessing both perceived quality and preservation of musical content?

  • Could the authors provide the complete FAD computation configuration, and verify that the reported rankings and differences still hold after unifying the decoding precision?

1 comment

We thank the reviewer for their comments and valuable feedback.

  • Question 1: The practical restoration benefit remains insufficiently validated. Even after excluding gain-only degradation categories, the unprocessed input still has a higher average SI-SDR than Bridge Adapters (2.32 dB vs 0.99 dB), and SonicMaster achieves a better Mel distance. These observations may partly stem from codec loss and waveform re-synthesis, but for that very reason a direct perceptual evaluation is all the more necessary. The authors’ informal listening is not sufficient evidence that the method improves perceived quality while preserving musical content.

  • Answer: Thank you for raising this important point. As detailed in Table 1, codec reconstruction and waveform re-synthesis inherently penalize phase-sensitive, time-domain metrics like SI-SDR, even when perceptual metrics clearly favor the restored audio. We agree that this underscores the necessity of an apples-to-apples evaluation to isolate the true restoration gains of our method. Furthermore, we fully agree that informal listening is insufficient to validate perceptual quality and content preservation. To address this, we conducted a formal human listening study, the results of which are presented in Table 2.

**Table 1: Objective results across all 19 degradation **

ConditionSI-SDR (dB) ↑\uparrowSNR (dB) ↑\uparrowMel spectrogram distance ↓\downarrowFMA-Pop FAD ↓\downarrow
Degraded original6.624.090.56870.3392
Degraded decoded-2.351.760.66040.2723
Bridge restoration1.113.270.48300.2165

Table 2: Human restoration preferences.

Method AMethod BPrefer ANo PreferencePrefer B
BridgeDegraded input65.1%22.6%12.4%
BridgeSonicMaster68.2%27.3%4.5%
BridgeConditional LoRA30.7%44.0%25.3%
Conditional LoRASonicMaster76.5%11.8%11.8%
Prespecified subset: input with audible degradation
BridgeDegraded input73.9%24.0%2.1%

  • Question 2: The main perceptual evaluation is under-specified for reproduction, and its configurations are inconsistent.

  • Answer: Thank you for pointing out both the missing reproduction details and the inconsistency in decoding precision between the conditional baseline and Bridge Adapters. To ensure full reproducibility and a fair comparison, we have standardized decoding to FP32 across all methods and provided the complete evaluation setup in the revised text. As shown in Table 3, the updated results preserve all rankings from the original submission while widening the performance margin in favor of Bridge Adapters.

Table 3: FMA-Pop FAD (↓\downarrow) averaged across all degradations by decoder precision.

ConditionFMA-Pop FAD (↓\downarrow)Decoder precision
Degraded original0.3392No codec
Degraded decoded0.2723FP32
0.2707BF16
Bridge restoration0.2165FP32
0.2087BF16
Conditional LoRA0.2059FP32
0.1998BF16

  • Question 3: Could the authors provide a small-scale blinded listening comparison on representative mild and severe degradations, comparing the method against the unprocessed input and the main baselines, assessing both perceived quality and preservation of musical content?

  • Answer: We have added new audio samples across representative mild and severe degradations—comparing Bridge Adapters against the unprocessed input and main baselines—on our demo page (https://bridgeadapter.github.io/audio-examples/). In addition, we conducted the blinded human listening study to evaluate perceived audio quality and content preservation, as reported in Table 2.

Human Evalaluation: We conducted a blinded online restoration-preference study using ten held-out music excerpts and the 19 synthetic degradation families, forming a pool of 190 song–degradation pairs. Each excerpt lasted 15 seconds. Listeners could hear the clean reference and degraded input, then indicated whether the degradation was audible before hearing the anonymized candidates. Candidate identities and presentation order were randomized, the degradation type was not disclosed. For each pairwise comparison, listeners selected the version that best restored the original recording or indicated no clear preference. The results are based on 26 independent sessions totalling 547 pairwise judgments. Table 2 reports the pairwise results. The subset in which listeners reported an audible input degradation was specified in the study protocol.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.