acceptodds
Under review as a conference paper at ICLR 2027

One-Step Is Enough to Judge: Hybrid Noise Skimming for Full-band Text-to-Audio Generation

Abstract

Text-to-audio (TTA) systems based on latent diffusion and flow matching achieve high fidelity in the full-band scenarios, but their iterative sampling nature hinders practical deployment in real-world applications. Although few-step generation methods such as MeanFlow substantially reduce the number of function evaluations (NFEs), two fundamental challenges still remain in extremely low-NFE models. The first is unreliability: in extremely low NFE environments, quality of generated samples becomes highly sensitive to the initial noise. The second is quality inconsistency: increasing the number of ODE steps does not guarantee monotonic improvements in general sample quality. We propose a hybrid noise skimming framework to overcome both issues through better initial noise selection in few-step generation. The proposed framework combines two complementary components: a learnable Noise Booster amortizes the initial noise selection at the distribution level by refining random noise toward more favorable regions and Single-Step Search which improves the sample quality as instance-wise adaptation using only a single additional search step. The learnable Noise Booster restores the reliability and raises the quality floor by refining initial noise and in contrast to increasing ODE steps, enlarging the search budget yields monotonic improvement in sample quality. Experiments at 44.1 kHz show that our framework achieves SOTA performance with only a one-step generation. The project page is available at: https://github.com/briskmuntjack/muntjack-20262027

Primary area:
generative models
Keywords:
Full-band Text-to-Audio Generation, Flow Matching, Consistency Model, MeanFlow, x-prediction, Inference Time Scaling

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

This paper studies text-to-audio generation at 44.1 kHz with very few sampling steps. The authors point out that few-step models depend heavily on the initial noise and do not reliably improve with more steps, so they propose spending the extra compute on choosing a better initial noise. Building on an x-prediction MeanFlow generator that gives a usable sample in one step, a small Noise Booster first shifts Gaussian noise toward prompt-dependent high-reward regions, and Single-Step Search then previews N candidates in one step, ranks them with CLAP and a perceptual-quality (PQ) verifier, and finishes the best one with two steps.

The method is evaluated on AudioCaps with objective metrics and a listening test against diffusion, flow matching and few-step TTA baselines. The authors report the best text-audio alignment among few-step systems at 1 NFE, competitive with several 50-NFE models, steady gains as the search budget grows, and ablations of the components and design choices. Inference code and checkpoints are released, but not the training code.

Critical strengths
  1. The paper addresses an important problem: few-step text-to-audio models are sensitive to the initial noise and do not improve monotonically with more steps. Spending the budget on noise selection rather than extra steps is an intuitive way to restore monotonic scaling, and the proposed solution is clearly laid out.
  2. The method is validated with complete experiments, including objective metrics, a listening test, and ablations of each component and design choice.
  3. The authors release inference code and checkpoints.
Critical weaknesses
  1. The experimental scope is narrow. All objective results and the listening test are on the AudioCaps test set alone, whereas comparable few-step TTA works (e.g., AudioLCM [1], FlashAudio [2], MeanAudio [3]) report on two or more test sets. This may partly follow from adopting the TangoFlux training setup (AudioCaps and WavCaps). Still, validating on a single dataset limits how general the conclusions are, and I am concerned that the proposed method may to some extent be overfitting to this benchmark. Extending to broader audio generation settings such as V2A or VT2A would be welcome but is not necessary, as it falls outside the T2A scope of the paper.
  2. The sensitivity of few-step models to the initial noise is supported only by citations from the image domain, with no seed-variance analysis on MeanAudio, other few-step baselines, or the authors' own x-MF. The non-monotonic behavior with more steps is likewise shown only on x-MF. Since the argument for spending budget on noise selection rather than steps rests on these two observations, their lack of direct support weakens the motivation.

[1] AudioLCM: Text-to-Audio Generation with Latent Consistency Models. ACM MM 2024.

[2] FlashAudio: Rectified Flow for Fast and High-Fidelity Text-to-Audio Generation. ACL 2025.

[3] MeanAudio: Fast and Faithful Text-to-Audio Generation with Mean Flows. ACL 2026.

Questions for authors
  1. Can the authors report results on at least one additional test set beyond AudioCaps, so that the generality of the method and the possibility of overfitting to this benchmark can be assessed? If results on another test set such as Clotho, or further experiments, are provided, I would be willing to raise my score.
  2. Can the authors provide direct evidence for the two motivating premises: seed-to-seed variance of CLAP and PQ as a function of NFE for x-MF, a few-step baseline such as MeanAudio, and a multi-step baseline, and whether the non-monotonic behavior with more steps also appears in other few-step models?
  3. Figure 2 lacks the configuration x-MF + Single-Step Search without the Noise Booster. Can the authors add it? Without it, the marginal contribution of the Booster once search is active, which is the only learned component, cannot be judged.
3 comments

We thank the reviewer for your thoughtful and constructive feedback. We appreciate the valuable comments and questions regarding our work. We provide responses to your weakness critics and questions. We have uploaded a revised mauscript during the rebuttal period. Please refer to the updated version.

W1&Q1: Reports on generalization beyond the original benchmark

We evaluated all systems on the Clotho evaluation split (1,045 clips, 10 s), which is not used in our training (Table 2 on the paper, we added it).

Text relevance generalizes. At 1 NFE, our model has the best CLAP (0.398 vs. 0.331), KL (2.36 vs. 2.71) and AQA (0.892 vs. 0.857) among the accelerated models. None of these Clotho data used in training; the evaluation CLAP model differs from the reward CLAP model as described in section 4.3. Our 1-step model also outperforms the 50-step TangoFlux in FD, KL, CLAP and PQ.

Smaller performance drop than other models. From AudioCaps to Clotho, our CLAP score drops by 16%, whereas the score of MeanAudio-L and SoundCTM-DiT drop by about 40%. Our model shows lower performance drop rate than other conventional methods. Thus the gain of the Noise Booster is preserved and we find no evidence of overfitting to AudioCaps.

Limitations. The Noise Booster improves FD-PANN (16.42 to 15.52) but worsens the full-band FD-OpenL3 (79.93 to 113.93), where SAO small-ARC shows the best. We think it comes from SAO small-ARC is trained with Clotho dataset. In addition, we attribute this to its rewards (CLAP, PQ), which target text relevance and quality and do not constrain full-band content. A full-band reward is left for future work.

W2&Q2: Analysis on seed sensitivity and non-monotonic NFE scaling across models We measured both premises directly in audio.

Seed sensitivity (Appendix D and Fig. 8). Each model generates 64 clips from random initial noises for 500 AudioCaps test prompts. We report the within-prompt standard deviation (std, noise only) and the 10th percentile (low 10%). We used LAION-CLAP 630AF which is unused in search verifier or reward function. The multi-step TangoFlux (50 NFE) has a CLAP std of 0.047, below all six few-step models, including MeanAudio-L and our base x-MF (0.064–0.089). Its PQ std is 0.40, below five of them (0.45–0.56); the exception is ConsistencyTTA (0.37). With noise selection, our full system (16 NFE) has the lowest std value among few-step models (CLAP 0.057, PQ 0.28) and the highest low10% (CLAP 0.438 vs. 0.136–0.433; PQ 5.86 vs. 4.90–5.55) which means that it raises the lower bound and narrowed the fluctuations. Its PQ std is also below TangoFlux, while TangoFlux remains lower in CLAP score.

Revisiting Non-Monotonic Scaling (Figs. 2, 5, 6). Beyond 4–8 NFE, x-MF degrades gradually (Fig. 2). Other few-step models behave similarly (Fig. 6): AudioLCM degrades gradually in CLAP and FD, and SoundCTM-DiT drops at 16 NFE, while spending the same budget on Single-Step Search avoids this. v-prediction MeanFlow saturates (Fig. 5). We revised "non-monotonic" to "saturates or gradually degrades with more steps".

Q3: Ablation study with x-MF and Single-Step Search without the Noise Booster

We have added the x-MF + SSS configuration (without the Noise Booster) to Fig. 2. The new curve separates the two roles. SSS alone removes the degradation at large NFE budgets, so it is the main driver of the NFE scaling. Once search is active, the Booster's marginal contribution is smaller but consistent on fidelity and text alignment: the full model is better than SSS-only at every budget on FD (by 0.7-1.2) and CLAP (by 0.010-0.015). In efficiency terms, the full model at 4 NFEs already reaches the FD that SSS-only attains at 256 NFEs. On AQA the two are comparable, and on PQ SSS-only is marginally higher at large budgets. We also note that the Booster alone degrades faster at large budgets, and this does not occur once SSS is applied. We have revised the ablation paragraph to state these roles explicitly.

Thank you for the detailed response and the revised manuscript. The additional Clotho evaluation, the analysis of initial-noise sensitivity, and the SSS-only ablation address several of my original concerns.

I remain concerned about single-step generation quality on Clotho. At 1 NFE, the proposed method trails MeanAudio-L on FD-PANNs (15.52 vs. 13.53) and IS (8.89 vs. 9.96). I give limited weight to the gains on the four Audiobox Aesthetics scores as evidence of better overall generation quality. All four scores come from the same evaluator, and PQ is itself a training reward for the Noise Booster. The substantial gap between the FD-OpenL3 score for MeanAudio-L in Table 2 (551.34) and the score reported for MeanAudio-S in UNITE-AUDIO [1] (121.79) also raises concerns about possible differences in evaluation protocols and their implications for the fairness of the baseline comparison.

I would also like to understand why AudioCaps and WavCaps were chosen as the training datasets. Was the main motivation to match the TangoFlux training setup? The choice of training data could partly account for the gaps observed on some metrics relative to other baselines.

Finally, if the ICLR revision guidelines permit it, please highlight the main changes in the revised PDF in a different color so they are easier to locate.

[1] UNITE-AUDIO: Joint Learning of Continuous Tokenization and Latent Flow Matching for Text-to-Audio Generation. https://arxiv.org/pdf/2609.28206

Clotho result w/ single-step generation in MeanAudio We resampled 16kHz model's results to 44.1kHz by default algorithm of our evaluation framework. After testing alternative resampling algorithms, we found that MeanAudio's FD-OpenL3 score decreased to 124.1 at 25 NFE, while its score at 1 NFE was 131.9, matching the NFE setting of our method. We thank the reviewer for pointing out this overlooked issue and will update the paper accordingly.

The choice of training data We selected AudioCaps and WavCaps to align our training setup with TangoFlux as you mentioned and we also expect fairer comparison under similar training conditions. We acknowledge the differences in training data settings across baselines may partly contribute to the observed performans gaps.

Highlighting the main changes in the revised PDF in a different color We uploaded modification-addition marked paper in supplementary material. The reviesed parts are highlighted in yellow and added parts are highlighted in green.

  • We remove "SOTA" word from the abstrac/intro/conclusion and replace with numerical comparisons against SoundCTM-DiT
  • We measure ALL CLAP scores with an evaluation-only checkpoint (630AF), we add FD-OpenL3 and RTF columns to table 1 and replace 64-NFE row with 50 and 200 NFE rows.
  • We perform new measurement Clotho evaluation (Table 2) and add figure 6 as an single-step search.
  • We add initial-noise sensitivity analysis on Appendix D
  • We add SCORE curve to figure 3
Reviewer_pJFgOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a hybrid noise skimming framework for efficient full-band Text-to-Audio generation. Built upon an x-prediction MeanFlow backbone, the framework aims to mitigate initial noise sensitivity and enhance generation quality, even under conditions with a low NFE. Experimental results demonstrate the framework's capability to generate 44.1 kHz audio at 1 NFE, with objective evaluations confirming that performance further improves as the search budget is increased.

Critical strengths
  • s1. This paper presents a clear and practical motivation for evaluating initial noise candidates using 1-step generation to reduce the search cost compared to full-step search.

  • s2. The model achieves competitive performance with 1 NFE in 44.1 kHz stereo TTA and shows better results than existing few-step TTA models.

  • s3. The samples provided on the demo page show results comparable to, and sometimes better than, TangoFlux and SAO small-ARC in terms of alignment with the text.

Critical weaknesses
  • w1. The abstract of this paper identifies the instability of generation quality caused by initial noise as a key problem and claims that Noise Booster restores reliability and improves the lower bound of generation quality. However, the experiments mainly focus on improvements in average performance, without analyzing the variance in generation quality across different initial noise samples or the frequency of low-quality outputs. So, it remains unclear whether the proposed method actually reduces sensitivity to initial noise.

  • w2. This paper address the Text-to-Audio generation task, but its main contribution lies in the optimization and selection of initial noise. However, Noise Booster builds on the residual noise transformation of Noise Hypernetworks, while Single-Step Search reduces the number of generation steps required for candidate evaluation in existing reward-based Best-of-N selection. The novel contribution beyond combining these existing approaches is unclear. In particular, the lack of comparisons with directly related prior works, such as Noise Hypernetworks, LENS, and SCORE, using the same TTA backbone makes it difficult to assess the effectiveness of the proposed noise optimization method itself. Furthermore, Fig 2 does not include an x-MF + SSS experiment, making it difficult to isolate the independent effect of SSS.

  • w3. The claims of monotonic quality improvement and inference efficiency are not sufficiently validated. Section 4.4 claims monotonic performance improvements across all evaluation metrics, but Table 1 shows intervals where FD, IS, KL, and AQA become worse, without providing any analysis. In addition, Fig 4 compares search efficiency based only on generator NFE, excluding the computational costs of VAE decoding and CLAP inference for each candidate. Since SSS and full-step search evaluate different numbers of candidates under the same NFE budget, it is difficult to determine whether the proposed method is also more efficient in terms of actual inference time or total computational cost.

Questions for authors
  • q1. In Table 1, TangoFlux and Resonate, which are diffusion/flow-based TTA models, use 50 NFEs. What are the evaluation results of the proposed model when using NFE=50? Comparing NFE=50 and 200 in the same table would be more appropriate than reporting NFE=32 and 64.

  • q2. Who participated in the MOS evaluation, and how was it conducted (e.g., through crowdsourcing or internal evaluation)?

1 comment

We thank the reviewer for the constructive feedback. We have revised the manuscript and conducted additional experiments to address the concerns.

Initial-noise sensitivity (W1). We conducted a multi-seed evaluation on 500 AudioCaps prompts with 64 trials per prompt (Appendix D, Fig. 8). At 1 NFE, the Noise Booster improves mean CLAP-630AF (0.452 → 0.474) and PQ (5.66 → 5.83), while reducing within-prompt standard deviation (CLAP: 0.079 → 0.070; PQ: 0.56 → 0.40). The lower 10th percentiles increase by 9.9% for CLAP and 8.3% for PQ, while low-quality rates decrease from 10.0% to 6.4% (CLAP) and 4.7% (PQ). All improvements are significant (paired bootstrap, p < 0.001). At 16 NFE, combining the Booster with SSS further reduces variability and improves lower-tail CLAP. These results demonstrate reduced initial-noise sensitivity beyond mean quality improvements.

Novelty and prior work (W2). We expanded Section 2.3 to clarify the relationship to DNO, ReNO, Noise Hypernetworks, LENS, and SCORE. Our framework combines two complementary mechanisms for few-step full-band TTA: the Noise Booster performs amortized distribution-level refinement, while SSS performs instance-wise selection using one-step previews.

We added the missing x-MF + SSS without Booster ablation (Fig. 2). SSS mitigates degradation at increasing NFE budgets even without the Booster, while their combination further improves FD and held-out CLAP. The multi-seed analysis (Fig. 8) demonstrates complementary benefits in lower-tail quality and variability, although improvements are metric-dependent. We also compare against SCORE-based full-step Best-of-N (Fig. 3) and evaluate SSS on AudioLCM and SoundCTM-DiT (Fig. 6), demonstrating applicability beyond x-MF.

Monotonicity and inference efficiency (W3). We agree that the original monotonicity claim was overstated and have revised Section 4.3 accordingly. CLAP-630AF and PQ improve consistently with increasing search budgets, whereas other metrics may fluctuate or saturate. We additionally report FD-OpenL3 for broader distributional evaluation.

To assess practical inference efficiency, Table 1 and Appendix B now report end-to-end RTF, including generator evaluations, VAE decoding, and CLAP/PQ verification. We further measured the search configurations in Fig. 4:

Table R1. End-to-end RTF (lower is better).

NFENo SearchSSS (1-step)Full-step (8-step)
80.0260—0.0248
160.04410.09360.0540
320.08050.18610.0776
640.15330.38090.1279

SSS: NFE = N + 8; Full-step Search: NFE = 8N.

The key advantage of SSS is broader initial-noise exploration under a limited generator budget. At 64 NFE, SSS evaluates 56 candidates, compared with only 8 for full-step search. As shown in Fig. 4, this broader exploration yields favorable quality improvements, particularly at low NFE budgets. Moreover, at matched candidate count (N = 8), SSS reduces generator evaluations from 64 to 16 and end-to-end RTF from 0.1279 to 0.0936 (26.8% reduction), although final generation quality may differ.

At matched NFE, however, SSS incurs higher wall-clock costs. To investigate this overhead, we additionally profiled both approaches at 16 NFE using the same x-MF backbone without the Booster, under identical server conditions. Generator execution times were nearly identical (0.246s vs. 0.250s), whereas VAE decoding increased from 0.082s to 0.583s, accounting for 79.7% of the additional runtime. VAE decoding also constituted 57.1% of SSS's total latency, compared with 18.3% for CLAP/PQ verification.

These results indicate that SSS's wall-clock overhead at matched NFE is primarily associated with repeated audio decoding rather than generator computation. While this cost must be included in practical efficiency comparisons, its magnitude depends on the audio codec and verification pipeline. We therefore distinguish the algorithmic benefit of broader noise exploration per generator NFE from implementation-dependent end-to-end latency, reporting both measures to characterize this trade-off transparently.

50-NFE comparison (Q1). We added 50- and 200-NFE results to Table 1 for comparison with TangoFlux, Resonate, and GenAU-L. At 50 NFE, our method achieves FD-PANNs 12.28, CLAP-630AF 0.508, and PQ 6.33, providing competitive generation quality with higher perceptual quality.

MOS evaluation protocol (Q2). The listening test was conducted internally rather than through crowdsourcing. As detailed in Appendix C, 24 participants rated six anonymized systems using five-point OVL and REL scales. Each participant evaluated 15 captions, with system order balanced by a Williams Latin square. This yielded 360 ratings per system and metric, with each caption assessed by 2–5 listeners.

Reviewer_XsJbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets one- and few-step 44.1 kHz text-to-audio. It trains an x-prediction MeanFlow generator (x-MF) and adds a small Noise Booster that nudges the initial noise (trained with CLAP and PQ rewards) plus Single-Step Search, which previews N boosted noises in one step, picks the best by CLAP/PQ rank, and regenerates it in two steps. Evaluation is on AudioCaps.

Critical strengths
  1. It's a practical system. One-step 44.1 kHz stereo generation works, the 16-NFE MOS (3.00) is on par with the best baselines, and the released code matches the paper.
  2. The search seems to pick genuinely better noise, not just better verifier scores. FD, IS, KL, CE, CU and AQA also improve overall from 1 to 64 NFE, and MOS goes up from 1 to 16 NFE.
Critical weaknesses

Points 1 and 2 drive my score.

  1. The headline numbers are on the metrics being optimized. The Booster rewards and search verifiers are the same CLAP and Audiobox PQ models used in Table 1, and that's where the "SOTA in one step" claim and the +23.3% CLAP / +5.8% PQ come from. On metrics nobody optimized, the 1-NFE model trails MeanAudio-L (FD 14.35 vs 11.02, AQA 0.865 vs 0.882, also IS and KL). In the listening test, relevance is still below TangoFlux at 16 NFE (3.42 vs 3.73, CIs don't overlap), even though TangoFlux has much lower CLAP.
  2. The "inconsistency" problem isn't really established, and the fix sidesteps it. Getting worse with more steps mostly shows up for the authors' own x-MF (Fig. 2). The v-prediction MeanFlow in Fig. 5 improves FD monotonically, and no other few-step TTA model is tested. Fixing K = 2 and spending the budget on search avoids the problem rather than solving it, and monotonic verifier gains are what best-of-N gives you anyway (Table 1 isn't monotonic on FD, IS, KL or AQA, despite p.8).
  3. It's close to existing work, and the obvious comparisons are missing. The Booster is amortized noise refinement in the spirit of Noise Hypernetworks and LENS (and Golden Noise/NPNet, ICCV 2025). The selection uses the same standardized CLAP + PQ rewards as SCORE, which is cited but not compared. I'd want to see x-MF + search without the Booster, and the search (or SCORE-style best-of-N) on another one-step model at the same cost.
  4. Speed is never actually measured, only NFE. Table 1 has no parameter counts and no latency or RTF, and NFE isn't comparable across models of very different size and architecture. For this method NFE also leaves out the VAE decode and two verifier passes per candidate, so 16 NFE really means 15 decodes and 28 verifier calls. For an acceleration paper, real inference time is what matters, and it only appears in the repo README.
  5. The evaluation is narrow. Everything is tested on AudioCaps, whose training split is in the training data, and the Booster is trained on those same caption distributions, so it's unclear how the method holds up on out-of-domain prompts. "Full-band" is also barely measured (most metrics run at 16 kHz, the listening test is mono).
Questions for authors
  1. Can you report text-audio alignment with a model you didn't train or select with?
  2. Can you add parameter counts and measured latency (RTF, peak memory) for every system in Table 1 on the same GPU, including decode and verifier cost for yours?
  3. How does x-MF + search do without the Booster, and how does the search do on another one-step model at matched cost?
  4. Do other few-step TTA models (e.g., MeanAudio-L) also get worse with more steps, and how does the method do on out-of-domain prompts?
3 comments

We thank the reviewer for the constructive feedback. We have expanded the evaluation and revised our claims to address the concerns below.

W1 / Q1. Independent evaluation

We agree that alignment should not be assessed only with the CLAP and PQ models used for training and selection. We now report CLAP (630AF), evaluated using a separate LAION-CLAP checkpoint (630k-audioset-fusion-best.pt) not used for optimization or candidate ranking, alongside AQA, fidelity metrics, and listening results (Tables 1 and 3; Appendix B).

Our initial-noise sensitivity study across 500 prompts and 64 trials per prompt further shows that the Booster improves held-out CLAP lower-tail scores and reduces within-prompt variability (Figure 8, Appendix D).

Nevertheless, MeanAudio-L remains stronger on FD, AQA, and held-out CLAP at 1 NFE, while TangoFlux retains higher human-rated relevance than our 16-NFE setting (Tables 1 and 3). We have therefore removed the broad 1-NFE SOTA alignment claim.

W2 / Q4. Step-dependent degradation and search scaling

We agree that degradation with additional sampling steps is not universal. We observe it for x-MF, AudioLCM, and SoundCTM-DiT under the evaluated settings (Figures 2 and 6), while the v-prediction MeanFlow variant shows a different step-dependent trend (Figure 5).

Applying SSS without the Booster to AudioLCM and SoundCTM-DiT at matched generator-NFE budgets within each model avoids the deterioration associated with allocating compute solely to extra steps, with larger gains for SoundCTM-DiT. AudioLCM benefits less because its sampler injects fresh noise. Thus, SSS is not restricted to x-MF, although its effectiveness depends on the sampler.

Our method does not restore multi-step consistency; it reallocates computation to initial-noise selection. Because best-of-N naturally favors higher verifier scores, we distinguish those gains from held-out metrics and clarify that improvements are not monotonic across all measures (Table 1).

W3 / Q3. Component ablation and novelty

We added x-MF + SSS without the Noise Booster, completing the four-way component ablation (Figure 2). SSS provides the main search-scaling benefit, while the Booster further improves FD-PANNs and held-out CLAP even with SSS. The benefits are metric-dependent: SSS without the Booster can obtain higher PQ at larger budgets. Figure 8 further shows reduced initial-noise sensitivity.

We acknowledge precedents for learned noise refinement and reward-based best-of-N selection. Our distinctive design combines an effective one-step x-prediction MeanFlow generator with one-step previews followed by few-step regeneration only from the selected noise, reducing generator evaluations from NK to N + K (Section 3.2, Equation 2).

We additionally compare against SCORE-reward-based full-step best-of-N selection (Figure 3), illustrating the quality-efficiency trade-off of one-step screening.

W4 / Q2. Measured inference cost

We agree that generator NFE alone does not characterize practical inference cost. We added measured end-to-end RTFs for the evaluated systems (Table 1), including the Noise Booster, candidate previews, VAE decoding, verifier scoring, and final generation.

The A100 GPU, precision, and timing protocol are specified in Appendix B; our generator and Booster parameter counts are reported in Section 4.2. The measured results show favorable runtime at moderate budgets but substantially increasing cost with larger searches. We therefore qualify efficiency claims using actual runtime rather than NFE alone.

W5 / Q4. Generalization and full-band evaluation

We added evaluation on the unseen Clotho test split (1,045 audio-text pairs; Table 2). The method retains strong held-out CLAP and AQA performance compared with TangoFlux, providing evidence of transfer beyond AudioCaps, although MeanAudio-L remains better on FD-PANNs and some metrics saturate with search.

We also added FD-OpenL3 computed from 44.1 kHz signals (Tables 1 and 2). It reveals that full-band distributional fidelity does not consistently improve through noise selection and remains behind some baselines. We now state this limitation explicitly rather than claiming uniform improvements.

We thank the reviewer for helping us strengthen the evidence and clarify the scope of our contributions.

Thanks for the additional experiments. The search-only ablation, Clotho evaluation, and runtime measurements address several of my concerns. I still see the main contribution as a refinement of existing noise search methods, and the one-step results on the held-out metrics are mixed. I’ll keep my original rating.

Thank you for your thoughtful follow-up and for recognizing the value of our additional experiments. We sincerely appreciate your constructive feedback, which has been invaluable in strengthening our paper and clarifying its contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.