- 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
- 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.
- The method is validated with complete experiments, including objective metrics, a listening test, and ablations of each component and design choice.
- The authors release inference code and checkpoints.
Critical weaknesses
- 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.
- 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
- 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.
- 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?
- 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
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
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.