{"id":"ae17d089-a31b-485e-b93c-4adcc6a45bc7","arxiv_id":"2412.08577","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Mel-Refine boosts text-to-audio spectrogram sharpness by amplifying skip-connection high frequencies and attenuating backbone high frequencies at inference, but its reported 25% improvement is based on test-set-tuned parameters.","lead":"The paper proposes Mel-Refine, a training-free inference tweak that adjusts high- and low-frequency U-Net features to make text-to-audio models produce sharper Mel-spectrograms. The authors report large quality gains, but the gains come from parameters chosen on the same test set used to measure them.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Tuning the five per-model parameters on the same evaluation subset and then reporting metrics on that subset makes the 25% Tango2 gain an in-sample fit; the paper provides no held-out validation of the claimed improvement.","rationale":"The reader's weakest_assumption correctly identifies the most load-bearing flaw: the five parameters are tuned on the evaluation subset, and the headline metrics come from that same subset. I agree with the REJECT recommendation. The method itself is plausible and the component analysis in Table I plus the ablations in Table III provide some internal support, so the appropriate remedy is a held-out parameter-selection protocol rather than dismissal of the direction of the effect. I find no more fundamental internal inconsistency that would make the method impossible; the FFT-shift convention in Equation (4) is underspecified and worth checking in the released code, but the evaluation flaw alone is enough to invalidate the stated 25% claim. Verdict remains REJECT, so no adjustment to the reader's verdict is needed.","tokens_in":7967,"tokens_out":7052,"duration_ms":71569,"concrete_test":"Partition the Tango2 Audiocaps evaluation subset into two disjoint halves. Run the Section III-C grid search on half A only, freeze the best parameters, and report FD/FAD/KL/OVL on half B; repeat with the halves swapped. If the held-out Tango2 FD does not reproduce the 37.48 to 28.13 improvement (e.g., relative improvement below half the claimed 25%) or FAD/KL/OVL are not consistently better across both splits, the headline effect is an in-sample fit rather than a validated improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is not supported by the experimental protocol. In Section III-C, the five parameters (s1, s2, b1, b2, m) are selected by a two-stage grid search, and Section III-A states that evaluation uses only a subset of Audiocaps and MusicBench. The metrics reported in Table II (e.g., Tango2 FD 37.48 to 28.13, FAD 2.48 to 1.69) are computed on the same subset used for parameter selection. With five free parameters per model, a grid search can exploit noise in a small evaluation subset, so the 25% improvement is a fitted quantity rather than a predictive claim. No confidence intervals, significance tests, or standard errors are reported, and the subjective OVL result (36 to 64) is based on 10 evaluators and 30 clips without protocol detail. The ablation in Table III also uses the same tuned configuration, so it does not independently validate the components. This does not show the method cannot work; Table I and Figure 4 suggest a plausible directional effect, but the stated magnitude is unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Mel-Refine, a training-free inference-time method to improve the quality of Mel-spectrograms generated by latent-diffusion text-to-audio (TTA) models. The method applies frequency-domain scaling to the high-frequency components of U-Net skip connections and backbone features in the first two decoder blocks, controlled by five scalar parameters (s1, s2, b1, b2, m). The authors report experiments on Tango, Tango2, and MusTango, showing improvements in FD, FAD, KL, and subjective OVL. The central claim is that high-frequency components in skip connections enhance texture, while low-frequency backbone components are critical for denoising, and that adjusting these weights during inference improves audio fidelity.","tokens_in":8225,"tokens_out":8651,"duration_ms":80681,"significance":"The idea of reweighting U-Net frequency components at inference is timely and draws on a known phenomenon from image diffusion models (FreeU). If the reported improvements were robust, the method would be a useful, training-free plug-in for LDM-based TTA models. The paper's strengths are the concrete mechanism and the qualitative spectrogram comparisons; the weaknesses are the in-sample parameter selection and lack of statistical rigor, which affect the credibility of all quantitative claims.","major_comments":[{"comment":"The parameter-selection and evaluation protocol is circular. In Section III-C, the five parameters (s1, s2, b1, b2, m) are determined by a two-stage grid search, and the resulting settings are used for the results in Table II. Section III-A states that evaluation is performed on 'only a subset of the data' rather than the full test set. Because the same subset is used both for tuning and for reporting, the improvements, including the headline Tango2 FD reduction from 37.48 to 28.13, are in-sample fits rather than predictive results. The authors must use a separate validation set for parameter search and a disjoint test set for final evaluation, or perform nested cross-validation, and they should report the subset size and the number of parameter configurations tried.","section":"III-A, III-C"},{"comment":"The paper reports no uncertainty quantification. All objective metrics are single numbers, so the small differences for Tango (FD 28.36 vs 27.78) and MusTango (FAD 1.67 vs 1.61) cannot be distinguished from noise. Similarly, the subjective OVL evaluations are based on only 10 evaluators and 30 clips, with no description of whether listening tests were blinded, randomized, or analyzed for significance. Confidence intervals, standard errors, or significance tests are needed before the claimed improvements can be accepted.","section":"III-D, Table II"},{"comment":"The frequency mask in Eq. (4) defines the 'central region' as low-frequency, with x in [1/4W, 3/4W] and y in [1/4H, 3/4H]. In a standard 2D FFT without fftshift, the low-frequency (DC) components are located at the corners, not the center. If the authors are using a library that does not shift the spectrum, this mask would amplify the wrong region. The manuscript must specify the exact FFT implementation and whether fftshift/ifftshift is used; otherwise, the mechanism described (enhancing high frequencies in skip connections, attenuating them in the backbone) may be operating on the wrong frequency bands.","section":"II-B, Eq. (4)"},{"comment":"The ablation study does not provide independent evidence for the contribution of each module. It removes one module at a time while keeping the parameters tuned for the full method, on the same evaluation subset used for tuning. A proper ablation should re-tune or at least evaluate on a held-out set, and it should report whether parameter re-optimization changes the pattern. The current table only shows that the method's components are interdependent under a particular tuning, not that each component is necessary.","section":"III-E, Table III"}],"minor_comments":[{"comment":"The size of the evaluation subset is not stated; knowing how many clips are used is essential for interpreting the stability of FD/FAD/KL estimates.","section":"III-A"},{"comment":"The FAD value of 9.58 for 'Backbone Low-frequency Attenuate' is far outside the range of the other entries; please verify this number and add a sentence of explanation.","section":"II-A, Table I"},{"comment":"The min and max in the normalization of α are not defined; it should be stated whether they are taken over the spatial dimensions of each channel or over the channel dimension.","section":"II-B, Eq. (6)"},{"comment":"The claim of full compatibility with 'any diffusion-based TTA architecture' is not supported by the experiments, which cover only LDM-based models; a more cautious statement would be appropriate.","section":"I"},{"comment":"The '25% boost' is not defined; if it refers to the relative decrease in FD, this should be stated explicitly, along with the fact that it is a single-metric, single-model result.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an interesting question, but the experimental evidence is currently insufficient. The main concern is the in-sample parameter tuning; if the authors can run a proper held-out evaluation and provide uncertainty estimates, the paper would be considerably stronger. I also recommend that the authors clarify the FFT shifting in Eq. (4). There is no indication of fabrication or prior publication concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe core idea is worth a look: the paper transfers FreeU's frequency-component manipulation to Mel-spectrogram TTA, and their analysis of how high-frequency components in skip connections and the backbone affect texture versus denoising is genuinely new for audio. The method itself is attractively simple—scale a few frequency bands at inference, no retraining—and the qualitative figures show visible changes in spectrogram detail.\n\nBut the central quantitative claim does not survive contact with the experimental protocol. Section III-C says the five parameters (s1, s2, b1, b2, m) were selected by a two-stage grid search, and Section III-A says evaluation used only a subset of Audiocaps and MusicBench. The same subset's metrics are then reported in Table II as improvements. That is in-sample fitting. With five free parameters per model, a grid search on a small evaluation set can easily chase noise, so the Tango2 FD drop from 37.48 to 28.13 (the '25% boost') is a post-hoc selected number, not a predictive result. No confidence intervals, error bars, or significance tests appear anywhere.\n\nThis is a load-bearing flaw, not a stylistic nit. The paper itself gives little reason to trust the magnitude: the ablation in Table III uses the same tuned configuration, so it does not independently validate the components. The subjective OVL comparison (36 to 64 on Tango2) rests on ten evaluators and thirty clips, with no protocol detail.\n\nNone of this proves the method is useless. The direction of the effect in Table I and Figure 4 is plausible, and the approach is cheap and model-agnostic. It could turn out to be a useful inference-time trick. But as presented, the evidence is not yet there.\n\nMy bottom line: this deserves a proper peer review, not a desk reject, because the idea is testable and the fix is straightforward—add a held-out set, report error bars, and show the parameter selection does not overfit. But I would not cite it in its current form, and I would not take the 25% number at face value.\n\nSend it out to a referee, with a request to focus on validation methodology.","headline":"A plausible no-training U-Net frequency hack for TTA, but the 25% headline is a fitted number because the knobs were tuned on the same evaluation subset used to report gains.","tokens_in":8794,"tokens_out":2499,"would_cite":false,"duration_ms":24486,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By reweighting high- and low-frequency components of the U-Net feature maps during inference, Mel-Refine improves text-to-audio quality without retraining, cutting Tango2's FD from 37.48 to 28.13 and raising listener preference from 36 to…","keywords":["text-to-audio generation","Mel-spectrogram","diffusion model","U-Net","frequency-domain refinement","plug-and-play inference","Tango2","audio quality metrics"],"falsifier":"Run Mel-Refine with the paper's published parameter settings on a held-out set of AudioCaps captions never used in the grid search, keeping all other inference settings fixed; if FD, FAD, and KL show little or no improvement over baseline, or listener OVL does not rise, the reported gains are tuning artifacts rather than a general inference-time effect.","tokens_in":7762,"feed_emoji":"🎧","tokens_out":6017,"duration_ms":59107,"temperature":0.7,"pith_summary":"Text-to-audio models that generate Mel-spectrograms with a U-Net-based diffusion backbone tend to blur or muffle complex sound scenes because they fail to render fine texture. Mel-Refine is a training-free, inference-only intervention that rebalances how much the U-Net's skip connections and backbone contribute at high and low frequencies: it amplifies high-frequency components in skip connections, applies a structure-aware amplification to the backbone, and suppresses high-frequency components in the backbone. On Tango, Tango2, and MusTango, these simple weight changes improve objective audio-quality metrics and subjective overall-quality scores. The point is that a model's quality ceiling is not fixed after training; the frequency structure of its own feature maps during inference contains recoverable detail.","feed_headline":"Tweaking U-Net frequencies at inference lifts Tango2 quality 25%","feed_subtitle":"No fine-tuning and no new data: the method reweights how U-Net frequency components act during diffusion sampling.","key_machinery":"The central object is the U-Net decoder feature map split by Fourier transform into high- and low-frequency regions. A scaling matrix $\\beta$ leaves the central low-frequency band unchanged while scaling the rest by $s$, applied to skip-connection features; the backbone gets a structure-aware scaling based on channel-wise means (controlled by $m$) followed by a second Fourier-domain mask that attenuates high frequencies with $b$. These operations are applied only in the first two decoder blocks, where features are densest. This machinery turns the empirical finding into a concrete inference-time recipe: reweight, rather than retrain, to recover high-frequency texture.","core_discovery":"This paper establishes that the frequency content of U-Net feature maps has separable roles in Mel-spectrogram generation: the high-frequency parts of both skip connections and backbone determine texture and edge detail, while the low-frequency backbone content drives the denoising process. Based on Fourier analysis of intermediate features, Mel-Refine adjusts these components during inference, multiplying skip-connection high frequencies by $s>1$, scaling backbone features with a structure-aware map controlled by $m>1$, and attenuating backbone high frequencies by $b<1$, all within the first two decoder blocks. The authors show that this restores detail and coherence in generated audio, with the largest gains when the base model has lost generation quality through preference alignment, and report consistent improvements in FD, FAD, KL, and listener OVL across three models. The intended contribution is a plug-and-play refinement stage that requires no additional training or fine-tuning and is compatible with any diffusion-based TTA architecture.","pith_inferences":["The paper does not test whether the same frequency-reweighting recipe transfers to other U-Net-based latent diffusion domains, such as image or video generation, but the mechanism is stated in terms of generic feature maps rather than audio-specific structure.","A stricter evaluation would fix the five parameters on a training subset and measure on a separate held-out caption split, since the paper's grid search and evaluation share the same subset.","The method could plausibly be combined with other inference-time interventions or applied at more than the first two decoder blocks; the paper does not explore whether later decoder stages would help or hurt.","If the proposed explanation is correct, the benefit should scale with prompt complexity: prompts with many sound events should show larger gains than simple single-event prompts, which is a directly testable prediction."],"forward_implications":["Any diffusion-based text-to-audio model with a U-Net can adopt Mel-Refine at inference time, with no retraining or fine-tuning, and improve FD, FAD, and subjective overall quality.","The frequency-decomposition result implies that amplifying low-frequency components in the U-Net backbone, rather than high-frequency ones, is what causes over-smoothing or degraded denoising in Mel-spectrogram generation.","The largest gains appear on models whose alignment fine-tuning traded away generation quality, so Mel-Refine offers a way to recover quality after alignment without retraining.","Because only the first two decoder blocks are modified, the added inference cost is small enough to make the method practical as a default post-processing step."],"supporting_citations":[{"why":"Supplies the structure-aware backbone scaling and the finding that enhancing the U-Net backbone improves denoising, which Mel-Refine adapts.","marker":"[22]"},{"why":"Defines Tango2, the primary model on which the largest FD, FAD, KL, and OVL gains are reported.","marker":"[21]"},{"why":"Defines Tango, one of the additional models used to demonstrate generalization.","marker":"[19]"},{"why":"Defines MusTango and the MusicBench test set used to evaluate the method on music generation.","marker":"[20]"},{"why":"Provides the AudioCaps test set used for the Tango and Tango2 evaluations.","marker":"[24]"},{"why":"Defines the FAD metric used to measure distribution-level audio quality.","marker":"[26]"},{"why":"Defines the FD metric, which the paper uses alongside FAD and KL.","marker":"[14]"},{"why":"Describes DPO fine-tuning, which the paper identifies as the cause of Tango2's reduced generation quality that Mel-Refine restores.","marker":"[27]"}],"fun_headline_variants":["No-training tweak sharpens Mel-spectrograms, lifts Tango2 25%","Inference-only frequency reweighting refines TTA audio","Rebalance U-Net frequencies for richer audio detail, no retraining","Plug-and-play frequency fix improves TTA models without data","Tweak U-Net weights at inference to boost Tango2 quality 25%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gains rest on the assumption that the five per-model parameters chosen by grid search on the evaluation subset generalize to audio beyond that subset, rather than merely fitting the test data.","fun_headline_variants_meta":{"raw":{"variants":["No-training tweak sharpens Mel-spectrograms, lifts Tango2 25%","Inference-only frequency reweighting refines TTA audio","Rebalance U-Net frequencies for richer audio detail, no retraining","Plug-and-play frequency fix improves TTA models without data","Tweak U-Net weights at inference to boost Tango2 quality 25%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001407,"raw_usage":{"total_tokens":5681,"prompt_tokens":937,"completion_tokens":4744,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":4646}},"tokens_in":553,"tokens_out":4744,"duration_ms":32235,"temperature":1.0,"reasoning_tokens":4646,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:43:33.619566+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Mel-Refine with the paper's published parameter settings on a held-out set of AudioCaps captions never used in the grid search, keeping all other inference settings fixed; if FD, FAD, and KL show little or no improvement over baseline, or listener OVL does not rise, the reported gains are tuning artifacts rather than a general inference-time effect.","supporting_citations":[{"cited_title":"Freeu: Free lunch in diffusion u-net,","cited_arxiv_id":null,"evidence_quote":"Supplies the structure-aware backbone scaling and the finding that enhancing the U-Net backbone improves denoising, which Mel-Refine adapts."},{"cited_title":"Tango 2: Aligning diffusion-based text-to-audio generative models through direct preference optimization,","cited_arxiv_id":null,"evidence_quote":"Defines Tango2, the primary model on which the largest FD, FAD, KL, and OVL gains are reported."},{"cited_title":"Audiocaps: Generating captions for audios in the wild,","cited_arxiv_id":null,"evidence_quote":"Provides the AudioCaps test set used for the Tango and Tango2 evaluations."}],"review_version":1}