{"id":"f5879f38-e7d7-4466-b3fc-c786ce52ad85","arxiv_id":"2506.03020","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A training-free inference strategy, FIFO sampling plus attention-guided curved denoising, enables diffusion text-to-audio models to produce long, temporally consistent audio with constant memory.","lead":"InfiniteAudio is an inference-only technique that lets pre-trained text-to-audio diffusion models generate arbitrarily long audio with a fixed memory footprint, by combining FIFO sampling from video diffusion with a new attention-based denoising step schedule. It reports comparable quality to standard generation on short clips while extending output length without additional training.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central 'infinite-length, consistent' claim is not stress-tested: quantitative evidence stops at 10-20 s, and the ~3 s sliding window cannot carry dependencies longer than itself, so unbounded consistency remains an extrapolation.","rationale":"I agree with the reader that the evaluation is not trustworthy enough for the quantitative 'comparable or superior' claim; the pruned test set, absent error bars, and test-set-strategy selection all warrant CONDITIONAL. But the single most load-bearing element of the central claim is not the exact CLAP numbers; it is the word 'infinite.' The mechanism's own design (fixed window, popped frames, about 3 s context) places an upper bound on the temporal dependencies it can express, and the experiments never reach durations where that bound would bind. The buffer-zone transfer to audio (the reader's weakest assumption) is a real but local out-of-distribution concern; even if the buffer completely fixes per-frame noise levels, long-horizon stability must be demonstrated separately. I therefore mark agreement as partial. I keep the verdict UNCHANGED because the reader's CONDITIONAL already requires stronger evaluation; my concern sharpens what that evaluation should include (a long-horizon drift test) rather than changing the verdict. If the 300 s stress test shows drift, the verdict should move to REJECT, or at least the claim should be weakened to 'constant-memory streaming with locally consistent transitions.' The constant-memory result itself is plausible and independently demonstrated, so no objection is raised there.","tokens_in":8338,"tokens_out":9536,"duration_ms":119282,"concrete_test":"Run a controlled long-horizon stress test with the released or re-implemented InfiniteAudio on a fixed 50-prompt validation set (not the pruned test set): generate 10, 30, 60, 120, and 300 s per prompt. Compute CLAP and FD (or KL) on consecutive non-overlapping 10 s segments and fit the per-segment scores versus time; if the slope is significantly negative or the worst-segment score degrades with length, the 'infinite-length consistency' claim is falsified. Also ablate buffer-zone size and window length (e.g., B=0 versus B>0; 3 s versus 10 s window). If per-segment scores stay flat to 300 s and are insensitive to B/window, the extrapolation is supported; if not, the paper must either weaken the claim to 'locally consistent streaming' or add a conditioning mechanism that maintains long-range context.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The pivotal claim is that FIFO sampling plus curved denoising produces arbitrarily long, temporally consistent audio from an unmodified pre-trained TTA model. Two conditions must hold: (1) the per-frame noise levels seen by the denoiser stay on-manifold at every queue position (the buffer-zone assumption of Section 2.2.1), and (2) the process is stable when frames are popped and replaced, i.e., errors do not accumulate and no long-range structure is lost. The paper provides no evidence for (2): Tables 1-3 cap quantitative evaluation at 10-20 s, Figure 5 shows only a few 20/60 s spectrograms with CLAP, and Section 2.2.2 states the input size is reduced to about 3 seconds. Once a frame is removed, the model can never again condition on it; a 3 s window cannot represent an acoustic event or phrase lasting longer than the window. Therefore 'arbitrarily long, consistent' is not derived from the mechanism and is not measured; it is an extrapolation. An additional confound: the focused-timestep schedule is selected post hoc on the same test set used for the CLAP/FD tables, so the reported 'superior' numbers may reflect selection, not a general property. The concern is not that the method is fake; it is that the headline property is exactly the part that has not been tested.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InfiniteAudio, a training-free inference method for diffusion-based text-to-audio models. It adapts FIFO-Diffusion to the audio latent space by treating mel-spectrogram latent frames as a fixed-size queue, denoising each frame with a different diffusion timestep, popping finished frames, and appending new noise. A second component, called curved denoising, selects which diffusion steps to emphasize based on self-attention maps in the U-Net, reducing the number of sampling steps. Experiments on AudioLDM and VoiceLDM report CLAP, FD, KL, and subjective scores for 10-second generation, a memory-consumption comparison, and qualitative spectrograms for 20- and 60-second audio. The central claim is that the method generates arbitrarily long, temporally consistent audio with constant memory and comparable-or-better quality than the base models.","tokens_in":8593,"tokens_out":6359,"duration_ms":67562,"significance":"If the central claim held, InfiniteAudio would be a practical plug-in extension for long-duration text-to-audio generation, requiring no retraining and offering a clear memory advantage. The paper's strengths are that the FIFO mechanics are plausible, the constant-memory result in Fig. 3 is clear and convincing, and the attention-based motivation for timestep allocation is interesting. However, the evaluation is not commensurate with the headline claim: quantitative results stop at 20 seconds, the test set and the focus-region strategy are selected post hoc on the same data used for the reported scores, Table 3 uses an incompatible CLAP checkpoint, and several reported metrics contradict the abstract's 'comparable or superior across all metrics' statement. As it stands, the contribution is a promising inference heuristic with a demonstrated memory advantage, not a validated method for infinite-length consistent audio.","major_comments":[{"comment":"The claim of 'comparable or superior performance across all metrics' is not supported by Table 1. For AudioLDM, the Ours (Last focused) row is worse than the original on CLAP (0.4559 vs 0.4908) and on both subjective metrics (OVL 2.63 vs 3.03, REL 2.80 vs 3.06), improving only on FD and KL. For VoiceLDM, the Ours (Initial focused) row is worse than the original on CLAP (0.4107 vs 0.4199), FD (51.5047 vs 51.4019), KL (2.3498 vs 2.2749), and OVL (2.38 vs 2.53). Please restate the claim to reflect the metrics that actually improve, or provide a corrected evaluation.","section":"Table 1 and Abstract/Conclusion"},{"comment":"The evaluation design selects both the test subset and the focus-region strategy using the same data that produces the headline numbers. Excluding the bottom 20% of CLAP pairs based on AudioLDM/VoiceLDM predictions removes the hardest examples from every reported CLAP/FD/KL figure, and the 'Ours' rows in Table 1 are chosen after comparing four strategies on that same test set (Last focused for AudioLDM, Initial focused for VoiceLDM). This makes the reported 'superior' performance a selected maximum rather than a validated design choice. Please evaluate on the full Audiocaps test set, or on a fixed held-out split determined before any filtering, and specify the strategy-selection rule independently of the test-set outcome.","section":"§3.1.1 and §3.2.2 / Table 1"},{"comment":"The quantitative evaluation for the 'infinite-length' claim stops at 10–20 seconds (Table 3), with only a few 20/60-second spectrograms in Fig. 5, while the input window is reduced to about 3 seconds. Once a frame is popped, the model can never condition on it again, so a ~3-second window cannot represent an acoustic phrase or event longer than the window, and no experiment tracks error accumulation over many frame replacements. The headline property of arbitrarily long, temporally consistent generation is therefore an extrapolation from short-window results. Please add long-horizon tests—for example, 60–120-second generations with prompt elements separated by more than 3 seconds, per-segment CLAP/FD over successive windows, and repeated-run stability—before claiming infinite-length consistency.","section":"§2.2.2 and Table 3 / Fig. 5"},{"comment":"The CLAP scores in Table 3 are computed with a different CLAP checkpoint than the ones in Tables 1–2, so the 'fixed 10 second' baseline is not directly comparable across tables. In particular, AudioLDM's fixed-10s CLAP is 0.4559 in Table 1 but 0.3207 in Table 3, an inconsistency that the reader cannot resolve from the numbers alone. Please recompute all CLAP scores with a single checkpoint, or clearly state that no cross-table comparison is intended.","section":"Table 3 and footnote 4"},{"comment":"No error bars, confidence intervals, or multiple-run results are reported anywhere in the quantitative evaluation. The differences invoked to support 'comparable or superior' are often small (e.g., CLAP 0.4908 vs 0.4559 for AudioLDM), and without variance estimates it is impossible to tell whether they are meaningful. Please include at least standard errors over dataset subsets or random seeds for all reported metrics.","section":"Tables 1–3"}],"minor_comments":[{"comment":"The term 'curved denoising' is never defined formally. Please specify the exact step-skipping rule, the skipping factor P, and the buffer-zone size used in the experiments, since these are free parameters of the method and are needed for reproducibility.","section":"§2.2.2"},{"comment":"The description of the 'initial region' as 'early stages of sampling, where diffusion timesteps are close to τ' is ambiguous because τ decreases during sampling. Please clarify whether 'initial' refers to the temporal position in the sequence or to the noise level.","section":"§2.1"},{"comment":"The numeric CLAP values in the subfigure table are not explained in the caption. Please state which model and which prompt or evaluation set they correspond to.","section":"Fig. 4 caption"},{"comment":"The subjective evaluation says '30 speech samples,' but the task is text-to-audio generation. Please clarify whether this should read 'audio samples' and describe how the 30 items were selected.","section":"§3.1.1"},{"comment":"The model name 'V oiceLDM' appears with a spurious space in several places; also, the phrase 'we first propose InfiniteAudio' in the contributions list is redundant and should be trimmed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is written in a compact conference format and would need substantially expanded evaluation and reproducibility details for a journal audience. The core idea is worth pursuing, but the evaluation currently falls short of the title's promise, particularly regarding long-range consistency and unbiased metric reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: InfiniteAudio is a faithful port of FIFO-Diffusion to the audio-latent domain, plus an attention-guided curved denoising schedule. The memory-constant claim is demonstrated, and the no-training, drop-in property is real. What the paper does not do is measure its headline property: 'arbitrarily long, consistent' audio is extrapolated from 10–20 s samples and a roughly 3 s sliding window.\n\nWhat is genuinely new is applying the FIFO queue to mel-spectrogram latents and using self-attention maps to decide which diffusion step regions to keep. That attention analysis actually shows a difference between AudioLDM (initial frames matter) and VoiceLDM (later frames matter), and it motivates the curved schedule with fewer than 150 steps. The memory plot is clean and the argument for constant memory is straightforward. The FIFO mechanics appear sound, and the paper gives FIFO-Diffusion proper credit.\n\nThe evaluation is the weak part. The test set is pruned by throwing out the bottom 20% of CLAP pairs according to the baselines themselves; the 'best' focused-timestep strategy is chosen per model on that same test set and then reported as the method. There are no error bars on the main CLAP/FD/KL tables, and Table 3 uses a different CLAP checkpoint, so the small differences there are hard to interpret. The authors also omit the values of P and the buffer-zone size, both needed to reproduce the results.\n\nMore fundamentally, no experiment goes beyond 20 s, and the design uses about 3 seconds of input latency. Once frames are popped, the model never sees them again, so any acoustic event longer than the queue cannot be conditioned on. The claim of infinite consistency is therefore an extrapolation, not a measured result. The buffer-zone assumption that pre-trained noise levels stay on-manifold is plausible but not directly verified.\n\nCitation pattern is fine. The paper is honest about borrowing FIFO-Diffusion, and the writing is clear.\n\nWho this is for: people working on long-form text-to-audio or diffusion inference tricks. It deserves a serious referee because the core idea is usable and clearly explained, but it needs major revision before publication: code, a fixed evaluation protocol without test-set selection, longer test lengths, error bars, and explicit hyperparameters. My advice: send it to review, with instructions that the 'infinite' claim must be either demonstrated or sharply qualified.","headline":"A useful inference-time adaptation of FIFO-Diffusion to audio with a real memory win, but the 'infinite consistency' headline is extrapolated from 10–20 s tests and the evaluation protocol is too selection-biased to trust the quantitative claims.","tokens_in":9176,"tokens_out":2618,"would_cite":true,"duration_ms":30122,"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":"A fixed-size sliding window lets pretrained text-to-audio diffusion models generate audio of any length at constant memory, without retraining.","keywords":["text-to-audio generation","long-form audio","diffusion models","FIFO sampling","curved denoising","self-attention","latent diffusion","inference-only"],"falsifier":"Generate a 60-second clip with a pre-trained TTA model using InfiniteAudio while recording, at each popping step, the per-frame diffusion timestep actually fed to the model. If a plausible accounting of the training noise schedule shows that frames outside the buffer zone receive timestep values the model was never trained on (or if quality measurably degrades as audio length grows beyond the window), the central claim fails. A simpler observable falsifier: run the method on the same prompt for 20, 60, and 120 seconds and check whether CLAP score drops with length; the paper's own data only shows stability up to 60 seconds.","tokens_in":8104,"feed_emoji":"🎧","tokens_out":2496,"duration_ms":28001,"temperature":0.7,"pith_summary":"InfiniteAudio claims that a pre-trained diffusion-based text-to-audio model can generate arbitrarily long audio without any additional training. The method keeps the model's input size fixed and slides a window through the latent noise, so memory use stays constant no matter how long the output is. It also introduces curved denoising, which spends sampling steps only on the time regions that matter most, cutting the step count below 150 while preserving quality. This matters because current text-to-audio models either run out of memory on long clips or produce inconsistent audio by concatenating short segments.","feed_headline":"One sliding window generates infinite audio at fixed memory","feed_subtitle":"FIFO sampling plus curved denoising lets pretrained text-to-audio models run without retraining.","key_machinery":"The machinery is FIFO sampling with a buffer zone, adapted from video generation to audio mel-spectrogram latents, combined with curved denoising. FIFO sampling assigns a different diffusion timestep to each frame in a fixed-length latent window and advances the window by one frame per denoising step; the buffer zone consists of leading frames that are left unperturbed so the pre-trained model always sees timestep values it encountered during training. Curved denoising partitions the diffusion trajectory into initial, middle, and final regions, measures each region's average self-attention score, and either keeps or skips steps accordingly, so a model that attends mostly to early steps spends its sampling budget there.","core_discovery":"The central claim is that a latent diffusion text-to-audio model can be driven as a first-in, first-out queue: at each inference step, the model denoises a fixed-length latent window in which different frames carry different noise levels, the cleanest frame is popped out as finished audio, and a new noisy frame is pushed in at the end. Because the window never grows, memory stays bounded regardless of output length. A buffer zone of unperturbed frames keeps the per-frame noise values inside the distribution seen during training, and a curved denoising schedule derived from self-attention maps focuses sampling effort on the steps each model actually cares about. On AudioLDM and VoiceLDM, the paper reports that this inference-only procedure matches or beats the original full-sampling baselines on CLAP, FD, KL, and human ratings, while enabling seamless 20-to-60-second generation.","pith_inferences":["The attention-based step-selection criterion may transfer to other diffusion domains such as music generation or long video, since it only inspects self-attention maps that any U-Net already computes.","The buffer-zone trick is a general way to keep sliding-window diffusion in-distribution, and its length is likely a hyperparameter worth tuning separately for each base model.","Because the input window stays at roughly 3 seconds of audio, the method may cap the effective temporal context a prompt can control; prompts requiring structure over, say, 10 seconds should be tested separately.","A direct test of the core assumption would be to measure the denoiser's noise-prediction error on FIFO 'mixed-noise' inputs versus standard single-timestep inputs; if the error spikes, the buffer zone is not actually protecting the training distribution."],"forward_implications":["Any existing latent-diffusion text-to-audio model can produce arbitrarily long audio by plugging in InfiniteAudio's inference schedule, with no retraining or fine-tuning.","Memory consumption for generation becomes independent of the generated audio length, removing the principal scaling bottleneck of diffusion TTA.","The curved denoising schedule reduces the number of sampling steps below 150 while achieving quality comparable to or better than the original full-step baseline and better than equally spaced or middle-focused schedules.","Long-form generations stay temporally consistent, with stable CLAP scores across lengths from 10 to 60 seconds, in contrast to concatenation-based baselines that show repetitive patterns and discontinuities.","The method works on at least two different TTA architectures (AudioLDM and VoiceLDM) even though their attention maps assign importance to opposite ends of the sampling trajectory."],"supporting_citations":[{"why":"Supplies the FIFO sampling mechanism and the observation that mixed timesteps create a performance gap, which InfiniteAudio adapts to audio with a buffer zone.","marker":"[36]"},{"why":"AudioLDM is one of the two pre-trained TTA models on which InfiniteAudio is demonstrated and compared against.","marker":"[24]"},{"why":"VoiceLDM is the second pre-trained TTA model used to show the method's generality across different attention patterns.","marker":"[26]"},{"why":"Provides deterministic denoising (DDIM), which curved denoising relies on to skip non-essential sampling steps while preserving sample quality.","marker":"[37]"},{"why":"Defines the CLAP model that supplies the CLAP score used in the paper's main quantitative evaluation and in filtering low-quality test pairs.","marker":"[34]"},{"why":"The AudioCaps dataset provides the audio-text pairs used to evaluate generation quality.","marker":"[38]"}],"fun_headline_variants":["Audio that never ends, memory that never grows","Infinite audio via sliding-window FIFO sampling","No retraining, fixed memory: endless audio from diffusion","Curved denoising and FIFO: infinite audio in one pass","Endless text-to-audio with a fixed-size window"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The buffer zone keeps the noise levels seen by the pre-trained model within the training distribution, so the denoiser is never asked to handle out-of-distribution mixed-noise inputs.","fun_headline_variants_meta":{"raw":{"variants":["Audio that never ends, memory that never grows","Infinite audio via sliding-window FIFO sampling","No retraining, fixed memory: endless audio from diffusion","Curved denoising and FIFO: infinite audio in one pass","Endless text-to-audio with a fixed-size window"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000749,"raw_usage":{"total_tokens":3281,"prompt_tokens":836,"completion_tokens":2445,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":2364}},"tokens_in":452,"tokens_out":2445,"duration_ms":22055,"temperature":1.0,"reasoning_tokens":2364,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:10:01.410675+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a 60-second clip with a pre-trained TTA model using InfiniteAudio while recording, at each popping step, the per-frame diffusion timestep actually fed to the model. If a plausible accounting of the training noise schedule shows that frames outside the buffer zone receive timestep values the model was never trained on (or if quality measurably degrades as audio length grows beyond the window), the central claim fails. A simpler observable falsifier: run the method on the same prompt for 20, 60, and 120 seconds and check whether CLAP score drops with length; the paper's own data only shows stability up to 60 seconds.","supporting_citations":[{"cited_title":"Audiogen: Tex- tually guided audio generation,","cited_arxiv_id":null,"evidence_quote":"Supplies the FIFO sampling mechanism and the observation that mixed timesteps create a performance gap, which InfiniteAudio adapts to audio with a buffer zone."},{"cited_title":"Seeing through the conversation: Audio-visual speech separation based on diffu- sion model,","cited_arxiv_id":null,"evidence_quote":"AudioLDM is one of the two pre-trained TTA models on which InfiniteAudio is demonstrated and compared against."},{"cited_title":"Align your latents: High-resolution video synthesis with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"VoiceLDM is the second pre-trained TTA model used to show the method's generality across different attention patterns."},{"cited_title":"Audiolcm: Text-to-audio generation with latent consistency models,","cited_arxiv_id":null,"evidence_quote":"Provides deterministic denoising (DDIM), which curved denoising relies on to skip non-essential sampling steps while preserving sample quality."},{"cited_title":"Diffsound: Discrete diffusion model for text-to-sound genera- tion,","cited_arxiv_id":null,"evidence_quote":"Defines the CLAP model that supplies the CLAP score used in the paper's main quantitative evaluation and in filtering low-quality test pairs."},{"cited_title":"V oicedit: Dual-condition diffusion transformer for environment- aware speech synthesis,","cited_arxiv_id":null,"evidence_quote":"The AudioCaps dataset provides the audio-text pairs used to evaluate generation quality."}],"review_version":1}