{"id":"f720108d-6f44-46ef-aeb1-f45ef0b2dc78","arxiv_id":"2509.06389","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A MeanFlow-based one-step generator with scaled classifier-free guidance synthesizes video-to-audio and text-to-audio about 2x-500x faster than prior iterative methods with comparable automatic-metric quality.","lead":"This paper describes a video-to-audio system that generates sound from silent video in a single fast step instead of many slow ones. It uses 'average velocity' flow matching and a small guidance adjustment, reporting 2x to 500x faster synthesis with comparable quality on automatic metrics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5)'s stop-gradient objective is underdetermined: spurious K(r)/(t-r) terms satisfy the MeanFlow identity, so the one-step map is not guaranteed to recover the true average velocity.","rationale":"The reader's weakest assumption points to the same load-bearing link: Eq. (5) may not actually train u_theta to the true average velocity field. My concern makes this more specific: the MeanFlow identity admits a one-parameter family of solutions, so the stop-gradient objective is a self-consistency condition with no visible uniqueness or convergence guarantee. If this degeneracy is real, the one-step map z0 = z1 - u(z1,0,1) can be off by a spurious shift, directly invalidating the headline RTF/quality claims. The paper's empirical results cannot resolve this because no code, data, or checkpoints are released, and the relevant analysis is only cited as [13], not included. I do not think this rises to rejection: the method may still work in practice, and the MeanFlow paper may indeed prove convergence. But a conditional verdict requiring either a proof sketch or a reproducible toy verification is justified. Hence I keep the reader's CONDITIONAL verdict unchanged.","tokens_in":8261,"tokens_out":17904,"duration_ms":206979,"concrete_test":"On a 2-D Gaussian target with linear interpolation z_t=(1-t)x+tε, train a small MLP with Eq. (5) using the paper's logit-normal r,t schedule and 10% r≠t ratio until loss plateaus. Compare the learned u_theta(z1,0,1) with the exact mean-flow solution u_char(z1,0,1) (computable for this Gaussian setup), and compare the induced z0 = z1 - u_theta(z1,0,1) distribution with the target N(0,I). Repeat with an intentionally added K(r)/(t-r) output term. If the two runs converge to different one-step maps, Eq. (5) is underdetermined and the 400k-step convergence claim lacks support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Eq. (6) with u_theta trained by Eq. (5) yields a valid one-step generation map. The paper gives no derivation or convergence analysis of Eq. (5), and the consistency equation itself has a homogeneity gap. Along a flow characteristic, the identity (4), u = v - (t-r) du/dt, has general solution u(t) = (1/(t-r))∫_r^t v(s)ds + K/(t-r). Thus if u_char is any solution, u_char + K(r)/(t-r) is also a solution: the added term satisfies dw/dt = -K/(t-r)^2, so v - (t-r)d(u_char+w)/dt = u_char+w. At inference (r=0,t=1), this spurious term contributes K(0), shifting z0 = z1 - u(z1,0,1) away from the desired sample. Eq. (5)'s stop-gradient target is a self-consistency condition and does not break this degeneracy: no boundary condition, initialization argument, or theorem is given showing that SGD selects K=0. Since the entire efficiency claim rests on z0 being valid audio, this is the key unsupported link. The citation [13] may contain the missing result, but it is not reproduced or summarized here.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MF-MJT, a video-to-audio (VTA) and text-to-audio (TTA) synthesis model built on the MMAudio backbone, replacing instantaneous flow-matching velocity with a MeanFlow average-velocity formulation to enable native one-step generation. A CFG-scaled mechanism adjusts the unconditional prediction with an adaptive scalar to mitigate overshooting under one-step classifier-free guidance. Experiments on VGGSound and AudioCaps report one-step RTF = 0.007 on an H800 GPU, with perceptual metrics (FAD, FD, KL, IS, CLAP/ImageBind, DeSync) claimed to be competitive with or better than baselines including Frieren, MMAudio, and AudioLCM. The central claim is that Eq. (6), z0 = z1 - u(z1,0,1), trained via the stop-gradient objective Eq. (5), yields valid one-step audio generation without distillation or iterative refinement.","tokens_in":8571,"tokens_out":2528,"duration_ms":29647,"significance":"If the MeanFlow formulation is sound and Eq. (5) indeed trains the network to approximate the average velocity field, the contribution is practically significant: native one-step generation with an RTF of 0.007, a 2x speedup over the fastest one-step baseline and orders-of-magnitude speedup over multi-step diffusion/flow baselines, with competitive quality on standard public benchmarks. The paper also proposes a simple adaptive scalar for CFG that is plausible and easy to adopt. However, the central training objective is not derived, and the empirical evidence is limited to single-run automatic metrics without uncertainty quantification. The novelty relative to prior MeanFlow methods (reference [13]) and to the MMAudio backbone is incremental, but the engineering contribution (joint training with MeanFlow for VTA/TTA) could be useful if the theoretical gap is closed.","major_comments":[{"comment":"The stop-gradient training objective LMF = E ||u_theta(zt,r,t) - sg(u_tgt)||^2, where u_tgt = v_t - (t-r)(v_t ∂_z u_theta + ∂_t u_theta), is stated without derivation, proof, or convergence analysis. The MeanFlow identity (4) is only a necessary condition, not a characterization of the average velocity. Indeed, if u_char solves (4), then u_char + K(r)/(t-r) also solves it, since d/dt[K/(t-r)] = -K/(t-r)^2 and (t-r) times that derivative cancels. This degeneracy means the self-consistency target does not uniquely determine the true average velocity field. At inference r=0,t=1, a spurious component K(0) directly shifts z0 = z1 - u(z1,0,1), potentially invalidating the generated audio. No boundary condition, initialization argument, or theorem is given to show that the proposed training selects K=0. The paper cites [13] without reproducing the essential result. This gap is load-bearing for","section":"§2.2, Eq. (5)"},{"comment":"The experimental evidence for the 'no compromise' claim rests on single-run automatic metrics with no error bars, confidence intervals, or significance testing. Table 1 reports FAD, FD, KL, IS, IB, and DeSync for each method as point estimates; Table 2 similarly. Given that the proposed method often improves metrics by small margins (e.g., IS 9.39 vs 9.14 for Frieren one-step; CLAP 0.20 vs 0.19 for AudioLCM one-step), it is impossible to assess whether these differences are meaningful or noise. The paper should provide at least repeated evaluations (e.g., over multiple seeds or test subsets) or statistical tests. The absence of uncertainty quantifies particularly weakens the claim that MF-MJT preserves quality while accelerating.","section":"§3.3 and Tables 1-2"},{"comment":"The ablation on the r≠t sampling ratio only tests 10%, 30%, 60%, and 90%, and observes monotonic improvement as the ratio decreases. This does not demonstrate that 10% is optimal; 0% may be even better, which would undermine the necessity of MeanFlow training pairs. The explanation that r=t pairs provide 'direct supervision' is post hoc and not quantitatively supported. Since the r≠t ratio is a free hyperparameter, the choice of 10% should be justified with a more complete sweep or a clear saturation argument.","section":"§4.2, Fig. 4"},{"comment":"The scalar s = (u_theta(c)^T u_theta(∅)) / ||u_theta(∅)||^2 is an adaptive projection of the conditional velocity onto the unconditional direction. The paper states this 'reflects their alignment' and 'stabilizes guidance,' but provides no analysis of why this scalar mitigates overshooting in one-step generation. Since CFG-scaled is a stated contribution and is used in all main results, a derivation or at least a geometric intuition with a controlled comparison (e.g., with CFG-Zero* or other rescaling heuristics) is needed. Without it, the mechanism remains an ad hoc fix, and the reported gains in Fig. 3 could be attributed to other factors.","section":"§2.3, Eq. (9)"}],"minor_comments":[{"comment":"Typo: 'Notebly' should be 'Notably'.","section":"§3.4"},{"comment":"Typo: 'oth single-step and multi-step' should be 'both single-step and multi-step'.","section":"Table 2 caption"},{"comment":"The text renders 'V AE' with a space; likely a formatting issue. Also, 'Meanflow' versus 'MeanFlow' capitalization is inconsistent across the paper.","section":"§2.1"},{"comment":"The sentence 'Notably, the network does not condition directly on (r,t), but rather on (t,Δt)' is unclear: the model is given t and Δt, but the logit-normal sampling described above draws r and t. How are these encoded and combined? A clearer description of the conditioning mechanism would help reproducibility.","section":"§3.2"},{"comment":"The claim 'this objective reduces to the standard FM loss when r=t' is stated without showing the algebra. Since r=t makes the denominator in Eq. (3) zero and Eq. (4) degenerate, the limiting case requires explicit handling. Please clarify.","section":"§2.2"},{"comment":"The baseline parameter counts and sampling steps are given, but no wall-clock times are reported for baselines beyond RTF. For multi-step methods, RTF depends on the number of steps (fixed at 25); reporting RTF as a function of steps would make the efficiency comparison more transparent.","section":"§4.1, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper builds directly on the MeanFlow preprint [13] and the MMAudio architecture [8], but it does not provide enough self-contained theory to verify that the proposed training objective actually learns the average velocity field. The reviewer's concern about non-uniqueness in Eq. (4) is serious and must be addressed. If the authors can reproduce the MeanFlow derivation or provide a proof of convergence, and add uncertainty quantification to the experiments, the paper would be a solid contribution. As it stands, the central theoretical gap prevents acceptance. Also, the empirical comparison would benefit from stronger baselines in the one-step regime (e.g., InstaFlow-style distillation baselines) to contextualize the claimed speed-quality trade-off."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a legitimate efficiency result, not a paradigm shift. MF-MJT applies MeanFlow to multimodal video-to-audio and reports RTF 0.007 for one-step generation, beating Frieren on most metrics. If the MeanFlow machinery works as advertised, that is a useful practical advance for dubbing and interactive tools. The CFG-scaled scalar (Eq 9) is a sensible heuristic—project the conditional velocity onto the unconditional direction—and the ablation suggests it helps.\n\nWhat's genuinely new: the combination of MeanFlow with a multimodal joint-training backbone, and the empirical speed/quality numbers for VTA and TTA. The paper is mostly honest about the speed/quality trade-off in its tables, but the abstract's 'no compromise' is overstatement: one-step FAD is 1.46 vs MMAudio's 0.81, a real quality gap.\n\nNow the soft spots. The central training objective, Eq (5), is stated without derivation. The mean-flow identity (Eq 4) admits a family of solutions: adding K(r)/(t-r) to any solution preserves the identity, since the derivative term absorbs it. So the one-step map z0 = z1 - u(z1,0,1) is only valid if the stop-gradient fixed point selects the K=0 solution. No boundary condition, initialization argument, or convergence theorem is given. The citation [13] may contain this, but the paper doesn't summarize it. This is the key unsupported link.\n\nAlso, the empirical claims rest on single-run metrics with no error bars. Hyperparameters—CFG strength 1.5 and the 10% r!=t ratio—are chosen on the test set via the ablations in Figs 3-4, which inflates the reported performance. No code or checkpoints are released, so the numbers can't be checked independently. Given the paper's own data, 'comparable perceptual quality' is only true relative to other one-step methods, not to iterative baselines.\n\nWho should read it: anyone working on efficient audio generation or video dubbing will find the speed result worth knowing. It deserves peer review because the application is real, the experiments are reasonably broad, and the central mechanism, though borrowed, is applied to a new domain. But I'd send it back for major revision: require the authors to state the conditions under which Eq (5) converges to the true average velocity (or show the missing argument from [13]), report variance or at least multiple seeds, and stop claiming no compromise when the one-step metrics say otherwise.","headline":"A plausible one-step VTA/TTA speedup built on MeanFlow, but the training objective's degeneracy is unaddressed and the empirical evidence is thinner than the 'no compromise' claim.","tokens_in":9111,"tokens_out":3447,"would_cite":true,"duration_ms":37699,"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":"One-step video-to-audio synthesis from silent video is claimed at RTF 0.007 on an H800 GPU by modeling average velocity instead of instantaneous velocity.","keywords":["video-to-audio synthesis","MeanFlow","one-step generation","flow matching","classifier-free guidance","multimodal joint training","text-to-audio synthesis","inference efficiency"],"falsifier":"A direct test: take the trained MF-MJT model, sample z_t at several (r,t) pairs, and compute both sides of the MeanFlow identity u_theta(z_t,r,t) ≈ v_theta(z_t,t) - (t-r)(v_t ∂_z u_theta + ∂_t u_theta). If the residual is large, the stop-gradient objective has not reached its fixed point. Alternatively, train the same architecture with the standard flow-matching loss (regressing v_t) but keep one-step sampling; if quality metrics stay similar, the average-velocity parametrization is not what drives the one-step result.","tokens_in":8141,"feed_emoji":"🎧","tokens_out":5316,"duration_ms":53558,"temperature":0.7,"pith_summary":"Video-to-audio synthesis normally pays a steep inference cost: flow-matching models estimate instantaneous velocity and need many sampling steps to turn noise into sound. This paper claims that modeling the flow field with average velocity instead removes the need for iteration, giving a native one-step map from noise to audio. On top of a multimodal joint-training backbone, the authors add MeanFlow and a scalar rescaling for classifier-free guidance, and they report a real-time factor of 0.007 on an H800 GPU for both video-to-audio and text-to-audio generation, with no loss in perceptual quality, semantic alignment, or temporal synchronization relative to multi-step baselines. The significance, if correct, is that real-time interactive dubbing and audio generation from silent video become practical on a single GPU.","feed_headline":"MeanFlow makes video-to-audio synthesis a single step","feed_subtitle":"Modeling average velocity cuts video-to-audio inference to RTF 0.007 without losing quality.","key_machinery":"The MeanFlow identity is the engine: u(z_t,r,t) = (1/(t-r))∫_r^t v(z_τ,τ)dτ, and differentiating it gives u = v - (t-r) du/dt. The network u_theta is trained to satisfy this identity via a stop-gradient loss, and at inference the identity is used as an exact map, z_r = z_t - (t-r)u(z_t,r,t), which for (r,t)=(0,1) is a single step from noise to audio. The CFG-scaled scalar s projects the conditional velocity onto the unconditional direction to restrain guidance.","core_discovery":"The paper's central claim is that the average velocity field u(z_t,r,t), defined as the time-average of the instantaneous velocity along the flow between r and t, satisfies a self-consistent identity, and a network trained with a stop-gradient objective to respect that identity can generate audio in one step: z0 = z1 - u(z1,0,1). The authors instantiate this in MF-MJT, a multimodal diffusion-transformer network jointly trained on video, text, and audio, and show that one-step generation reaches RTF 0.007 on VGGSound and AudioCaps while matching or beating Frieren in one step and remaining competitive with MMAudio in multi-step settings. They also introduce an adaptive scalar rescaling of the","pith_inferences":["Editorial inference: if the stop-gradient objective reliably reaches the MeanFlow fixed point, the same one-step trick could be applied to other latent diffusion backbones beyond MMAudio, e.g., image or long-form video generation.","Editorial inference: the ablation showing that low ratios of r≠t pairs help semantic and temporal alignment suggests a curriculum that anneals the r≠t ratio during training could improve both one-step fidelity and alignment.","Editorial inference: RTF 0.007 is measured on an H800 GPU with fixed sample lengths; the paper does not report latency on consumer GPUs or CPU, so the practical 'real-time' claim beyond H800 is untested."],"forward_implications":["One-step inference at RTF 0.007 on an H800 GPU, with 2x to 500x speedups over the compared VTA and TTA baselines, makes real-time audio synthesis from silent video feasible on a single GPU.","The same MeanFlow formulation supports variable numbers of sampling steps: multi-step MF-MJT remains competitive with MMAudio on VGGSound and beats it on several TTA metrics on AudioCaps.","Classifier-free guidance can be used in one-step generation without the usual artifacts because the scalar rescaling adaptively damps the unconditional component.","Because the network is jointly trained with multimodal conditions, the VTA-trained model transfers to text-to-audio synthesis without fine-tuning."],"supporting_citations":[{"why":"Supplies the mean-flow formulation, the MeanFlow identity, and the stop-gradient training objective that the paper adopts for one-step generation.","marker":"[13]"},{"why":"Supplies the multimodal joint-training backbone (MMAudio) that MF-MJT builds on, including cross-modal attention and audio-visual sync conditioning.","marker":"[8]"},{"why":"Supplies the idea that an optimized scalar can mitigate CFG overshooting in one-step generation, which the paper adapts as CFG-scaled.","marker":"[15]"},{"why":"Supplies the Synchformer visual encoder used to extract synchronization features that condition generation.","marker":"[20]"},{"why":"Acts as the primary one-step VTA baseline; its rectified flow and distillation approach is the efficiency paradigm the paper compares against.","marker":"[1]"},{"why":"Acts as the efficient TTA baseline whose one-step and multi-step results MF-MJT is compared against on AudioCaps.","marker":"[30]"},{"why":"Provides the VGGSound dataset used for joint training and for the VTA test evaluation.","marker":"[22]"},{"why":"Provides the AudioCaps dataset used for TTA evaluation.","marker":"[23]"}],"fun_headline_variants":["One-step video-to-audio: MeanFlow hits RTF 0.007","MeanFlow: video-to-audio in a single step","Average velocity enables one-step video-to-audio","Video-to-audio in one step: MeanFlow","MeanFlow: one-step video-to-audio"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the stop-gradient objective in Eq. (5) actually trains u_theta to approximate the true average velocity field, and that after 400,000 training steps this fixed point is reached; the paper presents this objective without derivation or convergence analysis.","fun_headline_variants_meta":{"raw":{"variants":["One-step video-to-audio: MeanFlow hits RTF 0.007","MeanFlow: video-to-audio in a single step","Average velocity enables one-step video-to-audio","Video-to-audio in one step: MeanFlow","MeanFlow: one-step video-to-audio"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001149,"raw_usage":{"total_tokens":4584,"prompt_tokens":713,"completion_tokens":3871,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":3800}},"tokens_in":457,"tokens_out":3871,"duration_ms":32097,"temperature":1.0,"reasoning_tokens":3800,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:40:02.530510+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test: take the trained MF-MJT model, sample z_t at several (r,t) pairs, and compute both sides of the MeanFlow identity u_theta(z_t,r,t) ≈ v_theta(z_t,t) - (t-r)(v_t ∂_z u_theta + ∂_t u_theta). If the residual is large, the stop-gradient objective has not reached its fixed point. Alternatively, train the same architecture with the standard flow-matching loss (regressing v_t) but keep one-step sampling; if quality metrics stay similar, the average-velocity parametrization is not what drives the one-step result.","supporting_citations":[{"cited_title":"MMAudio: Tam- ing multimodal joint training for high-quality video-to-audio synthesis,","cited_arxiv_id":null,"evidence_quote":"Supplies the mean-flow formulation, the MeanFlow identity, and the stop-gradient training objective that the paper adopts for one-step generation."},{"cited_title":"AudioLDM: Text-to-audio generation with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"Supplies the multimodal joint-training backbone (MMAudio) that MF-MJT builds on, including cross-modal attention and audio-visual sync conditioning."},{"cited_title":"Denoising diffusion probabilis- tic models,","cited_arxiv_id":null,"evidence_quote":"Supplies the idea that an optimized scalar can mitigate CFG overshooting in one-step generation, which the paper adapts as CFG-scaled."},{"cited_title":"MeanFlow-Accelerated Multimodal Video-to-Audio Synthesis via One-Step Generation","cited_arxiv_id":"2509.06389","evidence_quote":"Acts as the primary one-step VTA baseline; its rectified flow and distillation approach is the efficiency paradigm the paper compares against."},{"cited_title":"Scalable diffusion models with trans- formers,","cited_arxiv_id":null,"evidence_quote":"Provides the VGGSound dataset used for joint training and for the VTA test evaluation."},{"cited_title":"Learning transfer- able visual models from natural language supervision,","cited_arxiv_id":null,"evidence_quote":"Provides the AudioCaps dataset used for TTA evaluation."}],"review_version":1}