{"id":"210097a9-e592-40b3-bf72-aeceb847220b","arxiv_id":"2504.20334","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Modifying the flow-matching training target lets F5-TTS synthesize speech without classifier-free guidance at inference, halving per-step cost and improving measured WER, SIM-O, and MOS.","lead":"This paper trains a flow-matching text-to-speech model to predict an adjusted velocity field so that inference needs one conditional pass instead of classifier-free guidance's two passes. The authors report quality equal to or better than F5-TTS with CFG on LibriTTS, with the table showing up to 7.75x lower real-time factor.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq (10) has a fixed point that damps the conditional-unconditional difference, not the CFG extrapolation from Eq (7); the central mechanism is unsupported.","rationale":"The reader's weakest assumption is exactly the load-bearing concern: Eq. (10) is asserted, not derived, and its fixed point damps rather than amplifies the conditional-unconditional difference. My independent analysis confirms that the loss as written minimizes to v_c = (u_c + w v_u)/(1+w), which for w=0.7 is an interpolation between conditional and unconditional fields, not the CFG extrapolation v_CFG = u_u + w(u_c - u_u). The paper's claim that substituting Eq. (7) into Eq. (9) yields Eq. (10) is algebraically incorrect, and Algorithm 1 is consistent with Eq. (10) if the unconditional output in the loss is a typo for the conditional output. This is not a mere technical gap; it invalidates the central theoretical contribution. The empirical results are interesting and might indicate a useful training technique, but without a corrected derivation or released code to demonstrate the actual learned trajectories, the paper's main claim fails as stated. A rejection is appropriate, though a revised version that correctly characterizes the training objective and provides code could be reconsidered.","tokens_in":8787,"tokens_out":8006,"duration_ms":77500,"concrete_test":"Re-derive the optimality condition of Eq. (10) with the stop-gradient operation and confirm that the fixed point is v_c = (u_c + w v_u)/(1+w). Then run a controlled toy experiment: train a small flow-matching model on a 2D Gaussian mixture with the exact loss in Eq. (10) (w=0.7) and compare the learned conditional vector field to the CFG target from Eq. (7) with guidance scale 2. If the learned field matches the damped average rather than the CFG extrapolation, the paper's mechanism claim is falsified.","verdict_should_be":"REJECT","load_bearing_attack":"Section III.B defines the model-guidance loss as LMG-CFM = E||vt(x|y) + w·sg(vt(x|y) - vt(x)) - ut(x|y)||^2 (Eq. 10). Setting the gradient with respect to the conditional output to zero, treating the stop-gradient copy as constant, gives the fixed point (1+w)vt(x|y) - w vt(x) = ut(x|y), i.e. vt(x|y) = (ut(x|y) + w vt(x))/(1+w). For w=0.7, this is a weighted average of the true conditional velocity and the unconditional model output. It does not reproduce the CFG interpolation from Eq. (7), v_CFG = ut(x) + w(ut(x|y) - ut(x)), which for w>1 extrapolates the difference. The paper claims that substituting Eq. (7) into Eq. (9) yields Eq. (10), but the algebra does not work: substituting the model analogue of Eq. (7), namely vt(x) + w(vt(x|y) - vt(x)), would give ||vt(x) + w(vt(x|y) - vt(x)) - ut(x|y)||^2, not Eq. (10). The actual Eq. (10) has the opposite sign on the difference term relative to the intended CFG form. Therefore the central claim that the trained conditional model directly approximates the CFG-optimized trajectory is contradicted by the paper's own equations. The positive empirical results might stem from a different, possibly useful, regularizing effect, but the paper provides no theoretical or experimental evidence for the claimed mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a training objective for flow-matching-based text-to-speech that aims to eliminate classifier-free guidance at inference time by directly approximating the CFG trajectory. The method modifies the flow-matching loss with a stop-gradient difference term between conditional and unconditional outputs. Experiments on LibriTTS with F5-TTS show improved WER, SIM-O, and MOS at lower NFE, with substantial RTF reduction. However, the derivation of the loss is algebraically incorrect, and the algorithm as written does not match the equation, undermining the central claim.","tokens_in":9193,"tokens_out":6318,"duration_ms":62221,"significance":"If the central claim were correct, the method would be practically significant: it promises roughly a 9x inference speed-up for flow-matching TTS by removing the second forward pass required by CFG, without degrading quality. The empirical results are indeed striking—at NFE 7 without CFG, the proposed model achieves WER 2.02%, SIM-O 0.601, MAS 4.101, and RTF 0.04, compared with the baseline's WER 2.28%, SIM-O 0.592, MOS 4.026, and RTF 0.31 at NFE 32 with CFG. The paper also reports faster convergence, which is interesting. However, the theoretical justification given in Section III.B is not valid, and the training loss as written has a fixed point that moves the conditional field toward the unconditional field, the opposite of CFG's extrapolation. The stop-gradient also makes the target self-referential, so the method is better described as a self-distillation bootstrap than as approximating CFG. The empirical gains are therefore not explained by the paper's narrative, and the method's reproducibility is compromised by an inconsistency between Equation (10) and Algorithm 1.","major_comments":[{"comment":"The derivation of Equation (10) from Equation (7) is algebraically incorrect. Substituting the model analogue of Equation (7), i.e., v_t(x) + w(v_t(x|y) - v_t(x)), into the loss in Equation (9) would yield E||v_t(x) + w(v_t(x|y) - v_t(x)) - u_t(x|y)||^2, not Equation (10). Moreover, the fixed point of Equation (10) is found by setting the gradient with respect to v_t(x|y) to zero while treating the stop-gradient copy as constant, giving (1+w) v_t(x|y) - w v_t(x) = u_t(x|y), i.e., v_t(x|y) = (u_t(x|y) + w v_t(x)) / (1+w). For w = 0.7, this is a weighted average that damps the difference between conditional and unconditional fields, not the CFG extrapolation of Equation (7), which for w > 1 amplifies the difference. Thus the paper's central claim that the trained conditional model directly approximates the CFG-optimized trajectory is contradicted by its own equations.","section":"Section III.B, Eq. (10)"},{"comment":"Algorithm 1 does not implement Equation (10). Line 6 modifies the target to u'_t = u_t(x0|c) - w·Delta v_t, and Line 7 computes the loss as ||v_t(x0) - u'_t||^2, where v_t(x0) is the unconditional model output. Thus the algorithm trains the unconditional output to match a target that depends on the conditional output and the true conditional velocity, while Equation (10) trains the conditional output. Since the stop-gradient in Delta v_t prevents gradients from flowing through the conditional output, the two losses have fundamentally different gradient flows. This inconsistency makes the method ambiguous and prevents reproduction from the description alone.","section":"Algorithm 1"},{"comment":"The abstract claims a 9x inference speed-up, but the measured RTF in Table I gives a ratio of 0.31/0.04 = 7.75 for the key comparison. The 9x figure appears to be based on counting forward passes (32 steps x 2 for CFG vs 7 steps x 1), which should be stated explicitly. Additionally, the objective metrics are reported without confidence intervals or statistical significance tests; given that the method's mechanism is not supported by the derivation, these results should be interpreted cautiously.","section":"Table I and Abstract"},{"comment":"The ablation on stop-gradient is described too vaguely. The statement that without stop-gradient 'the model can't inference normally, and the inference results are all noise' is not quantified or analyzed. Since the stop-gradient is essential to the method and creates a self-referential training target (Delta v_t depends on the model's own outputs), the paper should provide a fixed-point or convergence analysis, or at least a clear explanation of why the stop-gradient is necessary. As written, this is a load-bearing unsupported assertion.","section":"Section IV.D.2"}],"minor_comments":[{"comment":"The caption refers to 'w (as defined in Equation (8))', but Equation (8) defines the conditional dropout training, not w; w first appears in Equation (10).","section":"Figure 1 caption"},{"comment":"Equation (10) uses Delta v_t(x,y) without a vertical bar, while the preceding text uses Delta v_t(x|y); the notation should be consistent.","section":"Equation (10)"},{"comment":"The abstract contains a grammatical error: 'we reformulated' should be 'we reformulate' in the present tense. Also, the sentence 'Furthermore, It can be seamlessly integrated' has an unnecessary capital 'I'.","section":"Abstract"},{"comment":"The text says 'establishing a linear interpolation between unconditional and conditional dynamics', but for w > 1 the expression is an extrapolation, not an interpolation; this wording is misleading.","section":"Section III.A"},{"comment":"The paper says 'the overall training overhead is approximately 1.5 times higher than that of the previous methods' and later 'this method does not increase training overall overhead'. The first statement refers to per-update overhead, and the second to total training time; this distinction should be made explicit to avoid apparent contradiction.","section":"Section IV.D.1"}],"recommendation":"reject","confidential_remarks":"The empirical results in Table I are intriguing and might reflect a genuinely useful training technique, but the manuscript's central theoretical claim is demonstrably incorrect: Equation (10) does not follow from Equation (7), and its fixed point damps rather than amplifies the conditional-unconditional difference. Moreover, Algorithm 1 implements a different loss from Equation (10), which suggests the authors may have a different method in mind than the one they describe. These are load-bearing errors that cannot be fixed by simple rewording; the paper would need a substantially revised derivation and a clear statement of the actual training objective to be publishable in its current form. I recommend rejection, though I would be open to reconsidering a revised manuscript that accurately describes the method and provides a sound analysis of why it works."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the empirical result is genuinely interesting—removing CFG at inference while matching or beating the CFG baseline on WER and MOS would be valuable. But the paper hangs its mechanism on Eq (10), and the algebra in the text does not support it. The fixed point of the loss as written is a weighted average of the conditional target and the unconditional model output, not the extrapolation in Eq (7). That's the opposite of what CFG does. So the central claim fails as stated.\n\nWhat's new and good: it's the first TTS application of Tang et al.'s model-guidance idea to flow-matching TTS, concretely to F5-TTS. The adaptation to joint text/audio conditioning is nontrivial. The experiments are reasonably thorough for a conference paper: LibriTTS train, LibriSpeech-PC eval, WER/SIM-O/MOS/RTF, ablations on w and stop-gradient, convergence curves. The NFE-7 no-CFG row (SIM-O 0.601, WER 2.02, MOS 4.101, RTF 0.04) is a strong empirical point if it reproduces.\n\nThe soft spots are real. First, Eq (10) is not obtained by substituting Eq (7) into Eq (9). With stop-gradient, the fixed point satisfies (1+w)v_c = u_c + w v_u, i.e. v_c = (u_c + w v_u)/(1+w). For w=0.7 that damps the conditional-unconditional difference, the opposite of CFG's extrapolation. Second, there's an internal inconsistency: Algorithm 1 writes the loss as ||v_t(x) - (u_t(x|c) - wΔv)||^2 while Eq (10) writes ||v_t(x|y) + wΔv - u_t(x|y)||^2. These are not the same objective. Third, the 9x speedup claim in the abstract overstates the table: the 0.31 RTF is at NFE 32 with CFG; at the same NFE 7 the CFG baseline has RTF 0.07, so the per-step saving is ~1.75x and the big speedup comes from fewer steps. Fourth, no code is provided, despite the abstract promising release, so the central mechanism can't be checked.\n\nVerdict: the paper deserves referee time because the empirical direction is useful and the flaw is fixable in principle—the authors may have an effective regularizer even if the CFG story is wrong. But a serious referee should demand a corrected derivation, an explicit statement of what the loss in Algorithm 1 actually optimizes, and code. If the derivation cannot be fixed, the paper should be reframed as an empirical study of a new training target, not as 'no-CFG'.","headline":"Useful empirical result obscured by a wrong derivation: Eq (10) does not yield CFG as written, so the central mechanism is unsupported.","tokens_in":9697,"tokens_out":3439,"would_cite":false,"duration_ms":30517,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that replacing the flow-matching training target with a CFG-adjusted target lets a TTS model synthesize speech with a single conditional forward pass, matching or beating the CFG baseline at a fraction of the compute.","keywords":["classifier-free guidance","flow matching","text-to-speech","inference efficiency","model guidance","F5-TTS","zero-shot TTS","conditional flow matching"],"falsifier":"The sharpest test is to compute, for fixed $(x,t,c)$, the global minimizer of the training loss in Eq. (10). If that minimizer is not identical to the CFG target $v^{\\mathrm{CFG}}_t(x|c)=u_t(x)+w\\,(u_t(x|c)-u_t(x))$, the objective does not train the model to reproduce CFG, and any quality gain on the LibriSpeech-PC test set must be attributed to something else. A complementary experiment would compare the single-pass no-CFG model directly with the CFG-guided baseline on identical prompts at matched numbers of function evaluations.","tokens_in":8631,"feed_emoji":"🎙️","tokens_out":6707,"duration_ms":62813,"temperature":0.7,"pith_summary":"This paper asks whether classifier-free guidance (CFG) can be removed from flow-matching-based text-to-speech at inference time without losing quality. It proposes to modify the training objective so the model learns the CFG trajectory directly: the target vector field becomes the conditional field minus a weighted, gradient-stopped difference between conditional and unconditional velocities. At inference the model needs only one conditional forward pass per sampling step instead of two, halving per-step cost. Using F5-TTS on LibriTTS, the method reports WER 2.02%, SIM-O 0.601, MOS 4.101 and RTF 0.04 at 7 steps without CFG, against the F5-TTS baseline with CFG at WER 2.28%, SIM-O 0.592, MOS 4.026 and RTF 0.31 at 32 steps, a 9x inference speed-up. If this holds, CFG's double forward pass can be dropped in real-time TTS systems.","feed_headline":"Training tweak drops CFG from flow-matching TTS","feed_subtitle":"Modified training target makes one conditional pass match two-pass CFG quality at 9x speed-up.","key_machinery":"The key mechanism is the model-guidance loss for conditional flow matching, which appends to the standard conditional flow-matching loss a weighted difference between conditional and unconditional velocity fields, with a stop-gradient on the unconditional branch. It is designed so that the learned conditional field itself matches the CFG-modulated field $v^{\\mathrm{CFG}}_t(x|y) = u_t(x) + w\\,(u_t(x|y)-u_t(x))$; at inference the sampler solves $dx/dt = v_\\theta(x_t,t)$ with only the conditional network evaluation. Gradient stopping protects the unconditional branch from back-propagated gradients, and Sway Sampling or any other ODE solver can be applied without retraining.","core_discovery":"The central discovery is that a flow-matching vector field can be trained to approximate the CFG-adjusted target by adding a gradient-stopped penalty term to the conditional velocity target. The training loss becomes $L_{\\mathrm{MG-CFM}} = \\mathbb{E}\\|v_t(x|c) + w \\cdot \\Delta v_t(x,c) - u_t(x|c)\\|^2$, where $\\Delta v_t(x,c) = \\mathrm{sg}(v_t(x|c)-v_t(x))$ and $\\mathrm{sg}$ is the stop-gradient operation. The stated effect is that after training, inference uses only the conditional model output, with no unconditional evaluation and no guidance-scale tuning. The paper verifies this on F5-TTS and finds that at 7 sampling steps the no-CFG model beats the 32-step CFG baseline on word error rate, speaker similarity, and predicted MOS, while running roughly 9 times faster.","pith_inferences":["The paper's 9x speed-up stacks two effects: dropping the second forward pass (2x) and reducing sampling steps from 32 to 7 (about 4.5x). The two are independent, and Table I suggests the no-CFG model's MOS advantage over CFG appears mainly at 7 steps; at 32 steps the no-CFG model scores lower MOS than the CFG baseline. A reader should not read the 9x as attributable to CFG removal alone.","A testable extension is to apply the same gradient-stopped model-guidance objective to other flow-matching generators, such as video-to-audio or music generation, where the conditional and unconditional network share parameters; the same one-pass advantage should transfer if the loss is the mechanism.","The paper implicitly assumes that the conditional and unconditional fields can be estimated by the same network with dropout. If a future architecture uses separate heads for the two fields, the stop-gradient term and the single-pass trick would need to be re-engineered."],"forward_implications":["Per sampling step, inference cost drops from two forward passes to one, eliminating the unconditional network evaluation and any guidance-scale tuning at decode time.","The modified model composes with step-reduction samplers such as Sway Sampling, so per-step savings multiply with step-count savings rather than competing with them.","The training budget is not increased overall: although each update needs roughly 1.5x compute, convergence takes about half the updates, so total training cost is comparable.","The method is architecture-agnostic within flow-matching TTS: any model trained with conditional/unconditional dropout can adopt the modified target, not only the F5-TTS case study.","At the reported operating point (7 steps, no CFG), the model reaches RTF 0.04 on a consumer GPU, which is comfortably real-time."],"supporting_citations":[{"why":"Defines flow matching and the optimal-transport conditional vector-field training objective that the paper modifies.","marker":"[13]"},{"why":"Introduces classifier-free guidance, the two-pass inference procedure the paper aims to remove.","marker":"[21]"},{"why":"Supplies the model-guidance training idea of absorbing guidance into the training target, which the paper extends to text-to-speech.","marker":"[20]"},{"why":"F5-TTS is the flow-matching TTS architecture used as the case study and baseline.","marker":"[6]"},{"why":"LibriTTS is the training corpus, and its evaluation setup defines the experimental comparison.","marker":"[34]"},{"why":"The WavLM-based speaker verification model provides the SIM-O metric used to compare speaker similarity.","marker":"[35]"},{"why":"The NISQA model provides the predicted MOS used to compare generated speech quality.","marker":"[36]"}],"fun_headline_variants":["No-CFG flow matching for TTS: 9x faster","Drop CFG in flow-matching TTS, keep quality","Train flow TTS without guidance, run 9x faster","CFG-free flow matching: one pass beats two","9x speedup for TTS by eliminating CFG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole fast-inference scheme stands on the premise that the modified training loss in Eq. (10) really teaches the model the classifier-free-guidance trajectory, so that one conditional pass at inference can stand in for two guided passes. If that premise is wrong, the method may still produce decent speech, but not for the reason the paper gives.","fun_headline_variants_meta":{"raw":{"variants":["No-CFG flow matching for TTS: 9x faster","Drop CFG in flow-matching TTS, keep quality","Train flow TTS without guidance, run 9x faster","CFG-free flow matching: one pass beats two","9x speedup for TTS by eliminating CFG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1445,"prompt_tokens":933,"completion_tokens":512,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":428}},"tokens_in":549,"tokens_out":512,"duration_ms":5251,"temperature":1.0,"reasoning_tokens":428,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:32:24.662677+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The sharpest test is to compute, for fixed $(x,t,c)$, the global minimizer of the training loss in Eq. (10). If that minimizer is not identical to the CFG target $v^{\\mathrm{CFG}}_t(x|c)=u_t(x)+w\\,(u_t(x|c)-u_t(x))$, the objective does not train the model to reproduce CFG, and any quality gain on the LibriSpeech-PC test set must be attributed to something else. A complementary experiment would compare the single-pass no-CFG model directly with the CFG-guided baseline on identical prompts at matched numbers of function evaluations.","supporting_citations":[{"cited_title":"Flow matching for generative modeling,","cited_arxiv_id":null,"evidence_quote":"Defines flow matching and the optimal-transport conditional vector-field training objective that the paper modifies."},{"cited_title":"Large-scale self-supervised speech representation learning for automatic speaker verification,","cited_arxiv_id":null,"evidence_quote":"The WavLM-based speaker verification model provides the SIM-O metric used to compare speaker similarity."}],"review_version":1}