{"id":"0ed3bd1d-200d-4c8e-ba39-28532792cd73","arxiv_id":"2505.00998","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A VQVAE plus optimal-transport flow matching plus a noise-injected SDE sampler yields diverse 3D human motions with fewer training parameters than diffusion baselines.","lead":"This paper proposes a two-stage human motion generator: a VQVAE compresses motions into a latent space, then a flow-matching ODE maps Gaussian noise to that space, followed by a training-free stochastic sampler that adds noise for diversity. The authors report state-of-the-art scores on HumanAct12 and HumanML3D with fewer parameters than diffusion baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DivSDE sampler in Eq. 13 is self-referential and relies on an unvalidated Gaussian proxy score, so the reported diversity gains are not reproducible as written.","rationale":"The reader identified the self-referential update in Eq. 13 and the unvalidated score proxy as the weakest assumption; my independent reading of the manuscript reaches the same conclusion. The deterministic flow-matching and optimal-transport components are standard and could be sound, so the central idea is not necessarily false. However, the stochastic sampler that is supposed to deliver diversity without additional training is not a well-defined reverse SDE as written: the update equation is self-referential, the score formula conflicts with the marginal stated in Proposition 3 unless η is silently set to 1, and the score is centered at a deterministic proxy rather than the true latent variable. Because every reported diversity and FID result depends on this sampler, the empirical claims cannot be verified from the paper alone. This supports the reader's CONDITIONAL verdict: the paper should not be accepted until the sampler is disambiguated, the corrected algorithm is implemented, and the experiments are reproduced. I do not recommend REJECT because the identified problems are localized and potentially fixable, and the remaining pipeline is built from established components.","tokens_in":17932,"tokens_out":15501,"duration_ms":164212,"concrete_test":"Implement two corrected variants of Algorithm 1 for unconditional generation on HumanAct12: (A) explicit Euler-Maruyama with the score evaluated at the previous state z_{t+Δt,i}; (B) implicit Euler where the linear score term is solved for z_{t,i}. Compare FID, KID, Precision, Recall, and Diversity against Table 1. Additionally, for the best variant, measure the distance from the pre-decoding latent samples to the nearest VQVAE codebook entry and compare the empirical latent distribution with the training latent distribution. If neither variant reproduces Table 1 within reported confidence intervals, or if the latents are far from the codebook distribution, the DivSDE mechanism does not justify the paper's diversity and accuracy claims.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim depends on the stochastic diverse output generation procedure (DivSDE). In Eq. 13 and Algorithm 1, the score term is computed from z_{t,i} before z_{t,i} has been updated; the right-hand side contains the left-hand side, so the discrete step is not an explicit Euler-Maruyama update and no implicit solve is provided. Even under the natural explicit reading (evaluate the score at the previous state z_{t+Δt,i}), the 'score' is a Gaussian centered at the deterministic endpoint ez0,i = DerODE(ez1,i), not the score of the true latent motion distribution. Proposition 3 states the marginal is p(z_t) = N((1-t)z_i, η^2 t^2 I), whose score is ((1-t)z_i - z_t)/(η^2 t^2), not the /t^2 expression used in the text. The sampler therefore has no well-defined output distribution as written, and the diversity and FID numbers in Tables 1, 2, and 5 cannot be independently reproduced or traced to the stated stochastic process. Since the paper's headline advantage over SGMs is diversity without extra training parameters, this step is the load-bearing part of the claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DSDFM, a two-stage generative model for conditional and unconditional human motion synthesis. Stage one trains a VQVAE to map human motion sequences into a discrete latent space. Stage two learns a deterministic transport map (DerODE) between a standard Gaussian prior and the latent motion distribution using an optimal-transport flow-matching objective with an additional drift-consistency loss. To increase sample diversity, the authors propose DivSDE, a training-free stochastic sampler that augments the deterministic ODE endpoint with a reverse-SDE-style update whose score term is computed from the DerODE output. Experiments on HumanAct12 and HumanML3D report state-of-the-art FID, KID, precision, recall, and diversity with fewer parameters than prior diffusion-based motion generators, together with reduced training and inference time.","tokens_in":18105,"tokens_out":5559,"duration_ms":60023,"significance":"If the proposed sampler were rigorously derived and reproducible, the core idea would be attractive: augmenting a deterministic flow-matching generator with a training-free stochastic refinement step that increases diversity without additional parameters is a practically valuable contribution. The use of optimal transport coupling and the drift-consistency regularizer for learning near-straight trajectories is also reasonable, and the authors provide proof sketches for Propositions 2 and 3. The conditional and unconditional evaluations cover standard benchmarks and compare against several recent methods. However, the central stochastic-sampling step, which underpins the headline diversity and parameter-efficiency claims, is not a well-defined stochastic process as written, and the reported numerical results therefore cannot be traced to the stated method.","major_comments":[{"comment":"","section":"Eq. (13), Algorithm 1"},{"comment":"","section":"Proposition 3 and Eq. (13)"},{"comment":"","section":"Tables 1, 2, 5"},{"comment":"","section":"Section 4.2 and Table 2"}],"minor_comments":[{"comment":"","section":"Eq. (13) vs. Algorithm 1"},{"comment":"","section":"Eqs. (6)-(7)"},{"comment":"","section":"Eq. (12), Appendix B"},{"comment":"","section":"Table 2"},{"comment":"","section":"Proposition 1"},{"comment":"","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern lands: the DivSDE update is self-referential and the score proxy is not derived from the stated marginal law. This is the load-bearing component of the paper, so the revision must fix the sampler definition and either justify the ODE-endpoint proxy or retract the no-additional-training claim. If the corrected sampler requires a learned score network, the parameter-efficiency advantage over SGMs largely disappears, and the paper's central contribution would need to be reframed. I also noted the missing conditioning mechanism for the conditional experiments, which is a separate but important gap."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis is a two-stage human motion generator: VQVAE reconstruction plus a flow-matching model (DerODE) trained with an optimal-transport path and a consistency-style drift loss, then a training-free stochastic sampler (DivSDE) at inference meant to buy diversity. The components are known; the combination is new and the paper is honest about that, pointing to Lipman et al. for Proposition 1 and to consistency models for the drift loss. What it does well: the training objective is simple and cheap, the parameter count is genuinely lower than SGM baselines, and the empirical gains on HumanAct12, while modest, are consistent across metrics. Proposition 3 is a correct solution of a linear SDE, and the paper's own math is largely standard.\n\nThe soft spot is load-bearing. The DivSDE update in Eq. 13 and Algorithm 1 is not implementable as written. The score term is evaluated at z_{t,i} before that variable is updated, so the discrete step is self-referential. And the score formula itself is wrong: Proposition 3 says the marginal is N((1-t)z_i, eta^2 t^2 I), whose score is ((1-t)z_i - z_t)/(eta^2 t^2), not /t^2 as written. The missing eta^2 matters because eta is exactly the diversity dial. Beyond the typo, the score is a heuristic proxy, a Gaussian centered at the deterministic ODE endpoint, with no derivation showing it is close to the true latent score. Since the paper's headline advantage over SGMs is diversity without extra training parameters, and since the diversity numbers are directly controlled by eta, this step is exactly where the claim lives.\n\nThe empirical section is otherwise okay but not as strong as the abstract claims. The improvements over the best SOTA are in the 1-5% range. Table 4 is an ablation, not a full comparison on HumanML3D conditional generation, so the state-of-the-art statement is only partially supported. Missing ablations for eta and lambda_cl further soften the diversity claim.\n\nIf I were referee, I would ask for: a corrected, implementable sampler (with eta^2 in the score), code and evaluation protocol, ablations on eta and lambda_cl, and a proper comparison table on HumanML3D. The central flow-matching training math holds up; it is the inference-time stochastic add-on that is underspecified.\n\nThis paper is for the human motion synthesis crowd and for anyone engineering efficient latent-space generators. It deserves a serious peer review, since the empirical package is plausible and the fix path is clear, but I would not cite the diversity numbers until the sampler is reproducible.","headline":"A competent flow-matching plus VQVAE motion synthesis pipeline whose diversity-generating sampler is written in a way that cannot be run as stated; the results are plausible but not reproducible until Eq. 13 is fixed.","tokens_in":18718,"tokens_out":2266,"would_cite":false,"duration_ms":21158,"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 deterministic ODE plus a training-free stochastic sampler beats prior score-based and diffusion baselines on human motion synthesis while using fewer parameters.","keywords":["human motion synthesis","latent feature mapping","deterministic ordinary differential equation","stochastic differential equation","optimal transport flow matching","vector quantized variational autoencoder","diversity in generative models","action-conditioned motion generation"],"falsifier":"Run the exact reverse update in Eq. 13 on a synthetic latent space with a known mixture-of-Gaussians target: if the two-sample distance between DivSDE samples and the true target does not shrink as $\\eta \\to 0$ and as the discretization step shrinks, or if the self-referential term in Eq. 13 makes the iteration unexecutable as written, then the claim that the stochastic sampler samples the latent motion distribution is not supported.","tokens_in":17669,"feed_emoji":"🏃","tokens_out":9514,"duration_ms":92531,"temperature":0.7,"pith_summary":"The paper claims that a deterministic-to-stochastic two-stage mapping can replace score-based diffusion for human motion synthesis while getting both accuracy and diversity at lower training cost. The first stage trains a VQVAE to encode motion sequences into a codebook latent space. The second stage trains a drift network, using optimal-transport coupling and a drift-consistency loss, so the path from standard Gaussian noise to that latent space is straight, then adds a training-free stochastic reverse sampler (DivSDE) that reuses the deterministic endpoint as a proxy score to inject diversity. Reported results on HumanAct12 and HumanML3D put DSDFM ahead of compared diffusion and latent-diffusion baselines on fidelity and diversity metrics while using fewer parameters and less training and inference time.","feed_headline":"Two-stage latent mapping beats diffusion models for human motion","feed_subtitle":"A straight deterministic ODE plus a training-free stochastic sampler lifts diversity and accuracy while cutting parameter counts.","key_machinery":"The load-bearing object is the pair (DerODE, DivSDE). DerODE is a deterministic ODE whose drift is learned through the drift-estimate loss (Eq. 8) and the drift-consistency loss (Eq. 9) on optimal-transport-matched pairs $(z_0, z_1)$, so the learned vector field approximates the constant displacement $z_1 - z_0$ and the trajectories are straight. DivSDE is the reverse SDE $dz_t = -\\frac{1}{1-t} z_t\\,dt + \\eta\\sqrt{\\frac{2t}{1-t}}\\,dw_t$ from Proposition 3, whose score term is replaced by the closed-form Gaussian score $\\frac{(1-t)\\hat{z}_{0,i} - z_t}{t^2}$ with $\\hat{z}_{0,i} = \\text{DerODE}(\\hat{z}_{1,i})$; this is what adds diversity without extra training. Proposition 1 supplies the drift formula $u(z,t) = \\sigma'(t)\\frac{z - \\mu(t)}{\\sigma(t)} + \\mu'(t)$, which turns the chosen linear interpolation between latent data and Gaussian noise into the constant drift $z_1 - z_0$, making the whole second stage a parameter-light flow model between two fixed distributions.","core_discovery":"On the paper's own terms, the central claim is that accuracy and diversity can be decoupled: a deterministic ordinary differential equation (DerODE) maps Gaussian noise to the VQVAE latent space along straight paths, and a reverse stochastic differential equation (DivSDE) run only at sampling time turns that deterministic map into a diverse generator. The drift network is trained to predict the constant displacement $z_1 - z_0$ along optimal-transport couplings, with a drift-consistency regularizer that makes the prediction time-independent, so no score network is trained. At inference, DivSDE reuses the DerODE output $\\hat{z}_{0,i}$ as the center of a Gaussian proxy score and adds controlled noise of strength $\\eta$; larger $\\eta$ gives more diversity. The paper reports that this two-stage recipe sets the best listed unconditional FID, KID, precision, recall, and diversity numbers on HumanAct12 and improves or matches conditional motion generation with tighter confidence intervals, all while using roughly half the parameters of the nearest baselines.","pith_inferences":["Editorial inference: because DerODE and DivSDE operate entirely in a VQVAE latent space and never touch motion-specific structure, the same two-stage recipe should transfer to other continuous latent spaces such as images, video, or audio; the paper only demonstrates it on human motion.","Editorial inference: the diversity strength $\\eta$ is a one-parameter knob the paper does not sweep systematically; a natural testable extension is to plot FID versus diversity as $\\eta$ varies and check whether the trade-off is monotone and where it saturates.","Editorial inference: the shortcut of using the DerODE endpoint as the center of a Gaussian score can be read as amortized score estimation, which connects DSDFM to distillation and consistency-model ideas; the paper does not pursue that reading."],"forward_implications":["On HumanAct12 unconditional generation, DSDFM reports the best listed FID (12.86), KID (0.10), precision (0.75), recall (0.85), and diversity (18.41) while using 15M parameters, fewer than every compared baseline.","On the action-to-motion task, DSDFM reports the best conditional FID (0.068 on HumanAct12) and accuracy (0.994) among compared methods, with tighter 95% confidence intervals on accuracy and diversity than MotionDiffuse.","In the 500-epoch ablation, DSDFM cuts training time on HumanAct12 to 25.33 minutes versus 42.93 for VPSDE and 40.57 for VESDE, and at 500 inference steps reaches FID 12.86 compared to 15.63 and 14.92 for those SDE baselines.","Because no score network is trained, the stochastic diversity stage can be tuned at sampling time by changing $\\eta$, so a single trained model can move along the diversity-fidelity trade-off without retraining."],"supporting_citations":[{"why":"Supplies the VQVAE architecture and codebook whose latent space is the target distribution that DerODE maps into.","marker":"[47]"},{"why":"Provides the Gaussian conditional path derivation behind Proposition 1 and the flow-matching drift objective that the DerODE loss adapts.","marker":"[19]"},{"why":"Supplies the reverse-time SDE and score framework that DivSDE reverses and against which DSDFM is compared.","marker":"[39]"},{"why":"Supplies the straight-flow and rectified-flow idea behind the paper's claim that DerODE has straight training trajectories.","marker":"[27]"},{"why":"Motivates the drift-consistency loss that makes the learned drift independent of interpolation time.","marker":"[40]"},{"why":"Provides the latent-diffusion baseline (MLD) that DSDFM must beat on action-conditioned motion generation.","marker":"[3]"},{"why":"Supplies the HumanAct12 benchmark used for the unconditional and action-to-motion comparisons and for the action recognition accuracy metric.","marker":"[9]"},{"why":"Supplies the HumanML3D benchmark used for the second set of conditional generation results.","marker":"[10]"}],"fun_headline_variants":["Two-stage latent mapping beats diffusion for human motion","Deterministic-to-stochastic mapping improves motion diversity","No score net needed: DSDFM boosts synthesis with half params","Straight ODE plus sampling noise enhances motion generation","Decoupled accuracy and diversity in human motion synthesis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the deterministic endpoint $\\hat{z}_{0,i}$ from DerODE is a faithful proxy for the true latent motion distribution, so the reverse SDE's score computed from the Gaussian around $\\hat{z}_{0,i}$ samples the correct distribution; the discrete update in Eq. 13 also references $z_{t,i}$ before it is updated, so as written the sampler is not directly implementable.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage latent mapping beats diffusion for human motion","Deterministic-to-stochastic mapping improves motion diversity","No score net needed: DSDFM boosts synthesis with half params","Straight ODE plus sampling noise enhances motion generation","Decoupled accuracy and diversity in human motion synthesis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1448,"prompt_tokens":943,"completion_tokens":505,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":428}},"tokens_in":559,"tokens_out":505,"duration_ms":5303,"temperature":1.0,"reasoning_tokens":428,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:30:22.959877+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the exact reverse update in Eq. 13 on a synthetic latent space with a known mixture-of-Gaussians target: if the two-sample distance between DivSDE samples and the true target does not shrink as $\\eta \\to 0$ and as the discretization step shrinks, or if the self-referential term in Eq. 13 makes the iteration unexecutable as written, then the claim that the stochastic sampler samples the latent motion distribution is not supported.","supporting_citations":[{"cited_title":"SCORE-BASED GENERATIVE MODELING THROUGH STOCHASTIC DIFFERENTIAL EQUATIONS","cited_arxiv_id":null,"evidence_quote":"Supplies the reverse-time SDE and score framework that DivSDE reverses and against which DSDFM is compared."},{"cited_title":"Ac- tion2motion: Conditioned generation of 3d human motions","cited_arxiv_id":null,"evidence_quote":"Supplies the HumanAct12 benchmark used for the unconditional and action-to-motion comparisons and for the action recognition accuracy metric."},{"cited_title":"Generating diverse and natural 3d human motions from text","cited_arxiv_id":null,"evidence_quote":"Supplies the HumanML3D benchmark used for the second set of conditional generation results."}],"review_version":1}