{"id":"8eddd64d-9360-4d4d-9a27-18a3af2cd7bf","arxiv_id":"2508.20604","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A text-to-motion model that combines noise-conditioned guidance with latent-space sampling to produce more varied 3D motions at little cost to text alignment.","lead":"Diverse-T2M generates many different 3D animations from one text description by adding a random noise signal as an extra condition and by sampling from a learned latent distribution. On standard benchmarks it reports much higher diversity than prior methods while keeping text-motion accuracy nearly unchanged.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unablated 'Length Stochastic Enhancement' on the noise branch may inflate MultiModality via random temporal scaling, not semantically meaningful aleatoric diversity, so the central claim is not yet established.","rationale":"The reader's weakest assumption is that random Gaussian noise fused through Eq. (4) carries meaningful diversity information rather than corrupting text consistency. My concern agrees with that general direction but identifies a more specific, concrete mechanism: the length-stretch augmentation applied only to the noise-conditioned branch can inflate MultiModality through random temporal scaling, independent of any learned aleatoric diversity. This is load-bearing because MultiModality is the headline evidence for the central claim, and the paper does not ablate this augmentation or report its parameters. It is also testable: a straightforward ablation that removes the augmentation and re-selects w on a validation split would determine whether the diversity gain persists. I do not move the verdict beyond CONDITIONAL because the paper could in principle survive this check; the concern is a missing control and a missing ablation, not a demonstrated contradiction. The reader's other flags, namely the test-set sweep over w and the unreported p_noise, remain relevant but are secondary to identifying what actually causes the diversity gain.","tokens_in":18582,"tokens_out":8699,"duration_ms":93197,"concrete_test":"Retrain or rerun Diverse-T2M with the Length Stochastic Enhancement disabled, fixing and reporting p_noise and all augmentation parameters, and select w on a validation split before computing test metrics. If MultiModality drops substantially while R Precision stays near 52% on HumanML3D, or if the w needed to reach 3.976 forces R Precision below the MoMask baseline, the headline diversity claim is driven by random temporal scaling. A complementary check is to generate motions for prompts containing 'fast'/'slow' and measure whether the intended speed is preserved in the diverse samples.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline result is the MultiModality gain (HumanML3D: 3.976 vs. 2.799 for the next best method). MultiModality is computed as the average pairwise feature distance among 30 generated motions per text, so any source of uncontrolled variation can inflate it. Section III-B2 introduces a 'Length Stochastic Enhancement' that randomly stretches and compresses target motion sequences, but applies it only to the noise-conditioned training branch, explicitly to avoid conflicts with speed words in the text. This creates a confound: the noise branch p_theta(z|epsilon) is trained to emit code sequences corresponding to random temporal scalings of the paired motion. At inference, Eq. (4) fuses I_text and I_noise with weight w=3, so I_noise can inject random speed/duration variation into the fused code distribution. That alone could raise MultiModality substantially without adding semantically justified pose-level alternatives. The paper neither ablates this augmentation nor reports its strength or sampling ratio, and it does not evaluate text consistency on prompts containing speed or duration terms. The few qualitative examples cannot rule out this mechanism. Therefore, the reported diversity gain may be an artifact of random temporal scaling rather than evidence for the proposed aleatoric latent diversity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Diverse-T2M, a two-stage text-to-motion generation method built on residual VQ-VAE and a masked transformer. To increase diversity, the authors add a noise-conditioned generation branch alongside the text-conditioned branch and fuse the two code distributions through I = (1+w)I_text - w I_noise (Eq. 4). They also replace the deterministic text-to-code mapping with a variational predictor that samples a latent embedding from a Gaussian and maps it to code probabilities. On HumanML3D and KIT-ML the method reports MultiModality 3.976 and 3.462, substantially above prior methods, while maintaining R-Precision and FID at roughly the level of MoMask. Ablations in Table III separate the contributions of the variational predictor and the noise signal, and confidence intervals are reported for all benchmark numbers.","tokens_in":18823,"tokens_out":5444,"duration_ms":47941,"significance":"If the reported diversity gains are real and attributable to the proposed mechanism, the paper makes a useful practical contribution: a simple modification to a standard two-stage architecture that substantially raises the MultiModality metric without sacrificing text consistency, with qualitative examples and an application story. The paper is also commendable for reporting 95% confidence intervals and for performing component ablations on two baselines. However, the central claim is not yet established because the diversity gain may be inflated by a training-only temporal-stretch augmentation on the noise branch and by test-set selection of the fusion weight. These issues are fixable within the manuscript's scope, so the contribution is promising but requires substantive revision.","major_comments":[{"comment":"In Section III-B2, the 'Length Stochastic Enhancement' randomly stretches and compresses the motion sequences only when training the noise-conditioned branch. Because MultiModality is computed as the average pairwise feature distance among 30 generated motions for the same text, random duration or speed variation alone can inflate it without adding semantically meaningful pose-level alternatives. The +NS rows in Table III therefore conflate the proposed noise signal with this augmentation. Please ablate the length enhancement (for example, train the noise-conditioned branch without temporal stretch, or apply the same augmentation to the text branch), report the stretch/compress range and sampling probability, and evaluate text consistency on prompts containing speed or duration terms such as 'fast' or 'slow'.","section":"III-B2, Table III"},{"comment":"Figure 5 sweeps w over a range and the paper then fixes w=3 for the reported results; if this sweep is performed on the test set, the headline MultiModality value is the result of optimizing the diversity metric on the evaluation split. The figure shows MultiModality increasing with w while R-Precision and MM-Dist deteriorate, so the chosen operating point directly trades away semantic control to maximize the reported diversity. Please clarify which split was used for the sweep, use a validation split for selecting w, and report the sensitivity of the main conclusions to this choice.","section":"IV-A, Figure 5"},{"comment":"Two training hyperparameters central to the method are not reported: the noise-mixing probability p_noise introduced in Section III-B2 and the text/noise mixing ratio used in training in Section III-D2. In addition, the KL weight is set to 1e-5, which is effectively no regularization of the latent distribution toward N(0,1); this weakens the claim that the predictor is variational and leaves open whether the latent sampler contributes anything beyond an unconstrained stochastic mapping. Please report p_noise and the training mixture ratio, and ablate the KL weight over a range of values.","section":"III-B2, III-D2, IV-A"},{"comment":"Equation (3) defines p_t^k as the probability of code k at time t, and I_text and I_noise are described as probability distributions. Equation (4) then computes I = (1+w)I_text - w I_noise, which with w>0 can produce negative entries and does not in general sum to one, so it is not a valid probability distribution. If I_text and I_noise are actually logits, Eq. (3) and the surrounding text should say so; if they are probabilities, the fusion step needs to be defined in logit space or normalized explicitly. Please resolve this inconsistency, as it determines how the method is implemented and whether the 'probability distribution' terminology is accurate.","section":"Eq. (3)-(4)"}],"minor_comments":[{"comment":"There are numerous typos and inconsistent naming: 'Diverser-T2M' in Section I, 'tockens' in Section III-C, 'infernce' in Section III-C, 'Digram' in Section IV-C, and 'three two benchmark datasets' in Section IV-A should be corrected.","section":"Throughout"},{"comment":"The caption of Figure 1 contains 'Derails' and 'Fid'; the Introduction states the MultiModality gain is '30% higher' than the second place, while Table I shows 3.976 vs 2.799, which is about 42% higher; please reconcile the numbers.","section":"Introduction, Figure 1"},{"comment":"In the related work, 'MDM [27]' is cited with the same reference number as MotionCLIP [27], while the actual MDM paper appears to be reference [51]; please correct this citation error and verify that all references are uniquely assigned.","section":"Section II, References"},{"comment":"The experimental settings report the codebook size as 512x512 with 512-dimensional vectors, but Eq. (2) denotes V residual codebooks; please clarify the relationship between the residual layers (V=6) and the codebook structure, and state the dimension d_c used in the experiments.","section":"Section IV-A"},{"comment":"In Eq. (13), the notation [mask]_t is used both as an indicator variable and as a token placeholder; please define the notation explicitly to avoid ambiguity about which positions are masked.","section":"Eq. (13)"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically interesting and the benchmark results are strong, but the main claims hinge on the confounds described in Major Comments 1 and 2. If the authors can provide an ablation that separates the length augmentation, move the w selection to a validation split, and clarify Eq. (4), I would be willing to reconsider favorably. The paper's fit to TCSVT is reasonable given its applications section, though the methodological novelty share is modest."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThe paper's headline result is a MultiModality jump from 2.799 to 3.976 on HumanML3D. I think you should know that part of this gain probably comes from an unablated trick, not from semantically meaningful aleatoric diversity. The 'Length Stochastic Enhancement' in Section III-B2 randomly stretches and compresses motion sequences, but only on the noise-conditioned training branch. MultiModality is computed as the average pairwise feature distance among 30 generated samples, so random speed variation alone can inflate it. The paper never ablates this augmentation, nor reports its strength or the mixing probability p_noise. That is a real confound.\n\nWhat is genuinely new here is the combination: a noise-conditioned guidance branch fused with the text branch via (1+w)I_text - w I_noise, plus a VAE-style latent sampler inside a masked transformer. The ablations do separate the contributions of these two components, and the benchmark tables include 20-run confidence intervals. That is more careful than many papers in this area. The 'noise as diversity carrier' framing is reasonable, even if it borrows from classifier-free guidance.\n\nThe soft spots in proportion: (1) the length augmentation confound is the biggest issue and is fixable with one ablation; (2) w=3 is selected on test-set metrics (Figure 5), so the reported numbers are partially tuned; (3) p_noise is unreported and no code is released; (4) the KL weight is 1e-5, making the latent-space sampler nearly unconstrained—which may be fine, but the paper does not analyze it. There is also no task-based or human evaluation of what the diversity looks like beyond a few qualitative examples.\n\nIf the authors ablate the length enhancement, report p_noise, and evaluate w on a validation split, the central claim could hold up. As it stands, I would not use these numbers as evidence for semantically meaningful diversity. The paper deserves a serious referee, but the referee should ask for those experiments before acceptance. For a reading group, it's worth discussing as a case of evaluation pitfalls in generative diversity. I would not cite it yet.","headline":"The MultiModality gains are likely inflated by an unablated temporal-stretch augmentation, so the headline diversity numbers should be treated with caution until the confound is resolved.","tokens_in":19357,"tokens_out":3181,"would_cite":false,"duration_ms":30234,"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":"Injecting noise nearly triples the diversity of text-driven 3D motion.","keywords":["text-to-motion generation","3D human motion","generation diversity","aleatoric uncertainty","noise guidance","variational transformer","residual VQ-VAE","masked generative transformer"],"falsifier":"Fix a text prompt and a fixed latent-space draw, then generate many motions using only different noise signals $g_\\epsilon$. If the spread of these motions, measured by pairwise distance or MultiModality, is no larger than when the noise signal is held fixed and the latent draw varies, then the noise carrier is not the diversity source the paper claims; checking whether those motions all satisfy the text through R-Precision would reveal whether the extra variety is semantically controlled.","tokens_in":18381,"feed_emoji":"🕺","tokens_out":4915,"duration_ms":43802,"temperature":0.7,"pith_summary":"Text-to-motion models that discretize motion into code sequences produce precise motions but tend to generate nearly identical outputs for the same prompt, because one text maps to one code sequence. Diverse-T2M claims that injecting a fresh Gaussian noise signal as a second conditioning input, together with sampling from a learned latent distribution, turns each inference into a draw from a range of text-consistent motions. On HumanML3D the method reports a MultiModality of 3.976 against 2.799 for the next-best method, while keeping Top-1 retrieval precision at 52.5% and FID at 0.057; similar gains appear on KIT-ML. The intended consequence is that animators get many usable, semantically faithful variants from a single sentence.","feed_headline":"Injecting noise nearly triples the diversity of text-driven 3D motion","feed_subtitle":"Diverse-T2M reports a 3.976 MultiModality on HumanML3D while keeping top-1 text retrieval at 52.5 percent.","key_machinery":"The load-bearing mechanism is the stochastic fusion of two code distributions. The same variational predictor is asked to map a CLIP text feature and a Gaussian noise vector separately into distributions over motion codes; the output distribution is $I=(1+w)I_{\\text{text}} - w I_{\\text{noise}}$, borrowing the classifier-free guidance combination. A latent space sampler inside the predictor, parameterized as a diagonal Gaussian via mean and standard-deviation transformer layers, injects reparameterized randomness $z=\\mu+\\sigma\\odot\\epsilon$ before the code-probability MLP. The fusion weight $w$ is the single dial that trades text control against diversity: the paper sweeps it on the test set and chooses $w=3$.","core_discovery":"The paper's central discovery is that the diversity bottleneck in two-stage text-to-motion generation is not the discrete codebook but the deterministic mapping from text to codes. By training a noise-conditional model alongside the text-conditional model and fusing their code distributions as $I=(1+w)I_{\\text{text}}-w I_{\\text{noise}}$, with $w=3$ at inference, the model obtains markedly higher MultiModality while preserving retrieval accuracy and low FID. The second mechanism, a variational text-to-codes predictor, replaces the one-to-one mapping with a distribution $q_\\phi(z|e)=\\mathcal{N}(z;\\mu,\\mathrm{diag}(\\sigma^2))$ and samples $z=\\mu+\\sigma\\odot\\epsilon$ with $\\epsilon\\sim\\mathcal{N}(0,I)$, adding stochasticity inside the transformer. On HumanML3D the ablation shows each component contributes, with the noise signal adding the larger part of the diversity gain over the RVQ-VAE baseline.","pith_inferences":["Editorial inference: if the noise signal is truly the diversity carrier, then conditioning on structured noise, such as a latent style code, rather than isotropic Gaussian could turn the method into a controllable diversity knob that lets users request variation along a chosen axis.","Editorial inference: the same fusion rule could be tested in other conditional token generators, such as speech or video codec transformers, where map-like architectures also suffer from low sample diversity.","Editorial inference: because $w=3$ is selected on the test set, a deployment-ready version would need validation-based selection or a per-prompt adaptive weight; otherwise the reported trade-off curve may rely on test-set information."],"forward_implications":["A single text prompt can yield a batch of distinct, semantically valid motions, which is what character animation pipelines need for downstream selection and editing.","Two-stage token-based generators can reach diversity comparable to or better than diffusion models without iterative denoising at inference.","Diversity and text consistency are not in inherent tension: on HumanML3D the method reports R-Precision rising from 0.518 to 0.525 while MultiModality rises from 1.221 to 3.976 over its RVQ baseline.","The fusion weight $w$ gives practitioners a tunable trade-off; raising $w$ increases variety at the cost of retrieval accuracy and multimodal distance, as shown in the paper's sweep."],"supporting_citations":[{"why":"Supplies the masked-transformer, two-stage baseline and residual VQ-VAE design that Diverse-T2M builds on and compares against.","marker":"[14]"},{"why":"Establishes the T2M-GPT two-stage paradigm of discrete motion codes plus text-to-code prediction that the method follows.","marker":"[15]"},{"why":"Provides the HumanML3D dataset, evaluation protocol, and the standard metrics (R-Precision, FID, Multimodal Distance, MultiModality) used for all comparisons.","marker":"[47]"},{"why":"Introduces residual vector quantization used for the motion codebook in the first stage.","marker":"[21]"},{"why":"Defines VQ-VAE discrete representation learning underlying the codebook and decoder.","marker":"[19]"},{"why":"Supplies the KIT-ML dataset used for the second benchmark evaluation.","marker":"[54]"}],"fun_headline_variants":["Noise signals triple diversity in text-to-3D motion","Uncertainty-driven text-to-motion: triple the diversity","Diverse-T2M: noise signals triple 3D motion diversity","Text-driven 3D motion: sampling noise for diversity","Aleatoric uncertainty triples diversity in text-to-motion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands or falls on the premise that a Gaussian noise vector, combined with the text signal through $I=(1+w)I_{\\text{text}}-w I_{\\text{noise}}$, adds useful variation rather than random semantic corruption; if the noise term mostly bypasses the text condition, the reported diversity gain would not be text-consistent variation.","fun_headline_variants_meta":{"raw":{"variants":["Noise signals triple diversity in text-to-3D motion","Uncertainty-driven text-to-motion: triple the diversity","Diverse-T2M: noise signals triple 3D motion diversity","Text-driven 3D motion: sampling noise for diversity","Aleatoric uncertainty triples diversity in text-to-motion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001133,"raw_usage":{"total_tokens":4718,"prompt_tokens":970,"completion_tokens":3748,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":3661}},"tokens_in":586,"tokens_out":3748,"duration_ms":24448,"temperature":1.0,"reasoning_tokens":3661,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:42:16.744074+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a text prompt and a fixed latent-space draw, then generate many motions using only different noise signals $g_\\epsilon$. If the spread of these motions, measured by pairwise distance or MultiModality, is no larger than when the noise signal is held fixed and the latent draw varies, then the noise carrier is not the diversity source the paper claims; checking whether those motions all satisfy the text through R-Precision would reveal whether the extra variety is semantically controlled.","supporting_citations":[{"cited_title":"Momask: Generative masked modeling of 3d human motions,","cited_arxiv_id":null,"evidence_quote":"Supplies the masked-transformer, two-stage baseline and residual VQ-VAE design that Diverse-T2M builds on and compares against."}],"review_version":2}