{"id":"bc9b9b75-c104-43b9-afca-ce7e6bb3d1c6","arxiv_id":"2411.19527","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Replacing the decoder of discrete motion generation models with a rectified flow decoder improves naturalness and FID in text-to-motion, co-speech gesture, and music-to-dance generation.","lead":"DisCoRD replaces the final token-to-motion decoder in discrete human motion generation with a rectified flow network that refines Gaussian noise into smooth motion under token guidance, and introduces a jerk-based metric, sJPE, for measuring motion naturalness. On HumanML3D and KIT-ML it reports state-of-the-art FID of 0.032 and 0.169 while keeping faithfulness for models with rich tokenizers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim of 'without compromising faithfulness' is contradicted by the paper's own Table 2: R-Precision drops for T2M-GPT on both HumanML3D and KIT-ML, and reconstruction MPJPE rises for every baseline in Table 1; the method's benefit is real but must be scoped to high-capacity tokenizers.","rationale":"The reader's verdict is CONDITIONAL, and my analysis supports that conclusion rather than overturning it. The strongest support for the paper is the consistent and large FID improvement in generation (e.g., 0.045 to 0.032 for MoMask on HumanML3D, 0.204 to 0.169 on KIT-ML) and the user-study evidence that DisCoRD is preferred over MoMask and MDM on naturalness. Those results are credible conditional on code release and independent verification. The load-bearing flaw is not an internal mathematical inconsistency; it is that the headline faithfulness guarantee is too broad and is contradicted by the paper's own tables. I partially agree with the reader's identified weakest assumption: the Condition Projection in Section 3.2 is a plausible information bottleneck, but the more fundamental issue is the stochastic rectified-flow decoding itself, which can map one token sequence to many motions. The T2M-GPT rows expose this directly, and the reconstruction MPJPE increases in Table 1 show the same trade-off even without generation. This does not require rejecting the method; it requires narrowing the central claim to tokenizers with sufficient representational capacity, as the paper's own Section 4 discussion partially acknowledges. The reader already asked for this qualification, so no verdict change is needed.","tokens_in":21010,"tokens_out":6721,"duration_ms":60869,"concrete_test":"Re-run the T2M-GPT + DisCoRD experiment from Table 2 with deterministic decoding: condition on the same generated tokens but fix X0 to a constant (or average outputs over 16 noise seeds) and recompute R-Precision and MM-Dist. If R-Precision recovers to the T2M-GPT baseline, the faithfulness drop is caused by stochastic decoding and could be mitigated; if it does not, the vanilla VQ tokenizer lacks the capacity needed for faithful conditional decoding. Either outcome shows that the abstract's unconditional faithfulness claim is unsupported and should be narrowed to high-capacity tokenizers. As a secondary check, recompute Table 1 MPJPE with ground-truth tokens; the reported increases already demonstrate that reconstruction faithfulness is compromised for all tested tokenizers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that DisCoRD 'enhances naturalness without compromising faithfulness to the conditioning signals on diverse settings.' The paper's own numbers contradict this. In Table 2 on HumanML3D, T2M-GPT R-Precision drops from 0.491/0.680/0.775 to 0.476/0.663/0.760; on KIT-ML it drops from 0.398/0.606/0.729 to 0.382/0.590/0.715. In Table 1, reconstruction MPJPE increases for every baseline (T2M-GPT: 60.0 to 71.5; MMM: 46.9 to 56.8; MoMask: 29.5 to 33.3), which is a direct faithfulness loss at the decoding stage even when the tokens are ground-truth tokens. The paper's own discussion attributes the T2M-GPT generation drop to 'a single token can map to multiple motions in DisCoRD,' meaning the rectified flow decoder is stochastic and low-capacity vanilla VQ tokens do not fully determine the motion. This is a structural trade-off, not merely an implementation detail: for tokenizers with insufficient capacity, naturalness is improved at the expense of faithfulness. The reader's concern about the Condition Projection (Section 3.2) is plausible but secondary; even an injective linear projection would not fix the stochastic-decoding bottleneck if the token sequence itself is not a sufficient statistic for the text-conditioned motion. The contributions list and abstract should explicitly state the high-capacity-tokenizer condition under which faithfulness is preserved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DisCoRD, a decoder replacement for discrete motion-generation models. Instead of decoding discrete tokens with a feed-forward VQ-VAE decoder, the method repeats each token to frame resolution, projects it to conditioning features, and trains a conditional rectified-flow model to map Gaussian noise to raw motion conditioned on those features. The authors evaluate DisCoRD on motion reconstruction and text-to-motion generation on HumanML3D and KIT-ML, and on co-speech gesture and music-to-dance generation, reporting FID improvements over T2M-GPT, MMM, BAMM, and MoMask while claiming that faithfulness is preserved. They also introduce a new metric, symmetric Jerk Percentage Error (sJPE), designed to detect frame-wise noise and under-reconstruction.","tokens_in":21329,"tokens_out":4178,"duration_ms":36826,"significance":"If the claims are properly scoped, DisCoRD is a useful and non-invasive contribution: it can be attached to any existing discrete motion generator, and the reported FID gains for MoMask and BAMM are substantial and internally consistent. The paper also addresses a real measurement gap by proposing a sample-wise naturalness metric and by providing a user study linking sJPE to human judgment, which is a constructive step beyond relying only on FID and MPJPE. The main weakness is that the abstract and contribution statements overclaim faithfulness preservation, because the paper's own tables show a faithfulness loss for low-capacity tokenizers and a universal increase in reconstruction MPJPE. The underlying mechanism is plausible and the empirical evidence for high-capacity tokenizers is strong, but the claims need to be narrowed and the sJPE validation needs more detail.","major_comments":[{"comment":"The abstract and Contribution 1 state that DisCoRD 'enhances naturalness without compromising faithfulness to the conditioning signals on diverse settings,' but the paper's own results contradict this as stated. In Table 2, T2M-GPT R-Precision drops on HumanML3D from 0.491/0.680/0.775 to 0.476/0.663/0.760 and on KIT-ML from 0.398/0.606/0.729 to 0.382/0.590/0.715; in Table 1, reconstruction MPJPE increases for every baseline (T2M-GPT: 60.0 to 71.5; MMM: 46.9 to 56.8; MoMask: 29.5 to 33.3). The discussion in Section 4.2 itself acknowledges the T2M-GPT faithfulness decline. The claims should be explicitly scoped to tokenizers with sufficient representational capacity, as Table 5 suggests, or the abstract and contributions must be revised to describe the trade-off rather than asserting no compromise.","section":"Abstract and Contribution 1; Tables 1 and 2"},{"comment":"The abstract's 'state-of-the-art performance, with FID of 0.032 on HumanML3D and 0.169 on KIT-ML' is not supported on KIT-ML, where Table 2 lists ReMoDiffuse with FID 0.155. Please qualify the SOTA claim (for example, as best among discrete-decoder methods, or as best on naturalness under the sJPE criterion) and reconcile the stated numbers with the full comparison table.","section":"Abstract; Table 2"},{"comment":"The paper's central naturalness improvement is measured primarily by sJPE, a metric proposed in this same paper and designed specifically to penalize the artifacts that DisCoRD targets. The supplementary user study reports a higher Pearson correlation between sJPE and human naturalness scores than between MPJPE and human scores (0.483 versus 0.181), but it does not report the number of participants, number of rated samples, confidence intervals, or the exact exclusion criterion beyond 'lowest 10% of samples in terms of human score standard deviation.' Because the naturalness claim rests heavily on this metric, a fuller validation protocol and statistical reporting are needed before sJPE can serve as the primary evidence.","section":"Section 4.1 and Supplementary Section C.4"}],"minor_comments":[{"comment":"The caption contains a typo: 'fasterer' should be 'faster'.","section":"Figure 6 caption"},{"comment":"The notation in Equation (5) is unclear about whether jerk is summed over joints or computed per joint; please define the joint aggregation used to obtain a scalar Jpred,t and Jtrue,t.","section":"Equation (5)"},{"comment":"The target values 'Distk →(9.780)' and 'Distg →(7.662)' are unexplained; please define the arrow notation and state explicitly that these are ground-truth reference values.","section":"Table 4"},{"comment":"The supplementary table reports FIDk/FIDg degradation for TM2D+DisCoRD (23.98/88.74 versus 19.01/20.09), which appears to conflict with the main-text statement that DisCoRD outperforms the baseline on standard metrics; please clarify which supplementary metrics are considered reliable and why they are reported despite the conflict.","section":"Supplementary Table D"},{"comment":"Several entries in Table B have unresolved citation placeholders ('Fg-T2M [?]', 'M2DM [?]', 'MotionGPT [?]', 'MotionGPT-2 [?]', 'AttT2M [?]'); these should be replaced with proper references.","section":"Supplementary Table B"}],"recommendation":"major_revision","confidential_remarks":"The central mechanism is credible for high-capacity tokenizers, and the FID improvements on MoMask and BAMM are the strongest evidence. The main issue is claim-calibration: the abstract's blanket 'without compromising faithfulness' is contradicted by the paper's own numbers, and the SOTA claim on KIT-ML is not accurate against Table 2. I would encourage the authors to reframe the contribution around the tokenizer-capacity condition and to strengthen the sJPE validation; with those changes, the paper would be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the decoder idea is real and the FID gains on MoMask and BAMM look solid, but the abstract's 'without compromising faithfulness' is contradicted by the paper's own tables. The good part: replacing a discrete motion decoder with a rectified flow model, using a repeat-and-project condition module and sliding-window training, is a genuinely useful contribution. The sJPE metric also targets a real evaluation gap—FID genuinely misses frame-wise jitter. The experiments span text, speech, and music, and for RQVAE-based tokenizers (MoMask, BAMM) R-Precision holds while FID improves substantially. That alone is worth publishing.\n\nThe soft spots are real, though. Table 2 shows T2M-GPT R-Precision drops on both datasets, and Table 1 shows reconstruction MPJPE rises for every baseline—including MoMask. The paper waves this away by saying MPJPE doesn't reflect naturalness, but it does measure positional fidelity, and these are reconstructions from ground-truth tokens. The paper's own explanation—'a single token can map to multiple motions in DisCoRD'—means the stochastic decoder is the cause. That is structural, not an implementation detail: low-capacity tokenizers do not give the flow model enough information. The authors acknowledge this in Section 4.2 but the abstract and contributions still make the universal claim. The fix is easy: scope the faithfulness claim to high-capacity tokenizers and report the MPJPE increase honestly. The Condition Projection concern is secondary; even an injective projection would not fix the token-sufficiency problem.\n\nThe sJPE metric is more self-referential than the authors let on. It is designed to detect exactly the artifacts DisCoRD reduces, and the correlation with human preference comes from a user study that excludes the lowest 10% of samples by agreement and is reported without confidence intervals. There is independent support—standard FID gains and a user study—so this is fixable, but the metric needs a more careful validation.\n\nAlso: no code or data are linked, so the numbers cannot be independently checked. The dance results are mixed in the supplementary (FIDk/FIDg get worse); the authors call those metrics unreliable, which is fair, but they still include Distk/Distg selectively.\n\nBottom line: this deserves a serious referee. The core mechanism works and the paper is well-engineered, but the claims need to be scoped and the metric work needs tightening before it is ready.","headline":"Useful decoder replacement with real FID gains on high-capacity tokenizers, but the 'without compromising faithfulness' claim is contradicted by the paper's own tables and needs scoping.","tokens_in":21913,"tokens_out":2840,"would_cite":true,"duration_ms":25040,"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":"This paper claims that decoding discrete motion tokens can be reframed as a conditional generation problem, and that a rectified-flow decoder operating in the raw continuous motion space yields smoother, more natural motion than the…","keywords":["discrete motion generation","rectified flow","motion tokens","text-to-motion generation","VQ-VAE","flow matching","human motion synthesis","motion naturalness"],"falsifier":"Measure how much token information survives the Condition Projection by training a small linear classifier that predicts the original codebook index from the projected features C on held-out motions. If classification accuracy falls far below the accuracy achievable from the raw codebook embeddings themselves, the projection is the faithfulness bottleneck, and the observed R-Precision drop for T2M-GPT is explained by information loss rather than by anything the flow decoder does.","tokens_in":20758,"feed_emoji":"🕺","tokens_out":5603,"duration_ms":44413,"temperature":0.7,"pith_summary":"Discrete motion generation models (VQ-VAE-based) map text or other signals to a small set of tokens, then decode those tokens back into motion with a feed-forward network. That decoder is the weak point: it under-reconstructs fine details and injects frame-wise noise. DisCoRD replaces it with a rectified-flow model that treats the tokens as conditioning and generates motion by transporting Gaussian noise to the raw motion distribution, iteratively. The paper reports that this swap improves FID to 0.032 on HumanML3D and 0.169 on KIT-ML for text-to-motion generation, while preserving or even slightly improving R-Precision for tokenizers based on residual quantization (MoMask, BAMM). The authors also introduce sJPE, a jerk-based metric, arguing that FID misses exactly the sample-wise noise and under-reconstruction that their method removes.","feed_headline":"Flow decoding turns choppy motion tokens into smooth movement","feed_subtitle":"Discrete motion models get a rectified-flow decoder that lifts FID to 0.032 while keeping text alignment.","key_machinery":"Two components carry the method. (1) Condition Projection: each token zt covering q frames is repeated q times, stacked into a vector, linearly projected, and unstacked into q frame-wise conditioning vectors, preserving the temporal correspondence between tokens and frames. (2) Rectified Flow Decoder: a conditional flow-matching model vθ(xt, t, C) trained with the least-squares objective min E||(X1 − X0) − v(Xt, t, C)||², where Xt = tX1 + (1 − t)X0, so that inference is an ODE integration from Gaussian noise to motion, conditioned frame-wise on C. Training on 64-frame sliding windows (rather than full 196-frame sequences) and without attention in the U-Net backbone is what lets the decoder generalize to unseen token sequences in stage 2.","core_discovery":"The central discovery is a decoder-agnostic upgrade path for discrete motion generation. By repeating each discrete token q times and applying a learned linear projection, DisCoRD obtains frame-wise conditioning features that are concatenated channel-wise with the noisy motion in a conditional flow-matching objective dxt = vθ(xt, t, C) dt, learned from sliding windows of 64 frames. At inference, tokens predicted by any pretrained discrete model are projected to C and integrated with Euler steps from Gaussian noise, yielding continuous-motion output. The paper argues this is why discrete methods can finally match continuous ones on naturalness: 0.032 FID on HumanML3D and 0.169 on KIT-ML (best among compared methods), with sJPE reduced by up to 25% in reconstruction; faithfulness holds when the tokenizer uses residual quantization, while a vanilla VQ-VAE (T2M-GPT) shows a small R-Precision drop.","pith_inferences":["If the condition projection is the true bottleneck, a non-linear upsampler (e.g., a small transformer decoder over tokens) is a testable replacement that could remove the small R-Precision losses seen with vanilla VQ-VAEs while keeping the flow decoder's naturalness gains.","The sJPE metric is not limited to motion: any temporally dense generative model (video, audio, facial animation) where FID-like distributional metrics miss per-sample jitter could adopt the symmetric jerk decomposition as a cheap diagnostic.","The combination of discrete token prediction for faithfulness and flow decoding for naturalness suggests a new default architecture for conditional motion generation, where the tokenizer is the main remaining quality lever; pushing codebook size or residual depth may yield further monotone gains.","One concrete test of the paper's causal story: train DisCoRD on tokens whose frame correspondence is artificially scrambled (interleave tokens across time). If FID stays good but sJPE degrades, the method truly relies on the temporal alignment the Condition Projection preserves; if not, the projection's temporal structure matters less than claimed."],"forward_implications":["Replacing the decoder of MoMask with DisCoRD improves reconstruction FID from 0.019 to 0.011 on HumanML3D and generation FID from 0.045 to 0.032, with sJPE dropping about 25%, while R-Precision stays level or improves slightly.","On KIT-ML, the same swap takes generation FID from 0.204 to 0.169 for MoMask and from 0.718 to 0.541 for T2M-GPT, showing the gain transfers to the smaller, noisier dataset.","The method carries over to co-speech gesture (FGD 5.21→4.83 for ProbTalk, 74.88→43.58 for TalkSHOW) and music-to-dance (Distk and Distg move toward the ground-truth spread).","Decoding speed at 16 Euler steps is on par with MoMask's one-step decoder (0.221s vs 0.244s per batch), and with 2 steps DisCoRD is faster while keeping FID 0.034 and sJPE competitive.","Residual-quantization levels are the faithfulness lever: as MoMask's RQ level rises from R0 to R5, DisCoRD's R-Precision gain over the baseline grows from −2.8% to +0.6% at Top-1, indicating richer tokens are decoded more faithfully."],"supporting_citations":[{"why":"MoMask supplies the main discrete baseline and its residual-quantized tokenizer; DisCoRD replaces its decoder in the headline experiments.","marker":"[12]"},{"why":"Rectified flow is the generative transport framework (flow straight and fast) that the DisCoRD decoder is built on.","marker":"[29]"},{"why":"Flow matching provides the least-squares training objective used to learn the conditional vector field.","marker":"[26]"},{"why":"T2M-GPT is the vanilla VQ-VAE discrete baseline that shows the faithfulness trade-off when its decoder is swapped out.","marker":"[64]"},{"why":"RQ-VAE residual quantization underlies the MoMask and BAMM tokens that let DisCoRD preserve faithfulness.","marker":"[24]"},{"why":"HumanML3D is the primary text-to-motion dataset and benchmark where the headline FID numbers are measured.","marker":"[11]"},{"why":"KIT-ML is the second, smaller benchmark dataset where the naturalness gains replicate.","marker":"[42]"},{"why":"The Human Motion Diffusion Model defines the metric protocol (motion encoder for FID and R-Precision) that all compared methods are evaluated under.","marker":"[51]"},{"why":"MLD is the continuous latent diffusion baseline in the comparisons, and its stage-1 FID/MPJPE evaluation practice is what the paper builds on and critiques.","marker":"[60]"}],"fun_headline_variants":["DisCoRD: discrete tokens to smooth motion via rectified flow","Rectified flow decoding smooths discrete motion generation","Discrete tokens to continuous motion via flow decoding","Flow decoder lifts discrete motion models to continuous realism"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned linear projection that expands each token into q frame-wise features must preserve every detail the decoder needs; if it discards token-specific information, no amount of flow-model capacity can restore faithfulness.","fun_headline_variants_meta":{"raw":{"variants":["DisCoRD: discrete tokens to smooth motion via rectified flow","Rectified flow decoding smooths discrete motion generation","Discrete tokens to continuous motion via flow decoding","Flow decoder lifts discrete motion models to continuous realism"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00114,"raw_usage":{"total_tokens":4738,"prompt_tokens":960,"completion_tokens":3778,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":3715}},"tokens_in":576,"tokens_out":3778,"duration_ms":22563,"temperature":1.0,"reasoning_tokens":3715,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:06:39.771976+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure how much token information survives the Condition Projection by training a small linear classifier that predicts the original codebook index from the projected features C on held-out motions. If classification accuracy falls far below the accuracy achievable from the raw codebook embeddings themselves, the projection is the faithfulness bottleneck, and the observed R-Precision drop for T2M-GPT is explained by information loss rather than by anything the flow decoder does.","supporting_citations":[{"cited_title":"Momask: Generative masked modeling of 3d human motions","cited_arxiv_id":null,"evidence_quote":"MoMask supplies the main discrete baseline and its residual-quantized tokenizer; DisCoRD replaces its decoder in the headline experiments."},{"cited_title":"T2m-gpt: Generating human motion from textual de- scriptions with discrete representations, 2023","cited_arxiv_id":null,"evidence_quote":"T2M-GPT is the vanilla VQ-VAE discrete baseline that shows the faithfulness trade-off when its decoder is swapped out."},{"cited_title":"Autoregressive image generation using resid- ual quantization, 2022","cited_arxiv_id":null,"evidence_quote":"RQ-VAE residual quantization underlies the MoMask and BAMM tokens that let DisCoRD preserve faithfulness."},{"cited_title":"Generating diverse and natural 3d human motions from text","cited_arxiv_id":null,"evidence_quote":"HumanML3D is the primary text-to-motion dataset and benchmark where the headline FID numbers are measured."},{"cited_title":"Human motion diffusion model","cited_arxiv_id":null,"evidence_quote":"The Human Motion Diffusion Model defines the metric protocol (motion encoder for FID and R-Precision) that all compared methods are evaluated under."},{"cited_title":"Executing your com- mands via motion diffusion in latent space","cited_arxiv_id":null,"evidence_quote":"MLD is the continuous latent diffusion baseline in the comparisons, and its stage-1 FID/MPJPE evaluation practice is what the paper builds on and critiques."}],"review_version":1}