{"id":"61b41ffa-14a3-4ff2-abe2-745da4e5952f","arxiv_id":"2504.16722","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ProMoGen generates human motion conditioned on both a trajectory and sparse anchor postures via a diffusion transformer trained with a dense-to-sparse curriculum.","lead":"This paper introduces a diffusion model that generates human motion from two controls at once: a global trajectory and a small number of user-chosen key poses. The authors add a curriculum that trains first with many key poses and then with progressively fewer, which they report improves stability and accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on author-adapted baselines whose text encoders were replaced with two linear layers; if that adapter handicaps the baselines, the Table 1 outperformance is not established.","rationale":"The paper's core contribution is a controllable motion synthesis system; the only quantitative evidence for 'significantly outperforming state-of-the-art' is Table 1. The comparison's validity hinges entirely on the adapted-baseline protocol in §4. The reader flagged exactly this in weakest_assumption, and my reading confirms it is the least secure premise. I also noticed two additional issues that reinforced the concern but did not replace it: (i) Table 3's 'w/. Reg Learning' row is numerically identical to Ours-v2 in Table 1, suggesting the headline comparison may not even include SAP-CL, the paper's main curriculum contribution; (ii) the Filtering Module write-up in §3.2.1 overwrites δ_{f_n-1} with two incompatible definitions, an off-by-one sign that the text-level algorithm is not self-consistent. Neither issue is as directly load-bearing as baseline fairness, because the curriculum omission would only strengthen the reported superiority and the FM typo may be confined to the write-up. The baseline adaptation, by contrast, determines whether the central outperformance claim is real. The paper does not report how many anchors the baselines were trained with, whether they were trained under the same curriculum, or whether the same anchor-loss supervision was used. Without this, Table 1 could be showing a comparison between a fully-supervised task-specific model and under-conditioned repurposed text-to-motion models. A single reconstruction experiment with the released code would settle this, so the current CONDITIONAL verdict is appropriate and no adjustment is needed.","tokens_in":14946,"tokens_out":10355,"duration_ms":94332,"concrete_test":"Use the released repo to retrain the MDM baseline under exactly the conditions used for Ours-v2: same f_n/f_s sampling ranges (including curriculum stages if Ours-v2 uses SAP-CL), same loss with L_anchor, same epochs/optimizer, and the two-linear-layer adapter. Compare its Table 1 MPJPE/FID (reported 5.510 / 1.093 on DS-1) to the retrained value. If the retrained MPJPE is below ~4.5 or FID below ~0.7, the reported baseline is under-conditioned and the SOTA claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is in §4 (Implement Details): baselines are 'reconstructed' by removing their text encoders and substituting 'two linear layers that serve as trajectory and sparse posture encoders.' For Table 1 to support 'significantly outperforming state-of-the-art,' this adapter must not be a bottleneck, and every other variable must be fixed. The paper reports no baseline hyperparameters, no number of anchor frames used to train baselines, no f_s/f_n sampling distribution, and no loss terms. ProMoGen is trained with a dedicated Anchor Motion Loss (λ2=1) and SAP-CL; if the reconstructed baselines are trained with plain denoising loss and a different anchor-sampling protocol, then the large MPJPE and K-MPJPE gaps (e.g., 3.257 vs 5.452 for CAMDM on DS-1) may reflect loss design or training distribution rather than generative superiority. The description of Ours-v1 as 'this same variant' is ambiguous: if Ours-v1 is ProMoGen's architecture with linear encoders, then it is not an adapted baseline, and the table does not isolate adapter quality for prior methods. Because the code has no commit hash and the appendix is missing, this protocol cannot be checked from the preprint.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ProMoGen, a diffusion-transformer-based framework for human motion generation conditioned jointly on a global trajectory and sparse anchor postures, together with SAP-CL, a curriculum learning strategy that progressively reduces the number of anchor poses during training. The authors report large gains over existing methods on HumanML3D and CombatMotion across MPJPE, K-MPJPE, FID, and related metrics, and they provide ablations of the main architectural components and of the curriculum strategy. The central claim is that a single model can combine trajectory control with user-specified sparse pose constraints and outperform prior trajectory-, keyframe-, and text-conditioned motion generators.","tokens_in":15193,"tokens_out":7841,"duration_ms":67957,"significance":"If the reported results are reproducible, ProMoGen would be a practically useful contribution: it addresses a real gap by allowing simultaneous trajectory and sparse-pose control, and the SAP-CL idea is a sensible response to the instability of training from very sparse anchors. The paper is also honest about its task formulation and evaluates on two datasets. However, the significance is currently limited by verification gaps: the state-of-the-art comparison rests on baselines that the authors re-implemented with text encoders replaced by two linear layers, and the paper does not supply code, official checkpoints, or a complete appendix. The quantitative story is therefore plausible but not yet established to the standard the abstract claims.","major_comments":[{"comment":"The headline claim of outperforming state-of-the-art methods is not established by Table 1 because all baselines are \"reconstructed\" by replacing their text encoders with two linear layers, and the reconstruction protocol is underspecified. The paper does not report the baselines' training losses, anchor-sampling distributions, curriculum or regularization hyperparameters, or whether the same Filtering Module and anchor set were used for every method. If the linear adapters are a bottleneck for methods whose architectures assume text conditioning, the comparison is biased toward ProMoGen. The description of Ours-v1 is also ambiguous: if it is ProMoGen with linear encoders rather than an adapted baseline, then Table 1 does not demonstrate that the adapted baselines were given their best possible configuration. Please clarify the experimental protocol, provide per-baseline settings, or re-run the comparison against official checkpoints/code.","section":"§4 (Implement Details)"},{"comment":"The description of the Initial Motion Generator is internally inconsistent. The text states that G \"synthesizes coarse motion priors conditioned solely on trajectory features,\" but the equation defining it is M_init = G(x_t, E_k(X_s)), which uses the anchor motion encoder and gives no trajectory input. Please correct either the text or the equation; this ambiguity makes the architecture non-reproducible.","section":"§3.2.2"},{"comment":"The anchor-mapping equations contain indexing errors. The constraint is written as sum_{i=0}^{f_n} δ_i = R, but only f_n deltas (δ_0 to δ_{f_n-1}) are defined, and their actual sum is T_total - f_n = R. Moreover, the recurrence x_j = x_{j-1} + f_s + δ_{j-1} + 1 implies x_j = p_j + (j-1) f_s, so adjacent anchors are separated by f_s + δ_{j-1} + 1 ≥ f_s + 1; the text says f_s is the \"minimum interval,\" which is off by one. Please verify the formulas and restate them precisely.","section":"§3.2.1"},{"comment":"The conclusion promises that \"More details and results are shown in the Appendix,\" but no appendix is present in the submitted manuscript. Important experimental details—dataset splits, baseline reconstruction settings, evaluation protocol, and the exact relation between \"Regular\" and \"Curriculum\" training in Tables 2 and 3—are therefore missing. The GitHub link is given without a commit hash or released code, so the experiments cannot be checked.","section":"§5 (Conclusion) and Experiments"},{"comment":"The ablation results are not fully consistent with the narrative. Removing the physical-constraint loss gives MPJPE 3.251 and K-MPJPE 3.221, which are slightly better than the full model's 3.257 and 3.247, while the text states that \"GAN and Phys Loss can further improve the precision.\" Also, Table 3's \"w/. Reg Learning\" and \"w/. Cur Learning\" rows do not match the corresponding rows in Table 2, so the reader cannot tell which anchor count and training schedule these numbers refer to. Please clarify these inconsistencies.","section":"Table 3"}],"minor_comments":[{"comment":"The abstract contains a comma splice (\"inherently unstable, we introduce\"); please rewrite the sentence.","section":"Abstract"},{"comment":"Reference [56] is a GitHub repository rather than a peer-reviewed dataset; please provide a stable citation and describe the train/test split, retargeting procedure, and preprocessing for CombatMotion.","section":"References / Datasets"},{"comment":"The table reports \"Diversity→ closer to the value of Real is better,\" but the real reference values are not reported; please add them.","section":"Table 1"},{"comment":"Several entries in Table 3 are missing the '±' separator (e.g., \"3.5310.066\" and \"3.3260.054\").","section":"Table 3"},{"comment":"The paper uses \"Dit\" and \"DiT\" inconsistently; please choose one spelling and use it throughout.","section":"Notation"},{"comment":"DPM-Solver++ is mentioned with an update rule but is neither cited nor defined; please add the appropriate reference and explain how the coefficients are obtained.","section":"Diffusion solver"},{"comment":"No statistical significance tests are reported; given the small standard deviations in Table 1, please provide p-values or confidence intervals for the key comparisons.","section":"Statistics"},{"comment":"There are several typos, including \"anchor posetures,\" \"On the on hand,\" and \"oberseved\"; please proofread the manuscript.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the adapted-baseline protocol. If the authors can provide the exact reconstruction details, release code with a commit hash, and supply the missing appendix, the empirical claim may become convincing; without those, the SOTA comparison should be substantially toned down. I would not reject outright because the task formulation and curriculum idea are sound and the errors in the text appear fixable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a plausible engineering contribution with a nice curriculum idea, but the headline 'significantly outperforming SOTA' is not supported by the evidence as written. The missing appendix and author-adapted baselines make the main comparison uncheckable. If the authors release code and fix the protocol, it's worth a look.\n\nWhat's genuinely new: the combination of trajectory conditioning, arbitrary sparse anchor postures via the Filtering Module, and the dense-to-sparse curriculum. The Filtering Module's sampling algorithm is clean and well specified. The decoupling of global trajectory and local pose is a sensible design. The ablations are internally consistent and show the curriculum helps their own model. That part is credible.\n\nSoft spots: the Table 1 comparison is the load-bearing claim. The baselines are 'reconstructed' by stripping text encoders and replacing them with two linear layers. That is not a faithful adaptation unless the rest of the training protocol (anchor sampling, loss terms, schedule) is identical. The paper doesn't say. Ours-v1 is described as 'this same variant'—if Ours-v1 is just ProMoGen with linear encoders, it's not a baseline reconstruction, and the table doesn't isolate whether the adapter handicaps prior methods. The appendix is promised but missing; the code repo has no commit hash. Figure 3's caption says upper right is better while both axes are lower-is-better; that's a sign of carelessness, not necessarily a wrong result. The missing significance tests are minor for this field, but the missing protocol details are not.\n\nThe core idea is sound and the ablations support it. The problem is not the method, it's the evidence for 'superior to SOTA.' That can be fixed with code release and a fairer baseline protocol.\n\nThis paper is for people working on controllable motion synthesis. A serious referee should engage with it, but only if the protocol is clarified. I'd want to see the appendix and code before trusting the numbers.","headline":"Plausible engineering contribution with a clean curriculum idea, but the SOTA claim rests on uncheckable baseline adaptations and a missing appendix.","tokens_in":15697,"tokens_out":1330,"would_cite":false,"duration_ms":12449,"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":"ProMoGen claims that one diffusion model can obey both a global trajectory and a handful of sparse anchor postures, and that this dual control beats single-condition motion generators.","keywords":["human motion synthesis","trajectory conditioning","sparse keyframe control","diffusion models","curriculum learning","motion in-betweening","controllable generation","animation"],"falsifier":"Run the same trajectory-and-anchor test set through a strong text-conditioned motion model using text prompts that describe the actions at the anchor poses, and compare its MPJPE and FID with ProMoGen's; if it matches or beats them, the claimed superiority of sparse-anchor control is not established.","tokens_in":14734,"feed_emoji":"🕺","tokens_out":5241,"duration_ms":50275,"temperature":0.7,"pith_summary":"The paper argues that the most useful way to control generated human motion is to combine a global trajectory, which sets where the character goes, with sparse anchor postures, which set what the body looks like at chosen moments. It claims these two signals can be cleanly decoupled, and that a single diffusion model can be trained on both at once. To make training stable when anchors are very sparse, it introduces a curriculum that starts with many anchor frames and gradually reduces their number. On HumanML3D and CombatMotion, the system is reported to beat existing text-, trajectory-, and keyframe-based methods on pose error, fidelity, and anchor faithfulness. If correct, the contribution is a unified control interface: a user specifies a path and a few poses and gets smooth, customized motion.","feed_headline":"Trajectory plus a few key poses drives a diffusion motion model","feed_subtitle":"ProMoGen's curriculum training starts dense and goes sparse, beating prior text- and trajectory-driven baselines.","key_machinery":"The load-bearing mechanism is the Sparse Anchor Posture Curriculum Learning (SAP-CL) schedule paired with a two-stream diffusion-transformer architecture. SAP-CL splits training into stages with anchor-count interval $[K^{(s)}_{\\min}, K_{\\max}]$, where $K_{\\max}=30$; the minimum starts at 20 and linearly decays to 1, and each stage samples temporal density $f_n^k$ and interval elasticity $f_s^k$ uniformly. This progressive move from dense to sparse supervision is what avoids unstable gradients from extremely sparse signals and lets the model generalize to arbitrary anchor placements at inference. The Filtering Module implements the placement by selecting virtual points, distributing elastic interval increments $\\delta_i$, and mapping them through $x_j = j+(j-1)f_s + \\sum_{i=0}^{j-1}\\delta_i$, which guarantees the minimum spacing $x_k-x_{k-1}\\ge f_s+1$ and makes user-specified anchor positions reproducible during training.","core_discovery":"The central claim is that motion synthesis is best driven by two decoupled conditions: a global trajectory $\\tau\\in\\mathbb{R}^{N\\times 3}$ and sparse anchor postures $X_s\\in\\mathbb{R}^{M\\times 6}$, and that this pairing yields more faithful motion than any single-condition baseline. The network assigns each condition its own transformer encoder, builds a coarse motion from the trajectory alone, then refines it by fusing both feature streams. A Filtering Module samples anchor frames under user-given density and minimum-interval constraints, while SAP-CL schedules the anchor count from 20 down to 1 during training. Reported numbers include MPJPE 3.257 on HumanML3D and 3.654 on CombatMotion, with FID 0.279 and 0.412, each best among the listed baselines; the curriculum-trained variant further lowers MPJPE to 3.047 at five anchors. The paper's claim is that sparse anchors plus trajectory is not merely a harder in-betweening problem but a controllable generation task, and that the easy-to-hard curriculum is what makes learning it stable.","pith_inferences":["The decoupling assumption is testable beyond the paper: if trajectory and local pose are truly independent, adapting the model to a new skeleton or motion style should require retraining only the anchor encoder, not the whole network.","The same framework could ingest anchor postures detected from images or video, since the Filtering Module already imitates arbitrary user placement by uniform sampling over valid anchor sets.","A fairer head-to-head would also run text-conditioned baselines prompted with descriptions of the anchor actions, because the paper's comparison strips text encoders from prior models and replaces them with linear layers."],"forward_implications":["A single unified model can serve both trajectory-only and trajectory-plus-anchor control, since dual and single control share one training process.","The reported anchor-position error K-MPJPE is below the global MPJPE on both datasets, meaning the model is especially faithful at the poses the user actually specified.","Using only a linear encoder for both conditions already beats all listed baselines, implying much of the gain comes from the decoupled conditioning design rather than the elaborate modules.","The curriculum evaluation shows that at the same inference-time anchor count, a model trained with the progressive schedule performs better, shifting the accuracy-versus-sparsity tradeoff in the user's favor."],"supporting_citations":[{"why":"supplies the direct-motion-prediction diffusion design the authors adopt and anchors the MDM baseline in comparisons.","marker":"[45]"},{"why":"establishes the curriculum-learning principle that SAP-CL adapts into a pose-density schedule.","marker":"[3]"},{"why":"provides the HumanML3D dataset and evaluation protocol used for the primary comparisons.","marker":"[16]"},{"why":"supplies the CombatMotion dataset used for the second evaluation split.","marker":"[56]"},{"why":"defines the DDPM forward and reverse process underlying the diffusion objective.","marker":"[19]"},{"why":"gives the Diffusion Transformer backbone used across ProMoGen's encoders and refinement modules.","marker":"[37]"},{"why":"is the prior anchor-conditioned motion-diffusion approach ProMoGen extends by decoupling trajectory and pose encoders.","marker":"[25]"}],"fun_headline_variants":["Sparse anchors plus trajectory: a decoupled recipe for motion synthesis","Curriculum learning from dense to sparse anchors stabilizes motion generation","ProMoGen: decoupled trajectory and anchor control for human motion synthesis","Sparse-anchor curriculum beats single-condition motion baselines","From 20 anchors to 1: progressive curriculum boosts motion generation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that ProMoGen beats state-of-the-art methods rests on the assumption that replacing the text encoders of existing models with two linear layers leaves those baselines unimpaired; if that swap handicaps them, the comparison is not meaningful.","fun_headline_variants_meta":{"raw":{"variants":["Sparse anchors plus trajectory: a decoupled recipe for motion synthesis","Curriculum learning from dense to sparse anchors stabilizes motion generation","ProMoGen: decoupled trajectory and anchor control for human motion synthesis","Sparse-anchor curriculum beats single-condition motion baselines","From 20 anchors to 1: progressive curriculum boosts motion generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1415,"prompt_tokens":1025,"completion_tokens":390,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":302}},"tokens_in":641,"tokens_out":390,"duration_ms":3618,"temperature":1.0,"reasoning_tokens":302,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:57:14.063197+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same trajectory-and-anchor test set through a strong text-conditioned motion model using text prompts that describe the actions at the anchor poses, and compare its MPJPE and FID with ProMoGen's; if it matches or beats them, the claimed superiority of sparse-anchor control is not established.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the direct-motion-prediction diffusion design the authors adopt and anchors the MDM baseline in comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the CombatMotion dataset used for the second evaluation split."}],"review_version":1}