{"id":"c2542a32-ef1b-401a-818f-ad3bff3a8c9c","arxiv_id":"2505.19377","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Using absolute 3D joint coordinates with a plain Transformer and velocity-prediction diffusion outperforms the standard local-relative motion representation, improving fidelity and enabling direct control.","lead":"A new model for turning text into 3D human motion shows that using simple, absolute joint positions works better than the complex, relative coordinates used by most current systems. This could simplify motion generation and make it easier to control where and how characters move.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central claim is not directly tested: Table 1 compares absolute vs absolute+redundancy, not HumanML3D local-relative vs absolute, and Table 2 confounds representation with recipe, so the SOTA comparison cannot isolate the representation effect.","rationale":"The reader's weakest assumption already points at the missing local-relative baseline with the full optimal recipe, and I agree that is the decisive gap. My refinement is that Table 1 does not merely test the local-relative representation with Concat conditioning; it tests an absolute-plus-redundancy representation, which is a different cell entirely. The paper's own ablation structure therefore never directly contrasts the two representations named in the central claim. The comparison to MARDM-v is a useful sanity check but it is a proxy, since architecture, latent tokenization, conditioning, objective, and scale all differ. The paper has genuine independent strengths: the scaling curves in Figure 4 and Table A1 are systematic, the control results are extensive, and the AR extensions in Table A5 show flexibility. However, those support scalability and generality, not the specific superiority of absolute over local-relative coordinates. I would keep the reader's CONDITIONAL verdict: the claim is plausible and well-engineered, but the decisive controlled experiment should be run before the representation-level conclusion is treated as established. The requested check is a single training run plus dual-protocol evaluation, which is feasible for the authors and would settle the matter.","tokens_in":25628,"tokens_out":5613,"duration_ms":51383,"concrete_test":"Train the exact optimal ACMDM configuration (e.g., ACMDM-S-PS22 or XL-PS2 with AdaLN conditioning, causal latent AE, velocity prediction, same CFG and data splits) on the standard HumanML3D local-relative representation instead of absolute coordinates, converting outputs to absolute evaluation features, and compute FID/R-Precision under both [66] and the original HumanML3D evaluation code. If the local-relative variant is statistically indistinguishable from or better than ACMDM, the representation-specific claim fails; if it is clearly worse, the claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and introduction frame the contribution as absolute global joint coordinates beating the dominant HumanML3D local-relative, kinematic-aware representation. The controlled ablation in Table 1 does not provide that comparison. Its two representation rows are 'Absolute+Redundancy' (InterGen-style absolute coordinates with redundant kinematic features) versus plain absolute coordinates, both under Concat conditioning. It never trains the same ACMDM recipe on the HumanML3D local-relative representation. The headline comparison to MARDM-v in Table 2 also changes many variables at once: representation, architecture, tokenization, latent AE, AdaLN conditioning, velocity prediction, and model scale. Because the paper attributes the gains specifically to the 'radically simplified' absolute representation, the missing cell is load-bearing: if a local-relative model trained with the full optimal recipe (AdaLN + causal AE + velocity prediction) matched or exceeded ACMDM, the central claim would collapse to a recipe/engineering result rather than a representation result. A second confound is the evaluation protocol: all metrics are computed under [66], the authors' own re-evaluation framework, and the paper does not report results under the original HumanML3D protocol, making cross-paper comparisons nonstandard.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ACMDM, a text-to-motion diffusion framework built on absolute global joint coordinates instead of the local-relative, kinematic-aware representation popularized by HumanML3D. The authors argue that this simpler representation, combined with a DiT-style Transformer, velocity prediction, AdaLN conditioning, and an optional causal latent autoencoder, achieves state-of-the-art generation quality, naturally supports trajectory control and spatial/temporal editing, and extends directly to SMPL-H mesh vertex generation. The headline results on HumanML3D are FID 0.058 and R-Precision Top-1 0.522 for ACMDM-XL-PS2, compared with FID 0.114 and R-Precision Top-1 0.500 for MARDM-v, all measured under the evaluation framework of [66].","tokens_in":25846,"tokens_out":3680,"duration_ms":37849,"significance":"If the central claim is fully substantiated, this paper would challenge a widely adopted design choice in text-to-motion generation and would open a simpler path to controllable generation and direct mesh-level synthesis. The paper's strengths include a systematic ablation of conditioning mechanisms, diffusion objectives, autoencoder variants, and model scaling; a clear demonstration of controllable generation without classifier guidance; and exploratory results on direct SMPL-H mesh generation, KIT-ML, and autoregressive diffusion variants. These are useful empirical contributions. However, the manuscript currently does not directly test the representation claim against the dominant HumanML3D local-relative representation under a matched recipe, so the significance of the headline attribution remains conditional on an additional controlled experiment.","major_comments":[{"comment":"The central claim that absolute coordinates outperform the HumanML3D local-relative representation is not directly tested. The controlled rows in Table 1 compare 'Absolute+Redundancy' (an InterGen-style absolute representation with redundant kinematic features) against plain absolute coordinates, both under concatenation conditioning. The local-relative, kinematic-aware HumanML3D encoding is never trained with the same ACMDM recipe, and the later rows that add AdaLN, latent AE, and velocity prediction are run only for the absolute representation. Because the abstract and introduction attribute the gains specifically to abandoning the HumanML3D local-relative formulation, this missing cell is load-bearing. Please add a matched experiment that trains the same ACMDM model (ideally with the full optimal recipe: AdaLN + Causal AE + velocity prediction) on the HumanML3D representation, or revise the claim to be explicitly about absolute coordinates versus absolute coordinates with redundancy.","section":"§4.2, Table 1"},{"comment":"The state-of-the-art comparison confounds the representation change with changes in architecture, tokenization, latent autoencoder, conditioning mechanism, diffusion objective, and model scale. The comparison of ACMDM-XL-PS2 (FID 0.058, R-Precision Top-1 0.522) against MARDM-v (FID 0.114, R-Precision Top-1 0.500) therefore cannot establish that absolute coordinates -- rather than the overall recipe -- are responsible for the improvement. A matched comparison is needed: either apply the absolute-coordinate representation within the MARDM recipe, or train the ACMDM recipe on the HumanML3D local-relative representation. Without such an experiment, the claim that 'absolute coordinates make motion generation easy' is not causally supported.","section":"§4.2 and Table 2"},{"comment":"All metrics are computed under the re-evaluation framework of [66], which is the authors' own prior work, and the paper does not report results under the original HumanML3D evaluation protocol used in most published baselines. Since metric details (including the choice of motion features and how they are extracted) directly affect FID and R-Precision numbers, the lack of the original-protocol numbers makes cross-paper comparisons nonstandard and difficult for readers to relate to the existing literature. Please report results under the original HumanML3D protocol as well, or provide a clear calibration between the two protocols.","section":"§4.1, Evaluation Metrics"},{"comment":"The controllable generation comparison also mixes multiple variables: the baseline methods use different architectures, different motion representations, and some rely on classifier guidance, while ACMDM adds a ControlNet branch to the main model. The conclusion that absolute coordinates 'naturally support' control without guidance is plausible, but the current evidence does not isolate the representation effect. Please include an ablation that uses the same ControlNet-based framework with a local-relative representation (or with the HumanML3D representation) to show that the representational change, rather than the ControlNet design, is responsible for the improved control accuracy and the elimination of classifier guidance.","section":"§4.4, Table 3"}],"minor_comments":[{"comment":"In the upper-body editing section, OmniControl is cited as [120] and MotionLCM V2+CtrlNet as [120], but OmniControl is [103] and MotionLCM V2 is [14]; please correct these reference labels.","section":"Table 3 and Table A2"},{"comment":"The text states that all mesh-generation variants achieve results 'competitive with the best text-to-joint generation models,' but Table 4 contains no direct mesh-generation baselines and no joint-level baseline under the same evaluation features; please add a comparative baseline or qualify the claim.","section":"§4.5, Table 4"},{"comment":"The third contribution bullet contains a grammatical issue: 'promising generalizes beyond joints' should be 'promising generalization beyond joints' or 'promisingly generalizes beyond joints.'","section":"Introduction, Contributions"},{"comment":"The table caption says 'The results indicate that kinematic-aware redundancy is not necessary,' but the table only compares absolute representations with and without redundancy; please rephrase to match the actual comparison, e.g., 'redundant kinematic features are not necessary for absolute-coordinate diffusion.'","section":"§4.2, Table 1 caption"},{"comment":"The caption text 'MARDMMotionLCM V2MDM' appears to be a formatting artifact from concatenated method names; please separate the method names clearly.","section":"Figure 3 caption"}],"recommendation":"major_revision","confidential_remarks":"This is a strong empirical paper with a clear message, but the central attribution is not yet isolated. The missing experiment -- training the full ACMDM recipe on the HumanML3D local-relative representation -- is straightforward and should be requested. The reliance on the authors' own evaluation framework [66] is not circular, but the paper should be transparent about protocol differences and ideally report both protocols. The mesh-generation section is exploratory and should be framed as such rather than as a competitive result without baselines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper is worth reading and worth refereeing, but the abstract sells a comparison the experiments don't actually run. The key result that does hold up is Table 1's controlled ablation: under identical Concat conditioning and the same v-prediction objective, plain absolute coordinates beat the same model with InterGen-style redundant kinematic features (FID 0.208 vs 0.276). That is a genuine, non-obvious finding, and the scaling curves in Figure 4 are clean and consistent. The control and mesh extensions are substantively interesting, and the authors are honest that mesh results are preliminary.\n\nThe soft spot is the one everyone will point to: the paper claims absolute coordinates outperform the HumanML3D local-relative representation, but it never trains the local-relative representation with the full optimal recipe (AdaLN + causal AE + v prediction). Table 1 only compares absolute vs absolute+redundancy. The SOTA comparison in Table 2 changes representation, architecture, tokenization, latent AE, conditioning, and model scale at once. So the central attribution of the gains to the representation is not actually isolated. It is possible that a local-relative model with the same recipe would match or beat ACMDM, which would reduce this to a recipe paper rather than a representation result. The second issue is the evaluation framework: metrics are computed under [66], the same group's re-evaluation, and results on the original HumanML3D protocol are not reported, so cross-paper comparisons with the rest of the literature are not apples-to-apples. Code is also not released, which makes the missing cell harder to fill externally.\n\nIn proportion: the paper is still a good piece of empirical work. The controlled ablation is the honest core, and it supports the directional claim that absolute coordinates are not the liability people assumed. The overclaim is in the framing, not in the data. If the authors add the missing local-relative baseline under the full recipe, report numbers under the standard protocol, and release code, this becomes a strong contribution.\n\nFor peer review: send it out, but ask for the missing cell. A serious referee would have a clear, actionable request. This is exactly the kind of paper that should see the light of day after revision, because the question matters.","headline":"A well-run empirical study that makes a plausible case for absolute coordinates, but the headline claim overreaches: the decisive comparison—local-relative trained with the full recipe—is missing.","tokens_in":26404,"tokens_out":2279,"would_cite":true,"duration_ms":23735,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Absolute joint coordinates in global space, trained with velocity prediction and adaptive normalization in a simple Transformer, are sufficient to beat the local-relative motion representation on text-to-motion generation.","keywords":["text-to-motion generation","motion representation","absolute joint coordinates","diffusion models","human motion synthesis","motion control","SMPL-H mesh generation","Transformer"],"falsifier":"Train the same Transformer architecture, with the same AdaLN conditioning, velocity prediction, causal latent autoencoder, and identical model sizes, on the standard local-relative representation; evaluate both models with the original benchmark's official evaluation code. If the local-relative model matches or exceeds the absolute-coordinate model, the central claim is not supported.","tokens_in":25386,"feed_emoji":"🕺","tokens_out":8852,"duration_ms":76671,"temperature":0.7,"pith_summary":"Text-to-motion generation has converged on a local-relative, kinematic-aware encoding that stores joint positions relative to the pelvis and previous frame, with extra features such as foot contacts. This paper argues that the redundancy is not just unnecessary but harmful for diffusion models, and that the long-abandoned alternative of absolute joint coordinates in global space is simpler and stronger. Using a plain Transformer with velocity prediction and adaptive layer normalization conditioning, without kinematic-aware auxiliary losses, the proposed model reaches state-of-the-art fidelity and text alignment, and its smallest variant already beats prior systems. The same formulation also makes trajectory control and spatial editing straightforward, without classifier guidance from control signals or input optimization, and extends naturally to generating SMPL-H mesh vertices directly from text. If correct, the paper shifts the field's default representation from engineered kinematic features back to raw coordinates.","feed_headline":"Plain 3D joint coordinates beat the standard motion encoding","feed_subtitle":"With only absolute coordinates, a simple Transformer reaches state-of-the-art fidelity and makes control and mesh generation direct.","key_machinery":"The workhorse is the absolute-coordinate motion representation: per frame, joint positions $X_i \\in \\mathbb{R}^{N_j \\times 3}$ in global space, with no pelvis-relative or previous-frame encoding and no redundant kinematic features. The model tokenizes this 2D motion field with a convolutional patchifier that preserves temporal resolution, feeds the tokens into a Transformer with rotary position embeddings, QK normalization, SwiGLU, and adaptive layer normalization (AdaLN) for text and timestep conditioning, and trains with the flow-matching velocity objective $v$ instead of predicting $x_0$ or noise. An optional causal autoencoder compresses motion into a latent space that improves fidelity. For control, a ControlNet branch takes the same noised absolute inputs plus control signals and injects additive residuals into the frozen main backbone; for meshes, a fully convolutional mesh autoencoder compresses SMPL-H vertices to 28 latent channels and the same Transformer operates on those latents.","core_discovery":"The paper's central claim is that the dominant local-relative motion representation—which encodes intra-frame kinematics and inter-frame transitions around the pelvis and includes redundant non-animatable features—is not what makes modern text-to-motion diffusion models work. With absolute 3D joint coordinates as the only motion signal, a simple Transformer trained with the velocity-prediction objective, adaptive layer normalization conditioning, and a causal latent motion autoencoder achieves higher motion fidelity and better text alignment than existing systems, and it scales well as model size grows and token patch size shrinks. The paper further claims that this representation is the reason control and editing become straightforward: because outputs are already in global space, user-provided absolute trajectories or partial-body constraints can be injected directly through a frozen-backbone ControlNet without classifier guidance from control signals or inference-time optimization. Finally, the same model, with a mesh autoencoder for spatial compression, can generate SMPL-H mesh vertices from text directly, bypassing joint-to-mesh fitting.","pith_inferences":["Editorial inference: if absolute coordinates remain competitive under the original benchmark's evaluation pipeline, the local-relative representation's remaining justification is historical, and new text-to-motion work should default to global coordinates.","Editorial inference: the same recipe could transfer to other structured-sequence generation tasks whose outputs are point sets or trajectories in a fixed global frame, such as robot motion or multi-agent coordination.","Editorial inference: because absolute coordinates expose global position natively, physics constraints, foot-contact regularization, or scene collision terms could be added directly on generated output, which the redundant representation makes awkward."],"forward_implications":["Absolute coordinates remove the need for relative-to-absolute post-processing, so trajectory control and spatial editing can be done by directly feeding user-specified joint positions as conditions.","Because control signals are injected as additive residuals into a frozen backbone, generation runs without classifier guidance from control signals, reducing control inference time from roughly 81 seconds to about 2.5 seconds per sample.","The same framework generalizes to direct SMPL-H mesh vertex generation, so text-to-motion can skip SMPL fitting and the errors it accumulates.","Scaling model depth, width, and spatial patch granularity consistently improves quality, indicating that absolute coordinates give a representation that benefits from larger models.","The formulation also works in autoregressive diffusion variants, so the benefit is not tied to a single diffusion scheduling or backbone choice."],"supporting_citations":[{"why":"introduces the local-relative representation the paper challenges and the benchmark used for all comparisons.","marker":"[25]"},{"why":"provides the evaluation framework and the previous state-of-the-art baselines that set the comparison numbers.","marker":"[66]"},{"why":"represents the prior approach that argued absolute coordinates produce unnatural motion and relied on classifier guidance.","marker":"[103]"},{"why":"supplies the velocity-prediction / flow-matching objective found to be the best diffusion target.","marker":"[57]"},{"why":"provides the Transformer design with adaptive layer normalization conditioning that the paper adopts.","marker":"[69]"},{"why":"provides the ControlNet module used for controllable motion generation.","marker":"[119]"},{"why":"provides the fully convolutional mesh autoencoder used for spatial compression of SMPL-H vertices.","marker":"[131]"},{"why":"is the diffusion baseline and an early adopter of the local-relative representation.","marker":"[92]"}],"fun_headline_variants":["Absolute coordinates beat the standard motion encoding","Text-to-motion: absolute coordinates outperform relative","Absolute coordinates simplify and improve motion generation","Ditch relative motion: absolute coordinates work better"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"If the evaluation protocol used to compare methods is not a fair stand-in for the original benchmark, or if the comparison local-relative model is trained with a weaker recipe, the claim that absolute coordinates cause the improvement would collapse.","fun_headline_variants_meta":{"raw":{"variants":["Absolute coordinates beat the standard motion encoding","Text-to-motion: absolute coordinates outperform relative","Absolute coordinates simplify and improve motion generation","Ditch relative motion: absolute coordinates work better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000981,"raw_usage":{"total_tokens":4149,"prompt_tokens":914,"completion_tokens":3235,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":3180}},"tokens_in":530,"tokens_out":3235,"duration_ms":25938,"temperature":1.0,"reasoning_tokens":3180,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:14:48.227914+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same Transformer architecture, with the same AdaLN conditioning, velocity prediction, causal latent autoencoder, and identical model sizes, on the standard local-relative representation; evaluate both models with the original benchmark's official evaluation code. If the local-relative model matches or exceeds the absolute-coordinate model, the central claim is not supported.","supporting_citations":[{"cited_title":"Fully convolutional mesh autoencoder using efficient spatially varying kernels.Ad- vances in neural information processing systems, 33:9251–9262, 2020","cited_arxiv_id":null,"evidence_quote":"provides the fully convolutional mesh autoencoder used for spatial compression of SMPL-H vertices."}],"review_version":1}