{"id":"0e9e84f6-0c14-483e-b4f7-fab4571fa68f","arxiv_id":"2502.02358","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MotionLab unifies text-based and trajectory-based motion generation with text-based editing, trajectory-based editing, motion in-betweening, and style transfer in one flow-based transformer.","lead":"The paper introduces MotionLab, a single model that handles both human motion generation and editing across six tasks, including text-based, trajectory-based, and style-based controls. It shows the model can match or beat separate specialist models on several benchmarks, suggesting that unified motion models are practical and efficient.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"In-between Table 8 contradicts the claim of outperforming CondMDI on all settings: 1-frame FID 0.755 vs 0.155.","rationale":"The reader's verdict is CONDITIONAL, and the reader flagged the Table 8 contradiction in the rationale as a red flag. However, the reader's stated weakest_assumption was the curriculum task-ordering assumption, whereas I find the internal contradiction in the supplementary in-between results to be the most load-bearing concern: it is a concrete, checkable inconsistency that directly bears on the paper's central claim of universal superiority. The curriculum-ordering issue is real but secondary: the ablations do not test alternative orderings, yet the paper attributes much of the gain to the chosen order. Still, the in-between FID discrepancy is a factual matter that can be settled by reproduction and, if reproduced, requires correcting the overclaim. Credit is due for the thorough ablations, the released code, and the promising results on several benchmarks; those are not disputed. But the reported 1-frame FID being 0.755 versus CondMDI's 0.155 means the abstract and Section 11 overstate the results. Since the discrepancy is fixable via correction or explanation, the existing CONDITIONAL verdict remains appropriate; my concern does not change the verdict but reinforces it.","tokens_in":18725,"tokens_out":3855,"duration_ms":38321,"concrete_test":"Run the released MotionLab checkpoint and the official CondMDI checkpoint through the same HumanML3D in-between evaluation protocol with 1 keyframe, using the same feature extractor, sample count, random seed, and CFG settings. If the numbers reproduce Table 8 (MotionLab FID ~0.755, CondMDI ~0.155), then the sentence 'outperforms CondMDI on all settings' is false and the central claim must be revised. Also verify whether the 1-frame FID was computed with the same number of generated samples and whether classifier-free guidance was applied for both models; if the protocols differ, recompute under identical conditions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that MotionLab achieves superior versatility, performance, and efficiency across all unified tasks, and Section 11 explicitly states that 'our framework outperforms CondMDI on all settings.' This is directly contradicted by the paper's own Table 8: for the 1-frame motion in-between setting, MotionLab's FID is 0.7547 versus CondMDI's 0.1551 (roughly 4.9x worse), R-precision is 0.6681 vs 0.6787, and Diversity is 8.9058 vs 9.5807 (ground truth 9.503). The main-text Figure 1 reports only average error for in-between, not FID, so the headline visualization masks this failure. This is not a matter of consensus or taste; it is an internal inconsistency in the reported evidence. If the 1-frame FID is correct, then the unified model is substantially worse than a specialist on a key setting, undermining the claim that MotionLab outperforms task-specific specialists across tasks. The concern is load-bearing because the paper's headline contribution rests on empirical superiority, and this contradiction means the empirical support is incomplete or misreported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Motion-Condition-Motion paradigm, in which every human motion generation or editing task is expressed as a mapping from source motion to target motion under a condition. Based on this paradigm, the authors introduce MotionLab, a unified framework built on rectified flows and a multimodal transformer (MotionFlow Transformer) with Aligned Rotational Position Encoding, task instruction modulation, and a curriculum learning schedule. The central claim is that this single model matches or surpasses task-specific specialists and prior state-of-the-art methods across text-based generation/editing, trajectory-based generation/editing, motion in-betweening, and style transfer, while also being efficient at inference. The main-text quantitative results (Tables 3–5) support several of these claims, and the supplementary provides additional comparisons for in-betweening and style transfer. However, the supplementary in-between table directly contradicts the paper's statement that MotionLab outperforms CondMDI on all settings, and the lack of variance estimates leaves some comparisons uncertain.","tokens_in":18906,"tokens_out":5515,"duration_ms":56832,"significance":"If the empirical results can be made internally consistent, the paper makes a valuable contribution to unified human motion modeling. The Motion-Condition-Motion formulation is conceptually clean and potentially scalable, and the proposed architectural components—especially Aligned ROPE and task instruction modulation—are well motivated. The paper also provides extensive ablations (Tables 6, 9–12) and reports inference-time and memory costs, which are welcome. The main weakness is that the evidence for the headline claim is not yet consistent: the supplementary Table 8 shows that MotionLab's 1-frame in-between FID is about 4.9 times worse than CondMDI's, and several comparisons lack multiple-seed statistics. These issues must be resolved before the claim of universal superiority can be accepted.","major_comments":[{"comment":"The text in Section 11 states that \"our framework outperforms CondMDI on all settings,\" but Table 8 shows the opposite for the 1-frame in-between setting: MotionLab's FID is 0.7547 versus CondMDI's 0.1551, R-precision top-3 is 0.6681 versus 0.6787, and Diversity is 8.9058 versus 9.5807 (ground-truth Diversity is 9.503). This is an internal contradiction in the reported evidence, and it is load-bearing because the paper's headline is that MotionLab surpasses specialists on all evaluated tasks. The authors should either correct the numbers, replace the metric with one for which the comparison is favorable, or explicitly qualify the claim to exclude or explain the 1-frame case. As written, the empirical support for the central claim is incomplete.","section":"Supplementary Material, Section 11, Table 8"},{"comment":"The success of Motion Curriculum Learning rests on the easy-to-hard ordering specified in Section 5.3, but no alternative ordering is tested. The ablations in Table 11 remove masked pre-training, remove ordering, or use random selection, but they do not reverse or permute the stages. The reader cannot tell whether the improvement comes from the specific ordering (e.g., spatial conditions before text and style) or from the multi-task training schedule itself. Please add an ablation with at least one reversed order (for example, text-based generation first, or style transfer before trajectory-based editing) to demonstrate that the claimed transfer of spatial knowledge depends on the proposed order rather than on the mere fact of multi-task fine-tuning.","section":"Section 5.3 and Section 12, Table 11"},{"comment":"No standard deviations or multiple seeds are reported for any quantitative result. The sampling procedure is stochastic (CFG with learned velocity fields), and some comparisons are close (e.g., Table 3, R@3: MotionLab 0.810 vs MotionCLR 0.831; Diversity 9.593 vs 9.607). Without multiple runs, it is unclear whether the reported differences are meaningful. Please report mean and standard deviation over at least three seeds for the central metrics, or state explicitly that the evaluation protocol is deterministic and explain why.","section":"Section 6, Tables 3–5; Section 12, Table 11"},{"comment":"The CFG strengths in Table 14 of the supplementary were \"obtained\" by ablations on the evaluation benchmarks, and the curriculum-learning variant uses FID from the last evaluation to set sampling probabilities. This means the reported numbers are partially the result of fitting hyperparameters to the test benchmarks. While this is common in the field, the paper should acknowledge it in the main text, report the validation split used for tuning, and indicate the sensitivity of the main results to the chosen CFG strengths. This also bears on the comparison with baselines, which may not have been tuned in the same way.","section":"Section 6.3, Table 6; Section 12, Table 14"}],"minor_comments":[{"comment":"The grouped bars in Figure 1 are difficult to interpret because the values for different metrics are on very different scales and the legend for the three groups (Previous SOTA, Specialists, Ours) is sparse; please clarify in the caption which metric each panel refers to and consider normalizing the axes.","section":"Figure 1"},{"comment":"The term \"condition\" is used broadly to include text, trajectory, keyframe poses, style motion, and even the empty set; please state explicitly that a condition may be a concatenation of multiple signals and that the source motion can be empty for generation tasks.","section":"Section 4, Table 2"},{"comment":"The \"Average Error\" metric for trajectory-based generation and motion in-between is described only qualitatively; please specify which joints are included, how keyframe errors are aggregated, and whether the error is computed in the HumanML3D feature space or in joint coordinates.","section":"Section 6, Evaluation Metrics"},{"comment":"The row label \"w/o motion curriculum learning\" is described in the text as \"directly learns all tasks based on their FID compared to the last evaluation,\" which is more precisely a random/adaptive task-selection strategy; consider renaming the row to \"random task selection\" to avoid confusion with removing the curriculum ordering.","section":"Section 6.3, Table 6"}],"recommendation":"major_revision","confidential_remarks":"The contradiction in the supplementary Table 8 is the most serious issue: the claim of outperforming CondMDI on all settings is directly contradicted by the 1-frame FID. The authors should be asked to correct the numbers or revise the claim, and also to add multiple-seed statistics and an ablation with a reversed curriculum order. I recommend major_revision rather than reject because the main-text results for text and trajectory tasks are strong and the issues are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, the short version: MotionLab is a serious attempt at a unified human motion model covering six tasks, and the Motion-Condition-Motion framing is actually a neat way to organize generation and editing. But the paper contains an internal contradiction in its own supplementary numbers that undercuts the central \"outperforms specialists everywhere\" claim. It needs a revision, not a desk reject.\n\nWhat's genuinely new: the paradigm itself. Casting every task as source motion → condition → target motion is simple and it works for text/trajectory generation, editing, in-betweening, and style transfer. The architecture is an assembly of known parts—rectified flows, MM-DiT joint attention, RoPE, CLIP text—but the aligned 1D RoPE for source/target synchronization and the task-instruction modulation are sensible, and the curriculum learning with two-stage pre-training/fine-tuning is a reasonable way to avoid catastrophic forgetting. The main tables show real gains on text generation (FID 0.167 vs 0.254 for CondMDI), trajectory generation (avg err 0.0334 vs 0.0404 for OmniControl), and editing R@1. The inference speed advantage is also credible.\n\nThe soft spots, in order of severity. First, the supplementary Table 8 directly contradicts the text. For 1-frame motion in-between, MotionLab's FID is 0.755 vs CondMDI's 0.155, R-precision 0.668 vs 0.679, diversity 8.91 vs 9.58—worse on all three. Yet Section 11 says \"our framework outperforms CondMDI on all settings.\" That is a load-bearing overclaim because the whole selling point is that a unified model beats specialists. It may be a typo or a bad checkpoint, but as written the paper's own evidence contradicts its headline claim. Second, no standard deviations or multiple seeds anywhere. For a field where FID differences of 0.05 are routinely compared, that's a real gap. Third, per-task CFG strengths are tuned on the evaluation benchmarks (Table 14, Section 15). That is common practice in this subfield, but it means the reported numbers are partly fitting the test set. Fourth, the curriculum learning's difficulty ordering is intuitive—fewer modalities, editing before generation—but no alternative ordering is tested; the ablations show it matters a lot, so we don't know how sensitive the result is to that choice.\n\nNone of this kills the core idea. The paradigm is credible and the architecture is sensible; I'd bet the unified model does work well with the right hyperparameters. But the paper needs a careful correction of the in-between numbers and honest reporting of variance before the superiority claim is defensible.\n\nWho should read it: anyone working on human motion generation/editing or unified multi-task generative models. It deserves a serious referee—a competent reviewer will catch the contradiction and push for a clean revision. I'd send it out.","headline":"A strong unified-motion architecture with a neat paradigm, but the paper's own supplementary table contradicts its headline superiority claim and needs a fix before publication.","tokens_in":19482,"tokens_out":2403,"would_cite":false,"duration_ms":22679,"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":"The paper's central claim is that a single model, MotionLab, can unify all six major human motion generation and editing tasks and outperform its own task-specific specialists on every one.","keywords":["human motion generation","human motion editing","Motion-Condition-Motion paradigm","rectified flow","MotionFlow Transformer","curriculum learning","multi-task learning","MotionLab"],"falsifier":"Run the identical MotionLab training with the fine-tuning task order reversed (style transfer and trajectory-based editing first, text-based generation last) or randomly permuted. If a different order also makes the unified model beat its own specialist models on the main metrics, the curriculum's specific ordering is not doing the claimed work; if the reversed order collapses text-generation FID while preserving editing gains, the difficulty assumption is supported.","tokens_in":18478,"feed_emoji":"🏃","tokens_out":16908,"duration_ms":147867,"temperature":0.7,"pith_summary":"MotionLab's thesis is that human motion generation and editing are one problem, not many: given a source motion (possibly none), a condition (text, trajectory, keyframes, or style), and a target motion, learn the map between them. The paper builds a single rectified-flow model around this Motion-Condition-Motion schema and trains it on six tasks with a two-stage curriculum. It reports that the unified model beats its own task-specialist versions on every task and shows strong results against published baselines on the same benchmarks, which matters because one network could then replace a zoo of separate motion models. The takeaway claim is that multi-task knowledge sharing, not task-specific architecture, is what buys both quality and speed.","feed_headline":"One motion model beats specialists at all six tasks","feed_subtitle":"One framework replaces six task-specific models and edges them out on quality and speed.","key_machinery":"The load-bearing object is the Motion-Condition-Motion triple (source motion, condition, target motion), which rewrites every task as a conditional transportation problem. The machinery that solves it: rectified flows, which learn a straight velocity field from a noisy target back to the true target; the MotionFlow Transformer (MFT), a joint-attention transformer with a dedicated modality path per input, so source motion, target motion, trajectory, style, and text all interact in one attention pass; Aligned ROPE, a 1D rotary positional encoding applied to each modality so that tokens at the same frame of source and target stay time-aligned; Task Instruction Modulation, which feeds a CLIP text embedding describing the task (e.g., \"edit source motion by given text\") as a modulation signal; and Motion Curriculum Learning, which first pre-trains on masked reconstruction and then fine-tunes tasks from easy to hard. The shared mechanism is that spatially concrete conditions (trajectory, keyframes, source motion) can partially represent abstract conditions (text, style), so the curriculum transfers spatial knowledge to the abstract modalities.","core_discovery":"The paper's claim is that the main human motion tasks — text-based generation, trajectory-based generation, motion in-between, text-based editing, trajectory-based editing, and style transfer — reduce to one schema, Motion-Condition-Motion: every task specifies a source motion (possibly empty for generation), a condition (text, trajectory, keyframes, or style motion), and a target motion, and the model learns a single rectified-flow velocity field that transports a noisy target into the correct target given the source and condition. On the benchmarks, MotionLab reports text-generation FID of 0.167 (specialist: 0.209), trajectory-generation average error of 0.0334 (specialist: 0.0398), text-editing R@1 of 56.34 (specialist: 41.44), trajectory-editing R@1 of 72.65 (specialist: 59.86), in-betweening average error of 0.0283 (specialist: 0.0371), and style-transfer CRA/SRA of 44.62/69.21 (specialist: 43.53/67.55). The ablations attribute the dominant share of these gains to Motion Curriculum Learning: without it, text-generation FID rises from 0.167 to 1.956 and editing retrieval drops by about half, while the full unified model beats its own specialist models on every task.","pith_inferences":["The paper's curriculum claim could be stress-tested by reversing the fine-tuning order; if a different order still beats the specialists, the explanatory mechanism would be multi-task reweighting rather than the specific easy-to-hard order.","Because the schema treats any input as a potential source motion and any output as a target motion, the same design could transfer to other time-aligned sequence-editing domains, such as facial-expression editing or gesture editing driven by language, where source and target must stay frame-aligned.","Task Instruction Modulation depends on CLIP's semantics; tasks whose instructions fall outside CLIP's text distribution would need a new conditioning vector, which is both a simple extension path and a likely failure mode for unusual editing commands."],"forward_implications":["A single MotionLab network covers text-based generation, trajectory-based generation, motion in-between, text-based editing, trajectory-based editing, and style transfer, so a practitioner no longer needs to train and run six separate models.","Data-scarce editing tasks inherit spatial knowledge from data-rich generation tasks; the editing retrieval numbers beat specialist models trained only on the editing data.","The Aligned ROPE component is what keeps source and target frames synchronized; removing it roughly triples in-betweening error (0.0756 vs 0.0283) and sharply drops editing retrieval.","With 50 inference steps and classifier-free guidance, text generation takes about 0.07 seconds per sample, faster than every diffusion-based baseline except the real-time MotionLCM."],"supporting_citations":[{"why":"Supplies the rectified-flow objective (linear interpolation between source and target, constant velocity) that MotionLab is trained to regress.","marker":"[38]"},{"why":"Provides the flow-matching formulation of generative transport that the method's velocity field is based on.","marker":"[37]"},{"why":"The MM-DiT architecture with joint attention and per-modality paths that the MotionFlow Transformer extends to motion modalities.","marker":"[15]"},{"why":"MoMask's random masking scheme (zero to all frames) that MotionLab uses for self-supervised pre-training.","marker":"[23]"},{"why":"MotionFix provides the paired text-editing dataset and the TMED baseline that MotionLab's editing results are compared against.","marker":"[6]"},{"why":"OmniControl defines the trajectory-based generation task and serves as the main spatial-control baseline.","marker":"[64]"},{"why":"MCM-LDM is the multi-condition style-transfer baseline that MotionLab compares against for style accuracy.","marker":"[55]"},{"why":"CondMDI is the flexible motion in-betweening diffusion baseline that MotionLab's keyframe-error numbers are compared with.","marker":"[11]"},{"why":"MotionLCM is the real-time controllable-motion baseline whose inference time and generation numbers MotionLab is compared against.","marker":"[12]"},{"why":"CLIP's text embeddings are used by Task Instruction Modulation to tell the network which task to solve.","marker":"[49]"}],"fun_headline_variants":["One motion model beats six specialists on every task","Unified framework outperforms task-specific motion models","MotionLab: one paradigm for generation and editing, all tasks","Six motion tasks reduced to one, with better scores","New paradigm lets one model handle all six motion tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the curriculum's difficulty ordering — fewer modalities before more, editing before generation, spatial conditions before text and style — is the right one; the paper justifies it by intuition and the ablations do not test an alternative ordering.","fun_headline_variants_meta":{"raw":{"variants":["One motion model beats six specialists on every task","Unified framework outperforms task-specific motion models","MotionLab: one paradigm for generation and editing, all tasks","Six motion tasks reduced to one, with better scores","New paradigm lets one model handle all six motion tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00058,"raw_usage":{"total_tokens":2810,"prompt_tokens":1099,"completion_tokens":1711,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":715,"completion_tokens_details":{"reasoning_tokens":1635}},"tokens_in":715,"tokens_out":1711,"duration_ms":14331,"temperature":1.0,"reasoning_tokens":1635,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T12:25:22.434284+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical MotionLab training with the fine-tuning task order reversed (style transfer and trajectory-based editing first, text-based generation last) or randomly permuted. If a different order also makes the unified model beat its own specialist models on the main metrics, the curriculum's specific ordering is not doing the claimed work; if the reversed order collapses text-generation FID while preserving editing gains, the difficulty assumption is supported.","supporting_citations":[{"cited_title":"Scaling recti- fied flow transformers for high-resolution image synthesis","cited_arxiv_id":null,"evidence_quote":"The MM-DiT architecture with joint attention and per-modality paths that the MotionFlow Transformer extends to motion modalities."},{"cited_title":"Momask: Generative masked model- ing of 3d human motions","cited_arxiv_id":null,"evidence_quote":"MoMask's random masking scheme (zero to all frames) that MotionLab uses for self-supervised pre-training."},{"cited_title":"Motionfix: Text-driven 3d human motion editing","cited_arxiv_id":null,"evidence_quote":"MotionFix provides the paired text-editing dataset and the TMED baseline that MotionLab's editing results are compared against."},{"cited_title":"Arbitrary motion style transfer with multi-condition motion latent dif- fusion model","cited_arxiv_id":null,"evidence_quote":"MCM-LDM is the multi-condition style-transfer baseline that MotionLab compares against for style accuracy."},{"cited_title":"Flexible motion in-betweening with diffusion models","cited_arxiv_id":null,"evidence_quote":"CondMDI is the flexible motion in-betweening diffusion baseline that MotionLab's keyframe-error numbers are compared with."},{"cited_title":"Motionlcm: Real-time controllable motion generation via latent consistency model","cited_arxiv_id":null,"evidence_quote":"MotionLCM is the real-time controllable-motion baseline whose inference time and generation numbers MotionLab is compared against."},{"cited_title":"Learning transferable visual models from natural language supervi- sion","cited_arxiv_id":null,"evidence_quote":"CLIP's text embeddings are used by Task Instruction Modulation to tell the network which task to solve."}],"review_version":1}