{"id":"5c2b5f2e-19d0-425e-beac-5656db3711d9","arxiv_id":"2412.04343","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"RMD improves text-to-motion generation by decomposing prompts with an LLM, retrieving and recombining body-part motions, and refining them with a pretrained diffusion model, all without additional training.","lead":"This paper introduces RMD, a method that generates 3D human motions from text by retrieving and recombining body-part motion clips from a database, then cleaning the result with a pretrained motion diffusion model. It needs no extra training, so it can be dropped on top of existing models to improve handling of unusual, out-of-distribution text prompts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Retrieval contribution may be washed out at t0=0.96, and the choice of t0 is made on the HumanML3D test set, so the reported SOTA gain may be an artifact of test-set hyperparameter selection.","rationale":"The reader identified the SDEdit conditioning assumption as the weakest point, but I see the stronger and more specific issue as the combination of that assumption with test-set hyperparameter selection. The paper's evidence does not currently rule out the possibility that at t0=0.96 the retrieval stage contributes little or nothing, and the only quantitative support for that operating point comes from a sweep directly on the evaluation set. The supplementary Table 7 experiment is actually a natural stress test, and it shows a different optimal t0 (0.8), which raises further doubt about the chosen 0.96 for the main comparison. Because the reported gains are small (Top-1 R-Precision 0.524 vs. 0.515) and the FID and MultiModality on Mixamo are worse than the base model, the claim of 'notable advantages on out-of-distribution data' is also weaker than the abstract suggests. These are addressable issues rather than fatal flaws: the method is plausible, ablations in Table 3 do show an internal ordering consistent with the design, and Table 4 shows consistent improvements across four MotionDiffuse variants. A validation-controlled re-evaluation would settle whether the central claim holds; if it does, the paper warrants acceptance. I therefore recommend keeping the reader's CONDITIONAL verdict and adding this specific validation requirement.","tokens_in":15782,"tokens_out":4962,"duration_ms":52160,"concrete_test":"Hold out a validation split (e.g., 10% of the HumanML3D training set), perform the t0 sweep on that split only, and then evaluate RMD and MotionDiffuse on the official test split at the validation-selected t0 with identical sampling seeds. If the Top-1 R-Precision advantage over MotionDiffuse is not preserved at that t0, the headline result is an artifact of test-set hyperparameter selection.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that a training-free retrieval stage adds semantic content that a pretrained MotionDiffuse prior then refines. This requires the SDEdit operating point t0=0.96 to preserve enough of the retrieved/composed motion to influence the final output. However, at t0=0.96 with 50 DDIM steps, the guided motion is first corrupted with near-pure noise (σ(0.96) is close to the maximum of the SDE noise scale), so the reverse process is overwhelmingly driven by the text-conditioned diffusion prior rather than by the retrieved composition. The paper's own supplementary Table 7 shows that when the retrieval database and training set are deliberately mismatched, the optimal t0 shifts to 0.8, indicating that the right operating point is distribution-dependent and that 0.96 is not a robust default. More importantly, Figure 6 sweeps t0 on the HumanML3D test set, and no held-out validation split is described. Thus the reported t0=0.96 is selected to maximize the exact R-Precision numbers in Table 1. If t0 were chosen on a validation split, the +0.009 Top-1 R-Precision gain over the authors' re-trained MotionDiffuse baseline could shrink or disappear, which would remove the empirical support for 'state-of-the-art' and for the claim that retrieval is causally load-bearing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RMD, a training-free retrieval-augmented pipeline for text-to-motion generation. Given a text prompt, an LLM-based decomposition agent decides whether to retrieve a full-body motion, half-body motions, or fine-grained body-part motions from an external database of motion-text pairs. The retrieved parts are composed into a guided motion, which is then refined by a pretrained MotionDiffuse model via SDEdit noise-and-denoise, conditioned on the original prompt. The authors evaluate RMD on HumanML3D, a cross-domain Mixamo test set, and a 33-prompt user study, reporting state-of-the-art R-Precision and MM Dist on HumanML3D and improved R-Precision on Mixamo. The central claim is that, without any training, RMD improves semantic alignment and generalization of text-to-motion models while reusing the base model unchanged.","tokens_in":16099,"tokens_out":6789,"duration_ms":60449,"significance":"If the claims hold, RMD is an appealingly simple and practical method: it requires no training, allows the retrieval database to be swapped, reuses body-part motions, and can improve an existing base model at inference time. The paper's explicit evaluation on a cross-domain Mixamo benchmark and on out-of-distribution user prompts is a useful contribution, and the authors are transparent about the base-model variants and ablations. The method's dependency on only frozen components (CLIP, GPT-4o, MotionDiffuse) makes it easy to build on. However, the empirical support for the central claims is currently weakened by test-set hyperparameter tuning, small and mixed gains, and a user study without statistical analysis.","major_comments":[{"comment":"The main hyperparameters t0, tau_half, tau_fine, and lambda are selected on the HumanML3D test set. Figure 6 sweeps t0 on the test set and the authors state they \"find the performance reaches a peak when t0=0.96 and choose this value for our main comparisons.\" No held-out validation split is described, and the same test set is then used to report Table 1. Because the reported improvement over the retrained MotionDiffuse baseline is small (+0.009 Top-1 R-Precision, +0.007 Top-2, +0.005 Top-3; FID 0.111 vs 0.141), it is possible that part or all of the gain is due to test-set tuning. The authors should either select t0 and thresholds on a held-out validation split and then report test results, or show a sensitivity analysis over a range of t0 values with confidence intervals and demonstrate that the conclusions are stable.","section":"Section 4.4, Fig. 6"},{"comment":"The causal role of the retrieval stage is not established. At t0=0.96 with 50 DDIM steps, the guided motion is corrupted with near-maximal noise, and the reverse process is largely driven by the text-conditioned diffusion prior rather than the retrieved composition. The paper acknowledges in Section 3.3 that t0 balances the guide and the prior, but it provides no quantitative evidence that the final output preserves the body-part semantics inserted by retrieval. The supplementary Table 7 shows that when the retrieval database and training set are mismatched, the optimal t0 shifts to about 0.8, which indicates the operating point is distribution-dependent and not a robust default. To support the claim that retrieval is load-bearing, the authors should include an analysis measuring the similarity between the final output and the guided motion, or an experiment that replaces the retrieved guided motion with a random or mean motion and evaluates the metrics at the same t0 values.","section":"Section 3.3 and Supplementary Table 7"},{"comment":"The cross-domain evidence for \"notable advantages on out-of-distribution data\" is weak and mixed. On the Mixamo benchmark, RMD improves R-Precision over MotionDiffuse (e.g., Top-1 10.015 vs 9.617) but its FID is worse (4.372 vs 4.318) and its Multimodality is worse (2.174 vs 2.333); the confidence intervals overlap on these metrics. The user study reports preference percentages from only 16 participants with no significance test, no error bars, and no inter-rater reliability. The paper should either present a formal significance analysis (e.g., bootstrap or paired test) or temper the claim from \"significantly outperforms\" to a more limited qualitative statement. Additionally, a cross-domain claim should be supported by more than one alternative dataset.","section":"Section 4.3, Table 2 and Fig. 5"}],"minor_comments":[{"comment":"Section 3.1 defines fine-grained decomposition into \"six body parts\" (head, torso, left arm, right arm, lower body, trajectory), but the prompt in Supplementary Table 9 asks for seven lines: head, spine, left upper limb, right upper limb, left lower limb, right lower limb, and trajectory. Please align the number and naming of body parts between the method description and the LLM prompts.","section":"Section 3.1 vs. Supplementary Table 9"},{"comment":"The sentence \"taking the highest match score as the similarity score si\" is ambiguous: it is unclear whether si is the score of the LLM-selected description or the maximum score across the k candidates. Please clarify the procedure.","section":"Section 3.2, Eq. (1)"},{"comment":"The terms \"startx\" and \"epsilon\" are not defined in the main text; please define them when the table is first referenced (and note the typo \"residule\" in the caption).","section":"Section 4.3, Table 4"},{"comment":"Please specify which version of MotionDiffuse is used as the base model in each table: the re-trained 12-layer variant or the original. The text says \"we re-train MotionDiffuse,\" but the reader needs to know whether the comparison numbers for other methods were obtained under the same inference protocol (e.g., 50 DDIM steps).","section":"Section 4.2"},{"comment":"Table 6 reports no confidence intervals for the SDS, DDS, and DNO baselines, unlike the other tables. Please report them for consistency.","section":"Supplementary Table 6"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is simple and plausible, and the cross-domain benchmark is a useful addition to the field. The main risk is that the reported gains may be inflated by test-set hyperparameter selection, and the paper does not yet demonstrate that the retrieval information survives the SDEdit refinement. If the authors can provide a validation-based hyperparameter choice and a direct analysis of the retrieval contribution, the paper could be acceptable. I would also strongly encourage releasing code, given the reproducibility value of a training-free pipeline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a genuinely training-free, plug-and-play retrieval baseline for text-to-motion, and it does show consistent small gains on the standard benchmark, on a cross-domain split, and on KIT-ML. The decomposition-retrieve-compose pipeline with an LLM and SDEdit refinement is not entirely novel component-wise (ReMoDiffuse, TMR++, PriorMDM cover parts), but the specific combination is new and is presented as a simple baseline rather than a full system. That is the honest value: it is easy to reproduce conceptually, and it gives instant improvement to an existing MotionDiffuse without retraining.\n\nWhat it does well: the ablations are systematic; they test different base models, different retrieval strategies, and they openly report that the optimal SDEdit starting time t0 shifts when the retrieval database exceeds the training set (supp Table 7). They also include a user study on OOD prompts, even though it is small. The paper is honest about the limitation that t0 is fixed and manual.\n\nSoft spots: the main one is the selection of t0=0.96 on the HumanML3D test set. Figure 6 sweeps t0 on the exact test split used for Table 1, so the reported SOTA numbers carry a selection bias. The margin over the re-trained MotionDiffuse baseline is modest (Top-1 0.524 vs 0.515), and with a proper validation split the gain may shrink. The stress-test concern that at t0=0.96 the retrieval contribution is nearly washed out is not fully resolved. At that setting the guided motion is heavily corrupted, so the output is close to the base model's free-running generation; the fact that performance still peaks there suggests retrieval provides only a weak prior that the base model largely overrides. That is not inconsistent with the method working, but it weakens the causal story that the retrieval stage is load-bearing. The cross-domain numbers on Mixamo are mixed: R-precision is better but FID and Multimodality are worse than the base. The user study has 16 raters and no significance test, so treat it as suggestive.\n\nOverall, this deserves a serious referee. The idea is simple, the evidence is mostly positive, and the flaws are addressable. A reviewer should ask for a validation split to set hyperparameters, a recomputed main table with the chosen t0, and code release before acceptance. It is a useful baseline paper, not a breakthrough, but it is a legitimate contribution to the motion generation subfield.","headline":"A simple, honest training-free retrieval baseline for text-to-motion, with gains that are real but smaller than claimed once you account for test-set t0 selection.","tokens_in":16612,"tokens_out":2895,"would_cite":true,"duration_ms":27333,"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":"Training-free RMD beats text-to-motion baselines on out-of-distribution data","keywords":["text-to-motion generation","retrieval-augmented generation","motion diffusion","out-of-distribution generalization","large language model","SDEdit","human motion synthesis"],"falsifier":"On the HumanML3D test set, replace the retrieved guided motion with a randomly selected motion of matching length from the retrieval database before the SDEdit refinement step. If R-Precision and MM Dist stay at the same level as the full RMD results instead of dropping toward the base MotionDiffuse numbers, then the diffusion refinement is ignoring the guided motion and the retrieval stage is not contributing. A second check: feed retrieval clips whose text contradicts the prompt (e.g., 'running' for a 'walking' prompt) and see whether the final output follows the prompt or the retrieved clip.","tokens_in":15615,"feed_emoji":"🏃","tokens_out":6620,"duration_ms":58357,"temperature":0.7,"pith_summary":"This paper introduces RMD, a training-free retrieval-augmented pipeline for text-to-motion generation. The method claims that by decomposing a text prompt into body-part descriptions with an LLM, retrieving matching motions for each part from an external database, recomposing them, and then refining the result with a pre-trained MotionDiffuse model under SDEdit-style noise-and-denoise, it achieves state-of-the-art R-Precision and MM Dist on HumanML3D without any parameter updates. Its main advantage is on out-of-distribution data: on the Mixamo cross-domain test and in a user study with real-world prompts, RMD outperforms existing baselines. Because motion datasets are small and real-world prompts are far from training text distributions, a training-free plug-in that improves generalization could make text-to-motion systems more practical.","feed_headline":"Training-free RMD beats text-to-motion baselines on OOD data","feed_subtitle":"It splits prompts into body parts, retrieves clips, and refines with a pretrained diffusion prior.","key_machinery":"The central machinery is a three-level hierarchical decompose-retrieve-recompose pipeline followed by SDEdit refinement. An LLM decomposition agent splits each prompt and each database motion description into half-body (upper/lower) and fine-grained (head, torso, left arm, right arm, lower body, trajectory) descriptions. A retrieval agent scores each query part against database entries using CLIP text embeddings with a length-aware similarity term (Eq. 1) and uses an LLM to pick among k=5 decomposition candidates. Retrieved parts are recombined into a full-body motion by concatenating joint quaternions, rescaled to the query length via SLERP and linear interpolation. The composed 'guided motion' is then refined by adding Gaussian noise up to time t0=0.96 and running the pretrained MotionDiffuse denoiser from t0 to 0, conditioned only on the original text prompt.","core_discovery":"The central claim is that a pretrained motion diffusion model can serve as a sufficient prior to repair motions assembled from retrieved body-part clips, and that this two-stage pipeline, without any training, outperforms both trained retrieval-augmented models and recent generative baselines. On HumanML3D, RMD reports the best top-1/2/3 R-Precision and Multi-Modal Distance, with a slight drop in MultiModality relative to the base MotionDiffuse. On the cross-domain Mixamo test, it reports the best R-Precision scores, and the percentage of queries requiring fine-grained decomposition rises from 5.9% on HumanML3D to 36.9% on Mixamo and 51.5% on user-collected real-world prompts, supporting the paper's argument that current benchmarks under-represent compositional OOD motions.","pith_inferences":["This suggests that any diffusion-based text-to-motion model with a reverse-SDE formulation could be upgraded at inference time by the same decompose-retrieve-recompose-refine recipe, without retraining.","The LLM-driven decomposition could be automated further, for instance by learning to predict the decomposition level and t0 per prompt, which the paper flags as future work.","The fact that retrieval from the same training set improves performance implies current training algorithms do not fully exploit the training data, pointing toward memory-augmented inference as a cheaper alternative to scaling model size."],"forward_implications":["RMD improves MotionDiffuse on HumanML3D across all metrics except a small MultiModality drop, reaching the best R-Precision and MM Dist among compared methods.","On the Mixamo cross-domain benchmark, RMD achieves the best R-Precision, showing a clear advantage over in-domain-strong baselines when tested on a different motion distribution.","Real-world OOD prompts trigger far more fine-grained decomposition than HumanML3D test prompts (51.5% versus 5.9%), indicating a large gap between standard benchmarks and practical usage.","RMD gives consistent gains across several MotionDiffuse variants (startx/epsilon, 8-layer/12-layer), suggesting it is a robust plug-in rather than a special-case tuning.","When the retrieval database is larger than the diffusion model's training set, RMD's improvement grows and the optimal t0 drops from 0.96 to 0.8, showing retrieval becomes more valuable as the database expands."],"supporting_citations":[{"why":"Supplies the SDEdit noise-and-denoise scheme that RMD uses to refine the composed guided motion with a pretrained diffusion prior.","marker":"[16]"},{"why":"The pretrained MotionDiffuse model that RMD wraps, and the primary baseline it improves.","marker":"[32]"},{"why":"Provides the frozen CLIP text encoder used to embed queries and database descriptions for retrieval.","marker":"[24]"},{"why":"Provides the HumanML3D benchmark dataset, the pose representation, and the evaluation protocol; its training set forms the retrieval database.","marker":"[8]"},{"why":"Supplies the Mixamo cross-domain test set used to measure out-of-distribution generalization.","marker":"[1]"},{"why":"The length-aware similarity scoring in Equation 1 is adopted from this work, and it serves as a key retrieval-augmented baseline.","marker":"[33]"},{"why":"A state-of-the-art generative baseline (MoMask) that RMD compares against on HumanML3D and Mixamo.","marker":"[9]"}],"fun_headline_variants":["Training-free RMD beats trained baselines on OOD motion","No-training retrieval-augmented motion: RMD leads OOD","RMD: zero-training retrieval for out-of-distribution motion","Training-free pipeline: retrieve, split, and diffuse motion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The refinement step assumes that at the chosen noise level t0=0.96, the pretrained diffusion model will remove the artifacts of compositional stitching while preserving the specific body-part semantics the retrieval stage added, even though the refinement is conditioned only on the original text prompt, not on the retrieved part descriptions.","fun_headline_variants_meta":{"raw":{"variants":["Training-free RMD beats trained baselines on OOD motion","No-training retrieval-augmented motion: RMD leads OOD","RMD: zero-training retrieval for out-of-distribution motion","Training-free pipeline: retrieve, split, and diffuse motion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000348,"raw_usage":{"total_tokens":1860,"prompt_tokens":856,"completion_tokens":1004,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":472,"completion_tokens_details":{"reasoning_tokens":933}},"tokens_in":472,"tokens_out":1004,"duration_ms":9972,"temperature":1.0,"reasoning_tokens":933,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:31:15.184199+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the HumanML3D test set, replace the retrieved guided motion with a randomly selected motion of matching length from the retrieval database before the SDEdit refinement step. If R-Precision and MM Dist stay at the same level as the full RMD results instead of dropping toward the base MotionDiffuse numbers, then the diffusion refinement is ignoring the guided motion and the retrieval stage is not contributing. A second check: feed retrieval clips whose text contradicts the prompt (e.g., 'running' for a 'walking' prompt) and see whether the final output follows the prompt or the retrieved clip.","supporting_citations":[{"cited_title":"Sdedit: Guided image synthesis and editing with stochastic differential equa- tions","cited_arxiv_id":null,"evidence_quote":"Supplies the SDEdit noise-and-denoise scheme that RMD uses to refine the composed guided motion with a pretrained diffusion prior."},{"cited_title":"Learn- ing transferable visual models from natural language super- vision","cited_arxiv_id":null,"evidence_quote":"Provides the frozen CLIP text encoder used to embed queries and database descriptions for retrieval."},{"cited_title":"Generating diverse and natural 3d human motions from text","cited_arxiv_id":null,"evidence_quote":"Provides the HumanML3D benchmark dataset, the pose representation, and the evaluation protocol; its training set forms the retrieval database."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Mixamo cross-domain test set used to measure out-of-distribution generalization."},{"cited_title":"Momask: Generative masked model- ing of 3d human motions","cited_arxiv_id":null,"evidence_quote":"A state-of-the-art generative baseline (MoMask) that RMD compares against on HumanML3D and Mixamo."}],"review_version":1}