{"id":"68efc63d-0c28-4daf-b92f-525edc82a904","arxiv_id":"2411.16964","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A diffusion model trained on discrete wavelet transform coefficients of motion sequences improves human motion prediction accuracy on standard benchmarks.","lead":"MotionWavelet converts observed and future human motion into wavelet coefficients, then trains a diffusion model to predict future poses by denoising those coefficients. It reports accuracy gains over prior motion prediction methods on Human3.6M and HumanEva-I, plus controllable prediction examples.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The conditioning history padding is unspecified: training uses DWT(x1:H) while inference uses an undefined padded condition, making the method irreproducible and the claimed gains potentially padding artifacts.","rationale":"The central claim is that wavelet-domain diffusion with the proposed guidances predicts human motion more accurately. For this to hold, the conditioning mechanism must be well-defined and consistent between training and inference. The paper's Eq. (5) and Algorithm 1 disagree on what the condition is: the former is the DWT of the observed history, the latter is a 'padded history' that is never defined. Because the DWT is a global transform, zero-padding the history before the DWT changes the boundary coefficients and injects future zeros into the condition. If the authors used the padded condition during training as well, they need to say so; if not, the CFG guidance is applied to an out-of-distribution condition and the reported gains are unexplained. The reader's verdict is CONDITIONAL for good reason; this concern alone prevents acceptance without code or a precise specification. The apparent copy of HumanMAC's numbers in Table 2 and the non-predictive RMSE in Table 3 are additional red flags, but the padding issue is the most fundamental because it concerns the validity of the method itself.","tokens_in":17328,"tokens_out":10034,"duration_ms":89695,"concrete_test":"In the released code (or by asking the authors), identify the exact construction of y(0)_pad in Algorithm 1. Verify that the training loss Eq. (5) uses the same padded condition. Then re-run the Human3.6M experiment under two conditions: (A) training and inference both use DWT of the zero-padded history; (B) training and inference both use a network that accepts the unpadded DWT(x1:H) directly. If the reported Table 1 numbers are only reproduced under (A), the method's success depends on the undefined padding, not on wavelet manifold learning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2.2 defines the training condition as y(0) = DWT(x1:H) (Eq. 5), which has temporal dimension K_H = floor((H+l-1)/2). Algorithm 1 at inference uses 'Conditioning padded history y(0)_pad' but never defines how the H observed frames are extended to the full H+F length before applying DWT. Appendix A states the padding mode is zero at the boundaries, but the convolution in Eq. (3) mixes boundary frames, so DWT([x1:H, 0:F]) is not equal to DWT(x1:H) with zeros appended in the wavelet domain. The noisy latent y(t) has temporal dimension K ≈ (H+F)/2, so the condition must be padded to length K to match the network input. If training uses the unpadded DWT(x1:H) and inference uses the padded version, the classifier-free guidance in Eq. (7) is evaluated with a condition distribution the network never saw, and the reported accuracy gains could be an artifact of the padding rather than the wavelet manifold. This is not a minor implementation detail: it determines the actual input to the denoiser and cannot be resolved from the paper as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MotionWavelet, a human motion prediction method that transforms motion sequences into a wavelet-domain manifold via a 2D discrete wavelet transform, trains a DDIM-based diffusion model on this manifold conditioned on the wavelet coefficients of the observed history, and adds two inference-time mechanisms: Wavelet Manifold Shaping Guidance (WMSG), which projects each denoised latent through an iDWT followed by a DWT, and Temporal Attention-Based Guidance (TABG), which re-noises and denoises temporally attended regions under an attention mask. Experiments on HumanEva-I and Human3.6M report improved FDE, MMADE, and MMFDE over prior methods, along with ablations on wavelet bases, diffusion schedulers, and guidance hyperparameters.","tokens_in":17617,"tokens_out":6033,"duration_ms":53484,"significance":"The central idea of performing diffusion in a wavelet-domain latent space is plausible and potentially useful: it explicitly separates frequency bands and may help capture non-stationary motion transitions. The manuscript contains a fairly extensive set of ablations, and the authors commit to releasing code and models, which would help reproducibility. However, the current evidence is weakened by an apparent data inconsistency in Table 2, an underspecified conditioning-padding protocol that directly affects the classifier-free guidance input, an ambiguous sampling algorithm, and an overstated accuracy claim on the ADE metric. These issues need to be resolved before the empirical claims can be accepted.","major_comments":[{"comment":"The row labeled 'Wavelet' in Table 2 reports APD=6.301, ADE=0.369, FDE=0.480, which are exactly the HumanMAC baseline numbers from Table 1 (APD=6.301, ADE=0.369, FDE=0.480), not the proposed MotionWavelet numbers (APD=6.506, ADE=0.376, FDE=0.408). Consequently, Table 2 does not demonstrate that the wavelet representation outperforms the phase representation; it appears to compare a phase-based variant with an unrelated published baseline. This undermines the conclusion of Section 4.7.1 and must be corrected with the actual MotionWavelet results.","section":"Section 4.7.1, Table 2"},{"comment":"The conditioning input is not defined consistently. Training uses y(0)=DWT(x1:H) (Eq. 5), whose temporal dimension is roughly K_H=floor((H+l-1)/2), while the noisy latent y(t) has temporal dimension K=floor((H+F+l-1)/2). Algorithm 1 refers to 'Conditioning padded history y(0)_pad' but never states how the H observed frames are padded to the full H+F length before applying the DWT, or how the wavelet-domain condition is padded to length K. Because the DWT in Eq. (3) is a convolution followed by downsampling, DWT of a zero-padded history is not equal to the zero-padding of DWT(x1:H). If training and inference use different padding schemes, the classifier-free guidance term in Eq. (7) is evaluated with a condition distribution the network never saw, and the reported gains could be artifacts of the padding scheme rather than of wavelet manifold learning. The paper must specify the exact padding procedure and confirm that it is identical in training and inference.","section":"Section 3.2.2 and Algorithm 1"},{"comment":"The sampling algorithm is internally ambiguous: the final return statement 'x = iDWT(y(0))' uses y(0) as if it were the predicted clean latent, but y(0) was defined earlier as the conditioning latent DWT(x1:H). The loop does not assign a new value to y(0), so it is unclear which latent is decoded to produce the final prediction. Please rename the conditioning variable, insert the missing assignment, and state explicitly how the final prediction is obtained.","section":"Section 3.2.2 and Algorithm 1"},{"comment":"The statement 'MotionWavelet achieves the overall best performance' is not supported by Table 1. On HumanEva-I, HumanMAC has lower ADE (0.209 vs 0.235); on Human3.6M, CoMotion (0.350) and HumanMAC (0.369) both have lower ADE than MotionWavelet (0.376). Claiming consistent superiority in accuracy requires more than winning FDE, MMADE, and MMFDE, especially since ADE is the primary whole-sequence accuracy metric in this literature. In addition, Table 1 reports no error bars or number of seeds, so the differences cannot be assessed statistically. Please report mean and standard deviation over multiple runs and revise the claim accordingly.","section":"Section 4.5.1 and Table 1"},{"comment":"The RMSE values in Tables 2 and 3 (e.g., 3.572e-8 for Position RMSE in Table 3) are physically implausible for pose coordinates and are not explained anywhere in the text. If these quantities are computed on normalized wavelet coefficients rather than on reconstructed positions, the metric definition and normalization must be stated explicitly; otherwise the ablation claims involving these values cannot be interpreted or reproduced.","section":"Tables 2 and 3"}],"minor_comments":[{"comment":"The notation for the noise predictor is inconsistent: Eq. (11) writes ϵθ(ˆy(t)) while Algorithm 1 writes ϵθ(ˆy(t), t). Please unify the notation.","section":"Algorithm 1 and Eq. (11)"},{"comment":"The symbol y(0) is overloaded: it denotes both the clean target wavelet DWT(x) and the conditioning wavelet DWT(x1:H). Please use a distinct symbol, such as c, for the condition.","section":"Section 3.2.2, Eq. (5)"},{"comment":"Appendix A states that for HumanEva-I, TABG and WMSG are not applied, while the abstract and introduction present both as core components of the method. Please state clearly which final results use which components and whether the ablation conclusions transfer to HumanEva-I.","section":"Appendix A"},{"comment":"The controllable prediction section blends in ground-truth motion during inference, so it should be explicitly framed as an editing/control demonstration rather than as a prediction task.","section":"Section 4.6"},{"comment":"Appendix A mentions that the decomposition level is set to 1, but this detail is not present in the method section. Adding it to Section 3.2.1 would aid reproducibility.","section":"Section 3.2.1"}],"recommendation":"major_revision","confidential_remarks":"The exact duplication of HumanMAC's reported metrics in Table 2's 'Wavelet' row, while the paper's own Table 1 reports different numbers for the same method, is a serious data-consistency issue that the editor may wish to investigate further. The undefined conditioning-padding procedure in Algorithm 1 is also a load-bearing reproducibility gap that should be addressed before the paper can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a fine incremental idea—diffusion in DWT coefficient space for human motion prediction, with two guidance mechanisms—but the paper as written is not reproducible, and one ablation table looks like a baseline copied by mistake. I would send it to review, but major revision, and I would not cite the numbers until code is out.\n\nWhat's genuinely new: putting the motion in the wavelet domain before diffusion, plus Temporal Attention-Based Guidance that masks neighboring wavelet time steps based on attention. The ablations mostly support the design choices: WMSG helps a bit, TABG helps a bit, and the scheduler study follows the usual pattern. The writing is clear and the method section is coherent.\n\nThe soft spots, in order of seriousness. First, the conditioning padding is the load-bearing one. Eq. (5) trains the condition as y(0) = DWT(x_{1:H}); Algorithm 1 samples with y(0)_pad, never defined. Appendix A says the padding mode is zero at the boundaries but never specifies how the H observed frames are extended to H+F before the DWT. With zero padding, DWT(iDWT(y)) is not the identity near the edges, and the condition distribution at inference can differ from training. This is not cosmetic: it determines the actual input to the denoiser and changes the CFG extrapolation in Eq. (7). The stress-test note is right.\n\nSecond, Table 2 is a red flag. The 'Wavelet' row reports APD/ADE/FDE = 6.301/0.369/0.480, which is exactly HumanMAC's Table 1 set (the HumanMAC row in this paper's own Table 1), not this paper's own Table 1 numbers (6.506/0.376/0.408). So the phase-vs-wavelet comparison appears to compare against a copied baseline, not their own wavelet model. That needs a correction.\n\nThird, Table 3 selects the wavelet base by reconstruction RMSE, not by prediction error. That is not a valid way to pick a component for the prediction pipeline.\n\nFourth, the text says 'overall best performance,' but in Table 1 HumanMAC has better ADE on HumanEva (0.209 vs 0.235) and CoMotion has better ADE on H3.6M (0.350 vs 0.376). It is a mixed win, not a sweep.\n\nNone of this proves the method is wrong; the idea is reasonable and the guidance bits are sensible. But with no code, no error bars, and the padding unspecified, the empirical claims are not yet supported. The duplicated row in particular needs an explanation.\n\nWho gets value: people actively working on diffusion-based motion prediction; it is a legitimate combination. I would send it to review, ask for code and a precise padding description, and fix Table 2. If the authors cannot clarify the padding, I would treat the reported gains as unverified.","headline":"A plausible wavelet-diffusion combination for human motion prediction, but the inference-time conditioning padding is unspecified and one ablation row appears to copy a baseline; the accuracy claims need major revision before they can be trusted.","tokens_in":18136,"tokens_out":3482,"would_cite":false,"duration_ms":31186,"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":"The paper claims that predicting future human motion is more accurate when a diffusion model is trained and sampled in a wavelet manifold rather than in raw pose or phase space, and backs the claim with error scores on two standard…","keywords":["human motion prediction","discrete wavelet transform","diffusion model","wavelet manifold","non-stationary motion","attention guidance","controllable motion prediction"],"falsifier":"Re-run the released system on Human3.6M with two different padding choices for the observed frames, zero padding and replication of the boundary values, keeping everything else fixed, and compare final displacement error. If the difference exceeds the gap between MotionWavelet and the second-best method, the reported accuracy gain is not attributable to wavelet manifold learning alone.","tokens_in":17157,"feed_emoji":"🕺","tokens_out":7636,"duration_ms":70791,"temperature":0.7,"pith_summary":"MotionWavelet sets out to improve short-horizon human motion prediction by moving the prediction problem out of raw pose space into a wavelet manifold. The paper argues that a discrete wavelet transform along both time and joint axes separates a motion into low- and high-frequency subbands, so a diffusion model trained on these subbands can represent subtle transitions and non-stationary dynamics that Fourier-phase or DCT representations miss. The reported experiments support this by showing the method outperforming earlier approaches on HumanEva-I and Human3.6M on final displacement error and multi-modal accuracy metrics, and by demonstrating controllable joint-level and motion-switch prediction. A sympathetic reader would care because the same frequency-domain framing might transfer to other problems where short observations must be extrapolated into structured, temporally coherent outputs.","feed_headline":"Wavelet manifold diffusion lowers pose error on two benchmarks","feed_subtitle":"A diffusion model that denoises wavelet coefficients outperforms pose-space methods on two motion benchmarks.","key_machinery":"The load-bearing object is the Motion Wavelet Manifold, defined as the concatenation of the four subbands from a 2-D discrete wavelet transform of a motion sequence: approximation, temporal detail, spatial detail, and spatio-temporal detail. It carries the argument because it is the space in which the diffusion model trains and samples, and because its explicit separation of high- and low-frequency content is what the paper claims lets the model handle non-stationary, abrupt motion. Two auxiliary mechanisms keep denoising on that manifold: Wavelet Manifold Shaping Guidance applies iDWT followed by DWT after every denoising step, and Temporal Attention-Based Guidance uses the denoiser's self-attention maps to re-noise temporally salient frames.","core_discovery":"The paper's central claim is that human motion prediction improves when a diffusion model operates on a manifold built from the 2-D discrete wavelet transform of the full observed-plus-future sequence. The motion tensor $x \\in \\mathbb{R}^{(H+F)\\times 3J}$ is decomposed into four subbands $y_{h,v}$ by low- and high-pass filtering along time and joints; concatenating them defines the wavelet manifold. A DDIM-style diffusion model is trained to denoise these latent coefficients conditional on the DWT of the observed history, and at sampling two mechanisms steer the denoising: Wavelet Manifold Shaping Guidance, which re-projects each intermediate denoised latent through iDWT then DWT to keep it on the manifold, and Temporal Attention-Based Guidance, which re-noises frames whose attention scores exceed a threshold so the model emphasizes temporally coherent motion. The paper reports best or second-best accuracy on both benchmarks, with the Bior2.8 wavelet, a small classifier-free guidance scale, and a cosine noise schedule.","pith_inferences":["Inference: the same wavelet-manifold construction could be applied to other structured extrapolation problems, such as hand motion, instrumented trajectories, or partially observed navigation paths, where abrupt transitions matter; the paper only evaluates full-body mocap.","Inference: the wavelet-basis sweep suggests a testable trade-off: a smoother basis may fit slow, fluid motions while a sharper basis captures impacts and stops, and pairing basis choice to motion style is not examined in the paper.","Inference: since controllable prediction is demonstrated by masking at joint or frame level, a natural next step is interactive editing of predicted motion, where a user can specify target joints or target frames after sampling."],"forward_implications":["On HumanEva-I, MotionWavelet reports the lowest final displacement error, multi-modal ADE, and multi-modal FDE among the compared methods.","On Human3.6M, it reports the best FDE and MMFDE while keeping diversity, measured by APD, comparable to leading methods.","Predicted motions track ground truth through abrupt transitions such as sudden stops and starts, a regime where periodic phase models are expected to fail.","The denoising process can accept a mask blending ground-truth or target motion at chosen joints or frames, yielding controllable motion prediction.","The method works best with a small classifier-free guidance scale ($w<1$), which the paper ties to the wavelet manifold's explicit high-frequency noise representation."],"supporting_citations":[{"why":"Supplies the TransLinear noise-prediction backbone and the masked-completion diffusion baseline that MotionWavelet extends and is compared against.","marker":"[7]"},{"why":"Defines the DDIM sampling procedure used to train and run the wavelet-manifold diffusion model.","marker":"[59]"},{"why":"Provides the Human3.6M motion-capture benchmark on which the main accuracy metrics are reported.","marker":"[29]"},{"why":"Provides the HumanEva-I benchmark, where the method reports its strongest accuracy results.","marker":"[57]"},{"why":"Provides the self-attention guidance idea that Temporal Attention-Based Guidance adapts to the temporal dimension.","marker":"[26]"},{"why":"Represents the phase-manifold frequency-domain approach that the paper argues fails on non-stationary motion and that wavelet manifolds are designed to surpass.","marker":"[61]"}],"fun_headline_variants":["Wavelet manifold diffusion cuts pose error in motion prediction","Diffusion on a wavelet manifold improves human motion forecasting","MotionWavelet: wavelet-space diffusion sharpens pose accuracy","Wavelet manifold learning reduces errors in predicting human motion","Wavelet domain diffusion enhances motion prediction on two datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the observed frames are padded to the full sequence length identically when training the diffusion model and when sampling from it; the paper does not state the padding scheme.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet manifold diffusion cuts pose error in motion prediction","Diffusion on a wavelet manifold improves human motion forecasting","MotionWavelet: wavelet-space diffusion sharpens pose accuracy","Wavelet manifold learning reduces errors in predicting human motion","Wavelet domain diffusion enhances motion prediction on two datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001157,"raw_usage":{"total_tokens":4791,"prompt_tokens":940,"completion_tokens":3851,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":3774}},"tokens_in":556,"tokens_out":3851,"duration_ms":27488,"temperature":1.0,"reasoning_tokens":3774,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:41:07.280829+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the released system on Human3.6M with two different padding choices for the observed frames, zero padding and replication of the boundary values, keeping everything else fixed, and compare final displacement error. If the difference exceeds the gap between MotionWavelet and the second-best method, the reported accuracy gain is not attributable to wavelet manifold learning alone.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Human3.6M motion-capture benchmark on which the main accuracy metrics are reported."},{"cited_title":"Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion","cited_arxiv_id":null,"evidence_quote":"Provides the HumanEva-I benchmark, where the method reports its strongest accuracy results."}],"review_version":1}