{"id":"2eef84cf-2f12-49d5-9992-acb1b01d675f","arxiv_id":"2505.06411","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"MAGE reconstructs full-body motion from 3-point HMD tracking by progressively generating 6, 11, then 22 joint poses, outperforming prior single-stage methods on AMASS.","lead":"This paper introduces MAGE, a system that creates a full-body avatar from just the head and wrist tracking data of a VR headset. It builds the body progressively, from a rough 6-part skeleton to the full 22-joint model, and reports improved accuracy and smoothness over previous methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scale-set ablation is confounded with model capacity: full MAGE uses 36 denoiser blocks while the S3-only baseline uses 12, so the coarse-to-fine gain is not isolated.","rationale":"The paper's core proposition is that progressive coarse-to-fine prediction, rather than a larger model, reduces inference uncertainty and error accumulation. Table 3 is the only experiment that directly varies the number of scales, and the paper's own implementation details show each added scale brings 12 extra denoiser blocks. Without parameter and compute matching, the observed gains cannot be assigned to the multi-scale mechanism. Table 4 does not resolve this because it holds capacity constant only across two multi-stage architectures, not against a one-stage baseline. A capacity-matched single-stage test is therefore the decisive check. I do not think this invalidates the paper's general plausibility or its empirical ranking; it makes the central mechanistic claim conditional on an additional experiment, so the CONDITIONAL verdict remains appropriate. My concern differs from the reader's under-specified merging concern; I agree partially because both point at the loose link between the proposed representation and the reported gains, but the capacity confound is a sharper, directly testable evidential gap.","tokens_in":12720,"tokens_out":11554,"duration_ms":123806,"concrete_test":"On AMASS D1, train a single-stage (S3-only) MAGE variant with 36 denoiser blocks (latent dim 512, N=120, T=1000, 4-step DDIM, same loss weighting and training schedule), evaluate across at least three seeds, and compare with the S1, S2, S3 row of Table 3. If the 36-block S3 model closes the gap (e.g., MPJPE and MPJVE within seed noise of 3.21 and 16.71), the multi-scale hierarchy is not the cause of the reported improvement; if it remains clearly worse, the coarse-to-fine claim survives this capacity check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing evidence for the central coarse-to-fine claim is the scale-set ablation in Table 3, but that comparison is confounded with model capacity. Section 4.1 fixes 12 denoiser blocks per stage (a=b=c=12), so the full S1-S2-S3 model contains 36 blocks, whereas the S3-only row—the effective one-stage baseline—contains 12 blocks (S1-S3 and S2-S3 contain 24). The monotone improvement in Table 3 is thus in the same direction as the increase in parameters and compute, and the paper never states that these variants were parameter-matched. The architecture ablation in Table 4 matches total layers, but it compares Sequential versus Gradual—two multi-stage designs—and never compares a capacity-matched direct 3-to-22 single-stage model. Consequently the experiments do not isolate the proposed mechanism: the gains attributed to reducing the inference space and cumulative kinematic errors could instead reflect a larger denoiser stack. This is related to, but distinct from, the under-specified joint-merging rule noted by the reader; it targets the evidential basis for the core assertion directly.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses full-body avatar generation from sparse HMD observations (head and two wrists). The proposed method, MAGE, is a conditional diffusion model that replaces the standard direct 3-to-22-joint mapping with three progressive stages: it first predicts a coarse 6-node composite skeleton, then an 11-node intermediate skeleton, then the full 22-joint SMPL pose; each stage's output is embedded and concatenated with the sparse observations to condition the next stage, and the final training loss is a weighted sum of per-stage regression losses. The authors evaluate on two AMASS splits (D1, D2) and report state-of-the-art or competitive results on MPJRE, MPJPE, MPJVE, and Jitter, together with region-specific errors (Hand/Upper/Lower/Root PE), ablations over the scale set, the multi-stage architecture, and the feature fusion scheme, and a real-time inference claim (0.36 ms/frame with 4-step DDIM). The central claim is that coarse-to-fine factorization reduces the inference space and cumulative kinematic errors, leading to better accuracy and temporal consistency.","tokens_in":12978,"tokens_out":12696,"duration_ms":110040,"significance":"If the central claim holds, the contribution is a useful and well-motivated empirical idea: factorizing an ill-posed 3-to-22 joint mapping through coarse composite skeletons is natural given the SMPL kinematic tree, and the paper supplies consistent ranking improvements on two benchmarks, a cross-distribution generalization test (D2), and three focused ablations. Credit is due for the clear formulation of the stage losses, the attempt to isolate the contribution of the multi-scale design, and the real-time performance report. The main caveat is that the headline mechanism is supported by an ablation (Table 3) that does not control for model capacity, and the coarse skeleton construction (Section 3.2) is not specified at a level that permits reproduction. The method is empirical and supervised by AMASS ground truth, so there is no circularity in the stage losses. The practical significance for AR/VR applications is real, provided the gains survive capacity-matched controls.","major_comments":[{"comment":"The scale-set ablation in Table 3 is confounded with model capacity. Section 4.1 fixes 12 denoiser blocks per stage (a=b=c=12), so the S1+S2+S3 model contains 36 blocks, whereas the S3-only row contains 12 blocks and the S1+S3 and S2+S3 rows contain 24; the improvement from S3-only to S1+S2+S3 is therefore in the same direction as the increase in parameters and compute. The S0+S1+S2+S3 row (48 blocks) shows the benefit is not strictly monotone in capacity, but the decisive comparison remains unperformed: no capacity-matched control (a 36-block S3-only model, or a multi-stage variant with shared weights giving 12 effective blocks) is reported, even though Section 4.4 (Table 4) demonstrates that the authors match total layer counts when they want a fair architectural comparison. As it stands, the experiment attributed to the coarse-to-fine mechanism could equally be explained by a larger denoiser stack.","section":"Section 4.1, Table 3 (scale-set ablation)"},{"comment":"The construction of the composite skeletons is under-specified to the point of non-reproducibility. The text states that adjacent SMPL joints are 'iteratively merge[d]' into the 6-node skeleton S1 and the 11-node skeleton S2, but it never specifies which joints form which composite nodes, nor how a composite node's 6D rotation is derived from the rotations of its member joints. This choice matters: averaging 6D rotation vectors is not a geometrically well-defined rotation average, while taking a member joint's rotation (e.g., the parent) would make the coarse losses L1 and L2 carry no information about the orientations of the merged children. Since L1 and L2 are regression targets that shape everything the later stages can constrain, the explicit 22-to-11-to-6 mapping and the composite rotation (and position) definition must be given.","section":"Section 3.2, Figure 2, Eqs. (6)-(8)"},{"comment":"The quantitative claims in the abstract are not consistently supported by the tables. The stated improvements of 5% MPJRE, 10% MPJVE, and 11% Jitter match D1 only for selected baselines (5.1% MPJRE relative to SAGE; 10.1% MPJVE relative to AGRoL), and no baseline yields the 11% Jitter figure (4.3% relative to SAGE and 13.6% relative to AGRoL on D1; 18.5% relative to SAGE on D2). In addition, several reported advantages are small relative to what a single run can resolve: on D1, MPJPE is 3.21 for MAGE versus 3.28 for SAGE; on D2, MPJRE is 4.26 for MAGE versus 4.30 for both AvatarJLM and AGRoL. No error bars, no number of training seeds, and no significance tests are reported. The claim that MAGE 'significantly outperforms state-of-the-art methods' should be backed by variance information or a significance test on the main comparison tables.","section":"Abstract, Tables 1 and 2"}],"minor_comments":[{"comment":"It is not stated whether the previous stages' outputs S1-hat and S2-hat are detached from the computation graph inside L2 and L3; the gradient-flow decision changes training dynamics and should be specified for reproducibility.","section":"Eqs. (6)-(8)"},{"comment":"Table 1's entry 'Avatorposer' should read 'AvatarPoser'; the Conclusion contains 'obsevations' for 'observations'; and 'V AEHMD' and 'VAE-HMD' are used inconsistently in Section 2.","section":"Table 1, Conclusion, Section 2"},{"comment":"The condition notation is inconsistent: Section 3.1 writes C^{1:N} in R^{N by 18 by M}, while Section 4.1 writes R^{N by 18 by 3}; since M=3, the two agree, but the identification should be stated once.","section":"Sections 3.1 and 4.1"},{"comment":"Figure 2 labels the three scales only by node count; annotating the composite-node groupings (or giving them in the caption) would resolve part of the reproducibility gap noted in Major Comment 2.","section":"Figure 2"},{"comment":"The provenance of baseline numbers should be documented: it should be stated whether all baseline results are taken from the original papers, from re-reported values in later papers, or from the authors' own reimplementation (only the D2 AvatarPoser result is currently attributed).","section":"Section 4.2, Tables 1-2"},{"comment":"The timing measurement behind the 0.36 ms/frame (2778 FPS) claim should state the batch size, sequence length, and whether it includes the overlapping-generation overhead.","section":"Section 4.1"},{"comment":"A few references are incomplete: the Kingma and Welling entry lacks a venue and year, and the Denoising Diffusion Implicit Models entry appears as a title-only citation.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the core idea is attractive, but the empirical foundation for the headline mechanism needs one additional experiment (a capacity-matched single-stage or shared-weight baseline for Table 3) and the method section needs the explicit joint-merging specification before I can support acceptance. I would also request variance reporting on the two main tables, since several differences are in the 0.05-0.1-unit range. The stress-test concern about Table 3 does land, as does the reader's identified weakness in the merging specification; the circularity concern correctly does not apply, since the stage targets come from external AMASS ground truth. The discussion of SAGE, the closest related work, could also be sharpened, given its stratified design."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"MAGE is a plausible step forward for full-body avatar generation from three HMD sensors. Instead of a direct 3-to-22 joint mapping, it denoises progressively through 6, 11, and 22 node representations within one diffusion model. The idea is clean, the implementation is credible, and the paper gives the method a fair shot: consistent top rankings on the D1 benchmark across all four headline metrics, competitive results on the more challenging D2 split, and real-time inference at 2778 FPS. The ablations, while flawed as discussed below, at least point in the direction of the design choices.\n\nThe main soft spot is the evidence for the central claim. Table 3 compares scale combinations, but each stage uses 12 denoiser blocks (a=b=c=12 in Section 4.1), so the full S1-S2-S3 model has 36 blocks while the single-stage S3 row has 12. The monotone improvement in that table tracks the number of blocks, not the hierarchy itself. Table 4 is parameter-matched, but it only contrasts two multi-stage architectures and never tests a capacity-matched one-stage model. So the paper does not actually isolate the proposed coarse-to-fine mechanism. This is a real gap and should be fixed with, say, a 36-block direct 3-to-22 diffusion baseline.\n\nTwo smaller but important issues. First, the construction of the 6- and 11-node composite skeletons is under-specified: 'iteratively merge adjacent joints' is not an algorithm, and the paper never justifies that these particular groupings preserve the kinematic constraints needed by later stages. Second, the abstract's quantitative claims do not match the tables: the MPJVE improvement from SAGE to MAGE is about 19% (20.62 to 16.71), not 10%, and Jitter improves about 4%, not 11%. The D2 MPJPE also trails AvatarJLM, so 'state-of-the-art' needs qualification. No error bars or significance tests are provided, but that is common in this literature.\n\nNone of this is fatal. The core idea is reasonable and the rankings are consistent across two benchmarks. It deserves a serious referee, but the referee should insist on a capacity-matched ablation and a precise description of the merging procedure.","headline":"Useful incremental architecture for sparse avatar generation, but the central coarse-to-fine claim is not isolated from model capacity; requires a parameter-matched ablation and a clearer description of the multi-scale skeletons.","tokens_in":13469,"tokens_out":2662,"would_cite":false,"duration_ms":24449,"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":"MAGE reconstructs full-body motion from three tracked joints by generating a 6-node skeleton first, then 11, then the final 22 joints, and reports that this progressive scheme beats one-stage mapping on accuracy and smoothness.","keywords":["full-body motion generation","sparse observations","head-mounted devices","multi-scale body representation","conditional diffusion model","coarse-to-fine generation","SMPL","motion continuity"],"falsifier":"Run an ablation where the 6- and 11-node skeletons are replaced by a different fixed merging, such as anatomical limb groups or randomly chosen joint clusters; if the coarse-to-fine advantage on MPJRE, MPJVE, and Jitter disappears, the reported gains depend on the exact merging, which is never specified.","tokens_in":12533,"feed_emoji":"🕺","tokens_out":8824,"duration_ms":82326,"temperature":0.7,"pith_summary":"The paper proposes MAGE, a multi-stage diffusion generator that reconstructs full-body SMPL motion from the three joints a head-mounted device observes: the head and two wrists. Instead of a one-stage 3-to-22 joint mapping, MAGE first generates a coarse 6-node body, then an 11-node refinement, then the full 22-joint skeleton, with each earlier stage feeding the next as a condition. The authors claim this coarse-to-fine scheme shrinks the inference space, reduces error accumulation along the kinematic tree, and improves both accuracy and temporal coherence. They report that MAGE outperforms previous methods on the AMASS benchmark across rotation, position, velocity, and jitter metrics, and runs in real time at about 0.36 ms per frame with 4-step sampling. If correct, this gives AR/VR systems a practical way to animate full avatars from the trackers already on a user's head and hands.","feed_headline":"MAGE builds full avatars from 3 trackers by going coarse to fine","feed_subtitle":"A staged diffusion model first infers a 6-node body, then 11, then 22 joints, cutting lower-body error and jitter.","key_machinery":"The multi-scale human motion representation is the central object: three skeletons $S_1$ (6 composite nodes), $S_2$ (11 composite nodes), and $S_3$ (22 SMPL joints), obtained by iteratively merging adjacent joints in the SMPL kinematic tree. Generation runs in reverse, through a single diffusion model divided into three cascaded phases: it outputs $\\hat{S}_1$, then $\\hat{S}_2$ conditioned on $\\hat{S}_1$, then $\\hat{S}_3$ conditioned on $\\hat{S}_2$. The objective is a weighted sum $\\alpha L_1 + \\beta L_2 + \\gamma L_3$ of per-scale losses, and each stage concatenates the sparse observations $C$, the intermediate denoised features $F$, and the re-embedded stage output $F^{\\mathrm{rec}}$. The coarse scales act as global constraints that narrow the inference space for the finer scales.","core_discovery":"The central discovery is a design strategy: full-body motion from sparse HMD observations should not be predicted as one direct 3-to-22 mapping, but progressively from a 6-composite-node skeleton that sets global motion, to an 11-composite-node refinement, to all 22 SMPL joints. Each stage's prediction is fed back as a condition for the next, so later stages see richer constraints and a narrower inference space. In the paper's experiments on AMASS D1, MAGE reports the best MPJRE (2.40 deg), MPJPE (3.21 cm), MPJVE (16.71 cm/s), and Jitter (6.27) among compared methods; on D2 it reports the best MPJRE, MPJVE, and Jitter. The paper interprets these results as evidence that coarse-to-fine generation reduces cumulative kinematic error, especially in the lower body, and balances static accuracy against dynamic continuity.","pith_inferences":["Editorial inference: The coarse-to-fine recipe is not tied to the specific 3-joint input; it could be applied to other sparse tracker layouts, and the paper's own S0 ablation suggests the benefit is not monotonic in the number of scales.","Editorial inference: Because the paper never specifies how SMPL joints are merged into the 6- and 11-node skeletons, a learned or data-driven hierarchy is a natural next step that could make the method less sensitive to the hand-chosen grouping.","Editorial inference: Directly predicting the clean target at each stage, as MAGE does, creates multiple error-correction points, so a failure at a coarse stage need not be fatal if a finer stage can revise it; this is a testable design principle beyond avatar generation."],"forward_implications":["A headset and two handheld controllers are enough to drive a full avatar in real time: MAGE reports about 0.36 ms per frame with 4-step DDIM sampling, far beyond AR/VR frame-rate requirements.","Lower-body prediction, a weak spot of one-stage methods, improves because the coarse 6-node stage fixes global motion before leg details are added; on D1 MAGE reports lower-body position error of 5.93 cm versus 6.01 cm for the next-best compared method.","Generated motions are smoother as well as more accurate: on D1 Jitter drops to 6.27 from 6.55 for the closest compared method, and on D2 MAGE reports 5.81 versus 7.13 for SAGE.","The multi-scale supervision acts as a regularizer, so the model can generalize across the different training and test motion distributions in D2 without overfitting the direct 3-to-22 mapping.","The staged design gives a flexible place to insert extra constraints or priors in future work, since each intermediate representation is itself a supervised prediction that can be checked or guided."],"supporting_citations":[{"why":"Supplies the input feature format (6D rotations, angular and linear velocities of the three observed joints) and the D1 train/test split used for evaluation.","marker":"[Jiang et al., 2022a]"},{"why":"Defines the 22-joint SMPL prediction target by excluding facial and hand joints from the skeleton.","marker":"[Dittadi et al., 2021]"},{"why":"Gives the DDPM forward/reverse process and the noise-prediction formulation underlying the three-stage diffusion objective.","marker":"[Ho et al., 2020]"},{"why":"Supports directly predicting the clean target instead of the noise, which MAGE uses at all three scales.","marker":"[Ramesh et al., 2022]"},{"why":"Provides DDIM sampling, letting MAGE use only 4 denoising steps for real-time inference.","marker":"[Song et al., 2021]"},{"why":"Provides the AMASS archive and the D1/D2 subsets used for training and testing.","marker":"[Mahmood et al., 2019]"},{"why":"Supplies the RepIn time-step embedding used in the denoiser blocks and the AGRoL diffusion baseline that MAGE compares against.","marker":"[Du et al., 2023]"},{"why":"Offers SAGE, the stratified sparse-observation generator that MAGE compares against and reports surpassing on most accuracy and continuity metrics.","marker":"[Feng et al., 2024]"}],"fun_headline_variants":["From 3 trackers to full body: MAGE adds stages, cuts jitter","Coarse-to-fine avatars: MAGE infers 22 joints from 3-tracker HMD","Sparse HMD to full body? MAGE predicts in three stages","Progressive MAGE: 6-node skeleton first, then 11, then 22 joints"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-designed merging of SMPL joints into the 6- and 11-node composite skeletons keeps enough kinematic structure that coarse-stage predictions constrain rather than mislead finer stages, and the paper never states the merging rule.","fun_headline_variants_meta":{"raw":{"variants":["From 3 trackers to full body: MAGE adds stages, cuts jitter","Coarse-to-fine avatars: MAGE infers 22 joints from 3-tracker HMD","Sparse HMD to full body? MAGE predicts in three stages","Progressive MAGE: 6-node skeleton first, then 11, then 22 joints"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000943,"raw_usage":{"total_tokens":4024,"prompt_tokens":934,"completion_tokens":3090,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":2995}},"tokens_in":550,"tokens_out":3090,"duration_ms":18698,"temperature":1.0,"reasoning_tokens":2995,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:43:12.631324+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run an ablation where the 6- and 11-node skeletons are replaced by a different fixed merging, such as anatomical limb groups or randomly chosen joint clusters; if the coarse-to-fine advantage on MPJRE, MPJVE, and Jitter disappears, the reported gains depend on the exact merging, which is never specified.","supporting_citations":[{"cited_title":"Cashman, and Jamie Shotton","cited_arxiv_id":null,"evidence_quote":"Defines the 22-joint SMPL prediction target by excluding facial and hand joints from the skeleton."},{"cited_title":"Hierarchical Text-Conditional Image Generation with CLIP Latents , April 2022","cited_arxiv_id":null,"evidence_quote":"Supports directly predicting the clean target instead of the noise, which MAGE uses at all three scales."},{"cited_title":"DENOISING DIFFUSION IMPLICIT MODELS","cited_arxiv_id":null,"evidence_quote":"Provides DDIM sampling, letting MAGE use only 4 denoising steps for real-time inference."},{"cited_title":"Troje, Gerard Pons-Moll , and Michael Black","cited_arxiv_id":null,"evidence_quote":"Provides the AMASS archive and the D1/D2 subsets used for training and testing."},{"cited_title":"Stratified Avatar Generation from Sparse Observations","cited_arxiv_id":null,"evidence_quote":"Offers SAGE, the stratified sparse-observation generator that MAGE compares against and reports surpassing on most accuracy and continuity metrics."}],"review_version":1}