{"id":"66c586b5-47fa-4dbf-8971-3355a2c9a3dd","arxiv_id":"2608.02214","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"VARPose applies next-scale visual autoregressive modeling to skeleton data, adding joints to sparse 2D poses so that 3D lifting and mesh recovery become more accurate.","lead":"VARPose trains an autoregressive model to turn a sparse 17-joint 2D skeleton into denser skeletons (48 to 768 joints) using a shared codebook. The extra joints improve downstream 3D pose and mesh predictions by small but consistent margins on standard benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pseudo-node relocation confounds the headline 2D densification metric: the HEM-generated dense skeletons contain an unreported number of joints relocated to the pelvis root, so after root alignment those joints trivially contribute zero error.","rationale":"The reader's weakest assumption identifies the same concern: pseudo-node relocation is not quantified and may deflate the 2D metric while also polluting the downstream input. This is the most load-bearing issue because the paper's headline SOTA number (7.8 px vs HiPART 9.9 px) is computed over a skeleton that Appendix B admits contains synthetic pseudo-nodes, all pinned to the root. The paper's own disclosure in Appendix B, combined with the absence of any count, makes this a genuine confound rather than a speculative one. I also give credit where due: the paper explicitly describes the relocation, provides a bone-interpolation control for downstream lifting (Table 11), and shows transfer to COCO skeletons; these facts limit the damage and keep the paper within a CONDITIONAL verdict rather than a REJECT. However, until the pseudo-node count and an exclusion-based recomputation are provided, the 2D densification SOTA claim should be treated as unverified. Since the reader already conditioned on this issue, my read does not move the verdict; it strengthens the condition.","tokens_in":26290,"tokens_out":7804,"duration_ms":67803,"concrete_test":"Run the released HEM/SMPL coarsening code on Human3.6M to enumerate pseudo-nodes at each granularity (48/96/161/768). Then recompute the Fig. 3 / Table 4 root-aligned 2D Mean Error on the GT test set twice: (i) all joints, (ii) excluding joints whose GT coordinate equals the pelvis root after root alignment (i.e., the relocated pseudo-nodes). If excluding these joints raises 7.81 px by more than ~0.5 px, or if pseudo-nodes are >5% of any skeleton used in the 7.8/9.9 comparison, the SOTA densification claim is confounded. As a secondary check, retrain the MixSTE 161-input experiment with the same number of non-duplicate dense joints (e.g., replace each pseudo-node with the nearest real mesh vertex) to test whether the 0.9 mm downstream gain survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that VARPose's dense poses are anatomically rich enough to beat HiPART by 2.1 px and to improve downstream 3D lifting. The load-bearing condition is that the root-aligned 2D Mean Error measures actual geometric accuracy of the generated dense joints. That condition is not secured. Appendix B states that HEM coarsening of SMPL meshes creates pseudo-nodes and that 'we relocate all pseudo-nodes present to the exact 3D coordinate of the root joint (pelvis) of the corresponding 17-joint ground-truth pose.' The paper never reports how many pseudo-nodes occur in the 48/96/161/768-joint skeletons. After root alignment, a pseudo-node relocated to the pelvis has zero distance to its target if the model emits the root coordinate. The model can learn this trivial mapping, so the reported 7.8 px (and the 9.61/7.81 ablation numbers) are diluted by an unknown fraction of zero-error joints. The same duplicated pelvis coordinates appear in the 161-joint input to MixSTE/GFPose; if they are numerous, at least part of the 0.6–0.9 mm MPJPE gains may reflect repeated uninformative input coordinates rather than 'fine-grained anatomical information.' The released code will allow counting, but as written the SOTA densification claim and the downstream attribution rest on an unquantified artifact. The bone-interpolation ablation (Table 11) partially mitigates the downstream concern, and the COCO 17-joint result shows genuine transfer, but neither quantifies the pseudo-node share.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VARPose, a two-stage framework that adapts visual autoregressive \"next-scale\" modeling to 2D human pose densification. A Granularity-agnostic Pose Tokenizer (GPT) with a shared residual codebook encodes 17/48/96-joint poses into multi-scale discrete token sequences; a unified autoregressive model (UniSkelar) generates denser token sequences in a coarse-to-fine manner. The resulting dense 2D poses are injected into off-the-shelf 3D lifting models via joint-wise concatenation or cross-attention fusion, and are further used for SMPL mesh recovery through SMPL-IKS. The paper reports state-of-the-art 2D densification accuracy (7.8 px vs. 9.9 px for HiPART on Human3.6M), downstream MPJPE improvements on Human3.6M, MPI-INF-3DHP, and 3DPW, and generalization to novel granularities including COCO skeletons. The authors release code and models.","tokens_in":26658,"tokens_out":5892,"duration_ms":49359,"significance":"If the reported results are robust, VARPose is a novel and useful bridge between generative visual modeling and 3D pose estimation. The unified codebook and the representation-projection decoupling are elegant ideas with broad potential, and the paper is commendably thorough: it includes ablations of the tokenizer, quantization densities, codebook size, token expansion, robustness to Gaussian noise, detection errors, motion blur, and cross-dataset transfer. The released code and models strengthen reproducibility. However, the headline 2D densification claim and the downstream attribution are currently weakened by an unquantified artifact in the data-generation process, and the empirical gains are reported without variance estimates. These issues are addressable and do not invalidate the framework's core design, but they must be resolved before the paper can be accepted.","major_comments":[{"comment":"The HEM coarsening pipeline creates pseudo-nodes that, as stated in Appendix B, are \"relocated to the exact 3D coordinate of the root joint (pelvis).\" The manuscript never reports how many pseudo-nodes are present in the 48/96/161/768-joint skeletons. Under the root-aligned 2D Mean Error metric, a pseudo-node relocated to the pelvis has zero target error if the model emits the root coordinate, and a model can learn this trivial mapping. Consequently, the claimed SOTA 7.8 px (and the 9.61/7.81 ablation numbers) may be diluted by an unknown fraction of zero-error joints. The same duplicated pelvis coordinates appear in the 161-joint input to MixSTE/GFPose, so the 0.6--0.9 mm MPJPE gains (Tables 2, 5, 7) could partly reflect repeated uninformative input coordinates rather than genuinely richer anatomical structure. Please report the number of pseudo-nodes per granularity, recompute 2D error","section":"Appendix B; Sec. 4.3; Tables 5 and 7"},{"comment":"All benchmark numbers appear to come from a single run, with no error bars, multi-seed variance, or statistical significance tests. Many claimed gains are small in absolute terms (e.g., Table 3: MPJPE 30.2 to 29.9, PCK +0.1; Table 2: 35.6 to 35.0; Table 12: 40.2 to 39.8). Since GPT and UniSkelar involve stochastic training and sampling, these differences may lie within run-to-run noise. The central claim that densification yields \"tangible performance gains\" requires at least mean +/- std over several seeds for the main comparisons, or an equivalent significance analysis. This is especially important for the 0.3 mm gain on MPI-INF-3DHP and the 0.1--0.4 px differences in ablation tables.","section":"Tables 2, 3, 12; Tables 5--8"}],"minor_comments":[{"comment":"The phrase \"unseen granularities\" is used to describe COCO and 192/384/768-joint skeletons, but for these novel granularities the lightweight decoder is retrained (Sec. 3.3, Sec. 4.2) and the COCO result uses 30% of the training data (Table 10). The text should explicitly state that the representation is frozen while the decoder is adapted, to avoid implying fully zero-shot generalization.","section":"Sec. 4.3 and Fig. 5"},{"comment":"The text says \"a single 48 GB RTX 4090\"; the RTX 4090 has 24 GB of VRAM. Please correct this factual error.","section":"Sec. 4.7"},{"comment":"The row \"17-GT-3D\" is not defined in the main text or table caption. Clarify how it differs from \"17-GT\" (presumably the 3D-lifted intermediate representation).","section":"Table 6"},{"comment":"The condition \"if n = 1 or k > K_{n-1}\" is ambiguous because K_0 is undefined for n = 1. Rewrite with a base-case condition.","section":"Algorithm 1, line 6"},{"comment":"The Domain column contains entries such as \"Z +\" and \"R^{...}\" that are not typeset consistently or defined; please clean up the notation.","section":"Table 1"},{"comment":"The magnitudes of L_Recon and L_VQ are reported as about 6e-7 and 2e-4, respectively, in Appendix E, but these values are not given in the main text. A brief note in Sec. 3.5 would help readers interpret the loss-balancing choice.","section":"Sec. 3.5, Eq. (9)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: VARPose is a legitimate new application of VAR's next-scale idea to pose densification, with a nice unified codebook and some genuinely promising generalization results, but the headline 2D metric is confounded by an unquantified pseudo-node relocation trick, and the downstream gains are small enough that the lack of error bars matters.\n\nThe genuinely new thing here is the granularity-agnostic pose tokenizer with a single residual-quantization codebook, plus an autoregressive model that predicts poses at increasing joint density. That is a real step beyond HiPART, which uses separate codebooks and fixed granularities. The ablation evidence supports the design: unified codebook beats pose-specific codebooks, next-scale beats next-token, and the model's ability to decode unseen granularities (COCO, 192/384/768 joints) with a frozen codebook and tiny decoders is a nice empirical result. The few-shot adaptation numbers (0.27 px extra with 30% data) are also credible. The code release should let people verify this, which is worth credit.\n\nThe soft spots: First, the pseudo-node relocation. Appendix B says all pseudo-nodes from HEM coarsening are moved to the pelvis root coordinate. The paper never says how many pseudo-nodes are in the 48/96/161-joint skeletons. If a meaningful share of the \"dense\" joints are just copies of the root, then the 2D mean error numbers are inflated by trivial zero-error joints, and part of the downstream lifting gain might come from feeding the network repeated pelvis coordinates rather than genuinely richer anatomy.\n\nSecond, no error bars anywhere. The Human3.6M gains are 0.6-0.9 mm, MPI-INF-3DHP is 0.3 mm, CPN is 0.4 mm. Those are within typical run-to-run variance, and several per-action deltas in Table 2 are negative. That doesn't kill the paper's central claim, but it means \"tangible gains\" is overstated until we see variance.\n\nThird, the encoding path for novel granularities is underspecified. The method says encoders stay frozen and only light decoders are retrained, but it's not clear how the frozen encoder, trained on 17/48/96, produces features for a 192-joint pose. The paper says interpolation is used, but the exact procedure is left to the reader.\n\nFourth, the SOTA densification claim rests on a single task-matched baseline (HiPART), reproduced rather than reported. Thin, but not fatal.\n\nWho is this for? People working on pose generation, 3D lifting with richer 2D inputs, or applying VAR-style modeling to non-image structured data. I'd want the authors to fix the pseudo-node counting and add seeds before believing the headline. The core direction is sound and worth referee time.","headline":"Genuine next-scale pose densification with a unified codebook, but the headline 2D metric is confounded by unquantified root-duplicated pseudo-nodes and the downstream gains need error bars.","tokens_in":27176,"tokens_out":3573,"would_cite":true,"duration_ms":30600,"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":"VARPose establishes that 2D pose densification can be reframed as next-scale autoregressive prediction, using one granularity-agnostic codebook to generate arbitrary — even unseen — joint densities and measurably improving 3D pose lifting a","keywords":["human pose estimation","2D pose densification","visual autoregressive modeling","next-scale prediction","granularity-agnostic codebook","3D pose lifting","human mesh recovery","residual quantization"],"falsifier":"Take the coarsened skeletons the paper uses (48, 96, 161, 192, 384, 768 joints) and count how many nodes are HEM pseudo-nodes relocated to the pelvis root. Then recompute the root-aligned 2D mean error and the 3D lifting / HMR deltas either (a) excluding those duplicated root joints or (b) after retraining on a coarsening that avoids forced pseudo-nodes (e.g., direct vertex subsampling). If the 2D error rises substantially above the reported 7.8 px or the downstream MPJPE gains disappear, the 'anatomical densification' claim collapses to a trivial root-repeat artifact.","tokens_in":26137,"feed_emoji":"🦴","tokens_out":7265,"duration_ms":51484,"temperature":0.7,"pith_summary":"VARPose argues that the 'next-scale prediction' idea that powers modern image generation applies directly to human skeletons if we read 'scale' as 'joint density.' To do this it introduces a Granularity-agnostic Pose Tokenizer (GPT), a single residual-quantized codebook shared by skeletons of very different joint counts, and UniSkelar, an autoregressive Transformer that generates the token sequence of each finer density from the sparsest pose. On Human3.6M the model densifies 17-joint inputs to 48/96-joint outputs at 7.8 px mean error, beating the previous hierarchical autoregressive approach (9.9 px), and it generalizes to novel granularities (including a COCO-style 17-joint skeleton and 192/384/768-joint skeletons) with a frozen codebook and retrained decoder. Feeding the densified 161-joint pose into off-the-shelf 3D lifting backbones improves MPJPE by 0.6–0.9 mm on Human3.6M and 0.3 mm on MPI-INF-3DHP, and improves human mesh recovery by 2.1 mm MPJPE-24 — evidence that the densification carries real anatomical information rather than just extra points.","feed_headline":"One codebook turns 17-joint poses into dense 768-joint skeletons","feed_subtitle":"VARPose treats joint density as image scale: 7.8 px densification error, better 3D lifting and mesh recovery.","key_machinery":"Two coupled components carry the argument. (1) GPT — Granularity-agnostic Pose Tokenizer: a single hybrid codebook shared by all pose densities, quantizing via multi-level residual error correction (each level removes the previous levels' quantized residuals) so different granularities land in one shared discrete space; a frozen codebook with retrained decoders handles novel granularities, decoupling representation from topology-specific projection. (2) UniSkelar — Unified Skeletal Autoregressive model: a decoder-only Transformer with rotary position embeddings that predicts the next density level's token sequence in parallel, conditioned on the sparsest 17-joint pose; block-wise causal atte","core_discovery":"VARPose claims that next-scale prediction from visual autoregressive modeling transfers to human skeletons when 'scale' is redefined as skeletal granularity. Its Granularity-agnostic Pose Tokenizer (GPT) learns one codebook (V=4096, D=128) shared across 17-, 48-, and 96-joint poses, using residual quantization with interpolation and an MLP-ResNet to strip away granularity-specific offsets. UniSkelar, a decoder-only Transformer, then predicts the next density level's token map in parallel, conditioned on the sparsest pose, replacing raster-scan next-token generation with global next-scale generation. The reported results: 7.8 px 2D mean error on Human3.6M densification (vs. 9.9 px for the hie","pith_inferences":["If the pseudo-node caveat is small in practice, the same granularity-agnostic codebook idea could transfer to other structured sparse-to-dense domains (hand/face landmarks, animal skeletons, even generic graph densification), since the core claim is that representation can be decoupled from projection.","The paper's downstream gains are modest in absolute terms (0.3–0.9 mm); a natural extension is to test densification where it should matter most — heavy occlusion or extreme poses — and quantify gains as a function of occlusion level, since the current evidence is strongest there.","The authors' own comparison with depth-based input enrichment (AugLift) suggests densification is orthogonal to other 2D cues; combining the two could yield additive gains in domain-shift scenarios like 3DPW, where depth cues generalize better.","A testable prediction of the representation–projection decoupling: a family of granularity-agnostic pose foundation models could be built where adding a new skeleton definition costs only a small projection head plus a few-shot data, as demonstrated by the 30%-data COCO adaptation."],"forward_implications":["Sparse 2D skeletons can be super-resolved to arbitrary joint densities with a single pretrained codebook; a retrained lightweight decoder — or even few-shot adaptation with 30% of data — handles unseen joint definitions (e.g., COCO-style 17-joint, 192/384/768-joint) without retraining the representation.","Densified 2D poses act as a plug-in prior for 3D lifting: modest but consistent MPJPE reductions across multiple backbones and datasets, with the largest relative gains under occlusion (3DPW-Occ, 3.2 mm MPJPE improvement).","Cross-granularity representation is granularity-invariant: codebook entries do not cluster by density, which explains why novel granularities decode without retraining the codebook and why densification error stays nearly flat from 48 to 768 joints.","Next-scale prediction outperforms next-token prediction for skeleton generation (7.81 vs. 9.61 px), indicating that global structural coherence matters more than local token order for human-pose synthesis.","The full pipeline runs near real-time (21.6 ms/frame) because parallel next-scale generation replaces sequential token-by-token decoding, despite higher FLOPs than the previous method."],"fun_headline_variants":["VARPose: autoregressive pose densification improves 3D lifting","Treat joint density as scale: VARPose upsamples skeletons","One codebook for 17 to 768 joints: VARPose's GPT","Coarse-to-fine pose densification with VARPose boosts 3D tasks","VARPose turns sparse 2D poses into dense ones for better 3D"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the dense skeletons used for training and evaluation are genuinely anatomical: the HEM-coarsening pseudo-nodes, which the paper relocates to the pelvis root without reporting their count, are a small enough share of each 48–768 joint skeleton that the reported 2D errors and 3D lifting gains reflect real structure rather than duplicated root coordinates.","fun_headline_variants_meta":{"raw":{"variants":["VARPose: autoregressive pose densification improves 3D lifting","Treat joint density as scale: VARPose upsamples skeletons","One codebook for 17 to 768 joints: VARPose's GPT","Coarse-to-fine pose densification with VARPose boosts 3D tasks","VARPose turns sparse 2D poses into dense ones for better 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1363,"prompt_tokens":826,"completion_tokens":537,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":435}},"tokens_in":570,"tokens_out":537,"duration_ms":4685,"temperature":1.0,"reasoning_tokens":435,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T11:18:12.956375+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the coarsened skeletons the paper uses (48, 96, 161, 192, 384, 768 joints) and count how many nodes are HEM pseudo-nodes relocated to the pelvis root. Then recompute the root-aligned 2D mean error and the 3D lifting / HMR deltas either (a) excluding those duplicated root joints or (b) after retraining on a coarsening that avoids forced pseudo-nodes (e.g., direct vertex subsampling). If the 2D error rises substantially above the reported 7.8 px or the downstream MPJPE gains disappear, the 'anatomical densification' claim collapses to a trivial root-repeat artifact.","supporting_citations":[],"review_version":1}