{"id":"802c044d-4a85-401b-852d-354b03324e88","arxiv_id":"2505.07333","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"TemPoFast3D reconstructs a 3D clothed human from monocular video at up to 12 FPS by propagating a canonical shape across frames and sampling only visible, surface-adjacent regions.","lead":"A video-speed 3D human reconstruction pipeline that reuses a canonical body shape across frames to cut per-frame computation. It runs pretrained single-image reconstruction networks as a plug-in, reaching up to 12 frames per second while keeping quality close to slower baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline claim 'matches or exceeds state-of-the-art' is unsupported for the monocular setting: the only rows in Table 1 that beat SOTA are three-view variants, while single-view and monocular-video results are worse than the backbones or near the bottom of Table 2.","rationale":"The central claim of the abstract is explicitly about monocular video. The strongest quantitative evidence the paper offers for 'matches or exceeds' (Table 1, TPF3D-GTA-3v / TPF3D-SIFU-3v) comes from Section 3.4's multi-view extension, which requires three synchronized views (0/120/240 degrees). In the monocular rows of the same table, the method underperforms the GTA/SIFU backbones on chamfer distance and normal error (e.g., THuman2.0 Chamfer: 0.5047 vs 0.4409; Normal: 0.0374 vs 0.0342). On monocular video (Table 2), TPF3D-GTA is at 27.60 PSNR and 0.022 LPIPS, below every optimization-based baseline except HumanNeRF on PSNR. The authors themselves note that TPF3D requires multiple views/frames for optimal results. This is not a disagreement with consensus; it is an internal mismatch between the advertised contribution and the reported numbers. The temporal-propagation assumption identified by the reader is also real, but it is secondary: even if propagation worked perfectly, the monocular quality numbers do not support 'matches or exceeds.' A single rerun separating monocular and multi-view rows would settle the overclaim. The speed results and plug-and-play architecture are plausible and partly supported by Table 3, but they do not rescue the central claim.","tokens_in":14229,"tokens_out":7094,"duration_ms":64988,"concrete_test":"Re-run Table 1 with TPF3D-GTA and TPF3D-SIFU in single-view mode using the exact pre-trained GTA/SIFU backbones, and separately recompute the 3-view rows, reporting per-metric deltas against GTA and SIFU on THuman2.0 and CAPE. If the single-view Chamfer and Normal errors remain worse than the backbone errors (as in the current Table 1), then the abstract's 'matches or exceeds state-of-the-art' is not supported for monocular input, and the claim must be restricted to the multi-view setting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is explicitly about monocular video, yet the quantitative evidence does not support it. In Table 1, the monocular single-frame rows TPF3D-GTA and TPF3D-SIFU underperform the GTA/SIFU backbones on most geometry metrics (e.g., THuman2.0 Chamfer: 0.5047 vs 0.4409 for SIFU; Normal error: 0.0374 vs 0.0342). The rows that do exceed SOTA, TPF3D-GTA-3v and TPF3D-SIFU-3v, use three synchronized views from the multi-view extension in Section 3.4, which is outside the advertised monocular setting. On monocular video, Table 2 places TPF3D-GTA at 27.60 PSNR / 0.965 SSIM / 0.022 LPIPS, near the bottom and well below ExAvatar (34.80/0.984/0.009). The paper itself concedes that 'TPF3D requires multiple view/frame for optimal result' and that the single-view result is slightly degraded. Thus the load-bearing condition that the method matches or exceeds SOTA in the monocular setting fails; the favorable geometry numbers are obtained outside the input modality claimed in the abstract.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"TemPoFast3D proposes a plug-and-play framework that wraps SMPL-guided pixel-aligned reconstruction backbones (GTA and SIFU) and temporally propagates a canonical shape representation across frames, aiming to achieve fast 3D clothed human reconstruction from monocular video without per-video optimization. The method introduces coordinate mapping between canonical and posed space, volumetric boundary filtering, visibility-guided and surface-adjacent sampling, and a frame threshold to switch from full reconstruction to efficient inference. Experiments are reported on THuman2.0, CAPE, and NeuMan, with ablations on the proposed optimizations. The paper claims a maximum speed of 12 FPS and that the method matches or exceeds state-of-the-art across standard metrics.","tokens_in":14512,"tokens_out":5689,"duration_ms":49632,"significance":"If the central claims were supported, this work would be a meaningful contribution to real-time 3D human reconstruction: it offers an algorithmic recipe for converting existing single-image pixel-aligned reconstructions into a temporal stream without per-video optimization, and the systematic ablation documents how each optimization affects speed and quality. The plug-and-play design is attractive and the idea of maintaining a propagated canonical shape is sound. However, the evidence in the manuscript does not support the headline claim for the advertised monocular setting: single-view results are worse than the backbone methods, the monocular-video results sit near the bottom of the comparison table, and the only rows that beat state-of-the-art use synchronized multi-view input.","major_comments":[{"comment":"The headline claim that TemPoFast3D \"matches or exceeds state-of-the-art methods\" is not supported for monocular input. In Table 1, the single-view rows TPF3D-GTA and TPF3D-SIFU are worse than their respective backbones on most geometry metrics: on THuman2.0, TPF3D-SIFU has Chamfer 0.5047 versus SIFU† 0.4409 and Normal error 0.0374 versus 0.0342; TPF3D-GTA has Chamfer 0.5247 versus GTA† 0.4625 and Normal error 0.0383 versus 0.0348. The rows that exceed the backbones, TPF3D-GTA-3v and TPF3D-SIFU-3v, use three synchronized views from Section 3.4, which is not the monocular setting promised in the abstract.","section":"Abstract; Table 1; Section 4.1"},{"comment":"The monocular-video results on NeuMan also contradict the \"matches or exceeds\" claim. TPF3D-GTA obtains PSNR 27.60, SSIM 0.965, and LPIPS 0.022, placing it near the bottom of Table 2; for example, ExAvatar achieves 34.80/0.984/0.009 and Vid2Avatar achieves 30.70/0.980/0.014. The manuscript's own phrasing that the method is \"comparable against early optimization-based approaches\" appears to concede this point.","section":"Table 2; Section 4.1"},{"comment":"The paper states that \"fair comparison is not possible as TPF3D requires multiple view/frame for optimal result while other methods only need single image to achieve max quality.\" This admission is load-bearing because it concedes that the favorable numbers in Table 1 come from a setting outside the abstract's monocular-video claim, while the monocular numbers are degraded relative to the backbones.","section":"Section 4.1, first paragraph"},{"comment":"There is an unexplained inconsistency between the reported video-quality numbers. Table 3 reports a baseline PSNR of 32.80 with GTA on the citron sequence and a final optimized PSNR of 31.13, but Table 2 reports TPF3D-GTA at 27.60 on NeuMan for the same method. The manuscript does not reconcile these numbers; if they come from different protocols, the protocol should be described, and if they come from the same protocol, the enormous gap needs an explanation.","section":"Table 2 versus Table 3; Section 4.2"},{"comment":"The hyperparameters K, the frame threshold n, and the surface-adjacent thresholds alpha and beta are selected by tuning on THuman2.0 evaluation tables and by development observations, rather than on a validation split. No sensitivity analysis for alpha and beta is provided on video data. This undermines the \"plug-and-play\" claim and leaves open the possibility that the reported speed-quality tradeoff is partially tuned to the evaluation benchmark.","section":"Supplementary, Tables 4-6; Section 4"}],"minor_comments":[{"comment":"The summation index K is overloaded with the later use of K as the nearest-neighbor count in Section 3.2; please use distinct notation for the number of joints.","section":"Section 3.1, Eq. (3)"},{"comment":"The composition of the per-vertex transformation Tm with the global scale s and translation t is not defined precisely; please clarify whether Tm already includes a global rigid component or whether it must be applied after the per-vertex transform.","section":"Section 3.2.2, Eq. (8)"},{"comment":"The sentence \"We run each methods in the same environment\" should be corrected to \"We run each method in the same environment.\"","section":"Table 2 caption"},{"comment":"The paper emphasizes 12 FPS in the abstract, while Table 2 reports average FPS of 8.900 and Table 3 reports a maximum of 12.301 FPS; please state clearly which reporting convention is used for the headline speed.","section":"Abstract and Tables 2/3"},{"comment":"The statement that FOF \"lacks texture inference capability\" is a strong claim; please provide a citation or a direct comparison, or soften the wording.","section":"Section 2, Related Works"}],"recommendation":"reject","confidential_remarks":"The paper's central claim is not supported by its own tables: the monocular single-image and video results are worse than the backbones or near the bottom of the comparison, and the only favorable geometry numbers come from multi-view input, which is outside the advertised setting. The underlying idea of canonical temporal propagation is interesting, and a revised manuscript that re-scopes the contribution to a speed-quality tradeoff or to multi-view reconstruction, with fair baselines and consistent evaluation, could be reconsidered. I also note that several hyperparameters appear tuned on the evaluation benchmark, which should be addressed in any resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTemPoFast3D is an engineering contribution that wraps existing pixel-aligned reconstructors (GTA, SIFU) in a canonical-space temporal propagation scheme. What's actually new is the inference-side machinery: coordinate mapping from SMPL to canonical space, visibility-guided and surface-adjacent sampling to skip redundant queries, and a frame threshold after which it reuses propagated geometry. That's a sensible system, and the supplementary analysis of K and view count shows the authors did their homework. The speed numbers are real: 12 FPS maximum after a 5-frame warm-up, with stable quality on the citron ablation.\n\nThe soft spot is the headline. The abstract says 'matches or exceeds state-of-the-art' but the monocular evidence says otherwise. On THuman2.0 single-view, TPF3D is worse than the unmodified GTA/SIFU backbones (Chamfer 0.5047 vs 0.4409 for SIFU). The rows that beat SOTA are the three-view variants, which use synchronized cameras — outside the advertised monocular setting. On NeuMan video, TPF3D-GTA sits near the bottom (27.60 PSNR vs ExAvatar 34.80). The paper itself concedes this in Section 4.1. That's a framing problem, not a design flaw; the system's actual niche is fast, approximate reconstruction from a single camera, with quality comparable to early optimization methods, not SOTA.\n\nThe evaluation has room to tighten. No error bars, no code release, which makes the 'plug-and-play' claim hard to verify. Hyperparameters n, alpha, beta, and K are tuned on the benchmark, so the numbers are partly self-selected. One thing the reader flagged — the PSNR gap between Table 2 (27.60) and Table 3 (32.80) — is not a real inconsistency: Table 3 is citron-only, Table 2 averages all four NeuMan sequences. Worth saying in a revision.\n\nBottom line: this is a useful speed-up trick that could matter for interactive applications, but it does not match the advertised quality. I'd send it to peer review with a request for major revision: reframe the claims, release code, add variance estimates, and stop leaning on multi-view results to justify the monocular abstract. The core idea deserves a look, just not with the current wrapping.","headline":"Useful temporal-propagation speed-up for pixel-aligned human reconstruction, but the SOTA quality claim is contradicted by the paper's own monocular results.","tokens_in":15045,"tokens_out":2085,"would_cite":false,"duration_ms":19017,"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":"TemPoFast3D reconstructs 3D clothed humans from monocular video at up to 12 FPS by reusing a canonical shape across frames.","keywords":["3D clothed human reconstruction","monocular video","temporal propagation","canonical space","pixel-aligned implicit function","SMPL","real-time reconstruction","multi-view reconstruction"],"falsifier":"Take a monocular video from a dataset with ground-truth body scans in which the subject wears a loose coat and removes it after frame five. Run TemPoFast3D and a per-frame full-reconstruction baseline on the same frames. If after the garment change the propagated canonical shape keeps the coat's geometry and the chamfer or point-to-surface error jumps far above the baseline's error, the time-invariance assumption that carries the speedup is falsified.","tokens_in":14008,"feed_emoji":"🎥","tokens_out":7680,"duration_ms":72444,"temperature":0.7,"pith_summary":"TemPoFast3D claims that full 3D reconstruction of a clothed human from every video frame is wasteful: because a person's body shape and clothing stay roughly constant over short time spans, a canonical shape learned from the first few frames can be propagated forward and refined with sparse local updates. The paper builds a plug-and-play pipeline around this idea, wrapping existing pixel-aligned reconstruction networks and mapping query points between canonical and posed space through body-model skinning. After a five-frame warm-up, the method queries only visible, surface-adjacent coordinates and propagates colors for occluded vertices, reaching a maximum 12 frames per second on a single GPU while keeping reconstruction metrics close to or better than per-image baselines and competitive with video-optimization methods that take hours. The authors position this as a step toward real-time 3D reconstruction from monocular RGB streams, with the caveat that the whole scheme leans on accurate body-model alignment and near-time-invariant clothing.","feed_headline":"Reuse a canonical body shape to hit 12 FPS reconstruction","feed_subtitle":"After five warm-up frames, the method skips full reconstruction and updates only visible surface regions, preserving quality.","key_machinery":"The load-bearing object is the canonical-space shape representation plus the skinning field used to move between canonical and posed space. A parametric body model supplies per-vertex transformation matrices $T_s$ via linear blend skinning; TemPoFast3D transfers these to arbitrary query coordinates by K-nearest-neighbor association, giving a deformation $x_p = T_x x_c$. A volumetric boundary mask discards query points outside a proximal volume around the canonical body mesh, visibility-guided sampling keeps only coordinates observable from the current view, and surface-adjacent sampling restricts queries to an occupancy band $\\alpha \\le o_c \\le \\beta$. Color for occluded vertices is propagated from the previous frame's canonical vertices by nearest-neighbor lookup. These mechanisms together turn a per-frame dense implicit reconstruction into sparse local updates on a persistent canonical mesh.","core_discovery":"The central claim is that temporal coherence can be converted directly into computation savings without retraining. Instead of re-reconstructing the whole volume each frame, TemPoFast3D maintains a canonical occupancy and appearance representation in a canonical pose, deforms query coordinates into posed space with per-vertex skinning transforms, and after the first five frames treats the propagated canonical shape as a geometric prior. Inference then samples only coordinates that are visible from the current camera and within a narrow occupancy band around the surface, and fills in occluded vertex colors by nearest-neighbor lookup in the previous canonical mesh. The paper reports that applying this wrapper to two pixel-aligned backbones keeps single-frame geometric error on the THuman2.0 and CAPE benchmarks in the same range as the base methods, improves on them in multi-view configurations, reaches 23.25 dB PSNR on CAPE textures, and on the NeuMan zero-shot video benchmark achieves 27.60 PSNR at 8.9 FPS average and 12.3 FPS maximum, compared with hours of training and slower inference for optimization-based baselines. The implied discovery is that the redundant work in per-frame human reconstruction is removable once a stable canonical body shape exists.","pith_inferences":["The fixed warm-up threshold of five frames is a heuristic; an adaptive trigger based on how much the propagated canonical shape changes between frames would likely make the method robust to sudden pose or clothing changes, and that is a testable extension the paper does not explore.","The recipe likely generalizes beyond humans: any articulated object with a parametric model and skinning deformation could receive the same temporal-propagation speedup, though color propagation quality for non-rigid texture is untested.","The reported plateau in multi-view quality suggests the canonical representation, not the number of views, is the bottleneck; further gains would require a richer per-frame appearance model rather than more cameras.","A direct stress test with loose clothing and accessories, where the paper itself expects degraded alignment, would quantify how much of the speed advantage survives outside tight-fitting apparel."],"forward_implications":["Real-time applications such as VR telepresence can reconstruct a clothed person from a single RGB camera at interactive rates without per-video optimization or templates.","Existing body-model-guided pixel-aligned reconstruction networks can be accelerated by swapping in this inference loop, so future improvements to those backbones carry over automatically.","Synchronized multi-view setups can improve geometry by merging canonical representations with no architectural change, with most of the gain coming by three views and diminishing returns after roughly seven views.","Speed has a warm-up cost: the first five frames are full reconstructions, so average throughput over short clips is lower than the 12 FPS maximum reported after warm-up.","Because color is stored per-vertex on the propagated mesh, texture fidelity depends on vertex density and view coverage, giving a concrete target for future appearance refinement.","The method's reported FPS includes both shape and color reconstruction, unlike some baselines that report only rendering speed."],"supporting_citations":[{"why":"Defines the pixel-aligned implicit function that maps 2D pixel features to 3D occupancy, the base reconstruction paradigm TemPoFast3D wraps.","marker":"[30]"},{"why":"Supplies the parametric body model and linear blend skinning transforms used to map between canonical and posed coordinates.","marker":"[20]"},{"why":"One of the two pixel-aligned networks used as a plug-and-play backbone in experiments.","marker":"[43]"},{"why":"The other pixel-aligned backbone, used with its pre-trained weights to demonstrate plug-and-play behavior.","marker":"[44]"},{"why":"Provides the NeuMan video benchmark used for zero-shot video evaluation and serves as an optimization-based baseline.","marker":"[15]"},{"why":"Optimization-based video reconstruction baseline requiring about 97 hours per sequence, used to frame the speed-quality trade-off.","marker":"[7]"},{"why":"Gaussian-splatting avatar baseline with high quality after four hours of training, used to compare reconstruction quality and speed.","marker":"[23]"},{"why":"THuman2.0 dataset used to train the pre-trained backbones and to evaluate single-image and multi-view reconstruction.","marker":"[40]"},{"why":"Used for body-model parameter regression on in-the-wild video frames.","marker":"[41]"}],"fun_headline_variants":["Temporal propagation hits 12 FPS for 3D human reconstruction","Skip full reconstruction: propagate canonical shape for 12 FPS","Canonical shape reuse gives 12 FPS 3D human reconstruction","Temporal coherence enables 12 FPS human reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes the body shape and clothing seen in the first few frames stay essentially unchanged in later frames, and that the estimated body model lines up accurately with the person in every frame.","fun_headline_variants_meta":{"raw":{"variants":["Temporal propagation hits 12 FPS for 3D human reconstruction","Skip full reconstruction: propagate canonical shape for 12 FPS","Canonical shape reuse gives 12 FPS 3D human reconstruction","Temporal coherence enables 12 FPS human reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000976,"raw_usage":{"total_tokens":4146,"prompt_tokens":945,"completion_tokens":3201,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":3127}},"tokens_in":561,"tokens_out":3201,"duration_ms":19175,"temperature":1.0,"reasoning_tokens":3127,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:18:59.017491+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a monocular video from a dataset with ground-truth body scans in which the subject wears a loose coat and removes it after frame five. Run TemPoFast3D and a per-frame full-reconstruction baseline on the same frames. If after the garment change the propagated canonical shape keeps the coat's geometry and the chamfer or point-to-surface error jumps far above the baseline's error, the time-invariance assumption that carries the speedup is falsified.","supporting_citations":[{"cited_title":"PIFu: Pixel-aligned implicit function for high-resolution clothed human digitiza- tion","cited_arxiv_id":null,"evidence_quote":"Defines the pixel-aligned implicit function that maps 2D pixel features to 3D occupancy, the base reconstruction paradigm TemPoFast3D wraps."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the parametric body model and linear blend skinning transforms used to map between canonical and posed coordinates."},{"cited_title":"Global-correlated 3D-decoupling transformer for clothed avatar reconstruction","cited_arxiv_id":null,"evidence_quote":"One of the two pixel-aligned networks used as a plug-and-play backbone in experiments."},{"cited_title":"SIFU: side- view conditioned implicit function for real-world usable clothed human reconstruction","cited_arxiv_id":null,"evidence_quote":"The other pixel-aligned backbone, used with its pre-trained weights to demonstrate plug-and-play behavior."},{"cited_title":"NeuMan: Neural human radiance field from a single video","cited_arxiv_id":null,"evidence_quote":"Provides the NeuMan video benchmark used for zero-shot video evaluation and serves as an optimization-based baseline."},{"cited_title":"Vid2Avatar: 3D avatar reconstruction from videos in the wild via self-supervised scene decomposition","cited_arxiv_id":null,"evidence_quote":"Optimization-based video reconstruction baseline requiring about 97 hours per sequence, used to frame the speed-quality trade-off."},{"cited_title":"Expressive whole-body 3D gaussian avatar","cited_arxiv_id":null,"evidence_quote":"Gaussian-splatting avatar baseline with high quality after four hours of training, used to compare reconstruction quality and speed."},{"cited_title":"Function4D: Real-time human vol- umetric capture from very sparse consumer rgbd sensors","cited_arxiv_id":null,"evidence_quote":"THuman2.0 dataset used to train the pre-trained backbones and to evaluate single-image and multi-view reconstruction."},{"cited_title":"PyMAF: 3D hu- man pose and shape regression with pyramidal mesh align- ment feedback loop","cited_arxiv_id":null,"evidence_quote":"Used for body-model parameter regression on in-the-wild video frames."}],"review_version":1}