{"id":"b5c6f607-0d96-4b4d-8889-a336d63c650e","arxiv_id":"2501.19088","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"JGHand is a joint-driven 3D Gaussian Splatting hand model that renders photorealistic hand images in real time, using a zero-error skeleton transform and depth-based shadow simulation to beat prior state-of-the-art on InterHand2.6M and HandCo.","lead":"This paper introduces JGHand, a hand avatar based on 3D Gaussian Splatting that is driven by 3D joint positions, and reports real-time rendering at 0.040 seconds per image with higher PSNR than prior hand avatar methods. It adds a zero-error skeleton transformation and a depth-based shadow simulation to handle finger self-occlusion.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The shadow-simulation function f in Eq. (12) is never defined, and without shadows JGHand falls below LiveHand, so the claimed SOTA gain rests on an unreproducible component.","rationale":"The central claim is that a keypoint-driven 3DGS hand avatar achieves real-time rendering and SOTA quality, enabled by a zero-error skeleton transformation and a depth-based shadow simulation. The zero-error transformation is essentially a deterministic inverse-kinematic computation: B is built from the target joints to reproduce them, so MPJPE=0 is partly by construction. This weakens the novelty of that sub-claim but is not the main risk. The main risk is the shadow module. The ablation quantifies its importance: without it, PSNR drops from 33.44 to 30.92, below LiveHand's 32.32. Therefore the 'surpassing SOTA' conclusion depends entirely on a function that the paper never defines. This is a load-bearing gap, not a preference for more detail. It is addressable: the authors can state f, its parameters, and the blending rule, and validate on a held-out dataset. If they do, the central claim is plausible; the method uses a standard 3DGS pipeline with a triplane feature decoder and Fast-SNARF weights, and the qualitative results are consistent. The reader's conditional verdict captures this correctly, so no verdict change is needed. We agree with the reader's weakest-assumption identification.","tokens_in":12581,"tokens_out":8079,"duration_ms":83632,"concrete_test":"Ask the authors to specify f in Eq. (12) (e.g., a smooth step or exponential falloff with explicit parameters), the sampling radius and offsets, and the blending rule for S with the rendered image. Then re-run the shadow ablation on a held-out split, such as HandCo sequence 0191 or a new InterHand2.6M subject, using the stated f without per-dataset tuning. If the PSNR gain over the no-shadow variant is well below the reported 2.5 dB, or if JGHand no longer exceeds LiveHand on that split, the SOTA claim is dataset-specific and the conditional acceptance should require a defined, transferable shadow model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.3, Eq. (12) defines the shadow mask S_x as the average over N=64 sampling points of f(d(x), d(x+Δx_k)), where f is described only as a differentiable mapping returning values in [0,1]. No functional form, kernel radius, or training procedure is given, and the pixel-wise combination of S with the rendered RGB image is likewise unspecified. This is not a cosmetic omission: Table 4 shows that removing the shadow module changes PSNR from 33.44 to 30.92 on test/capture0, a 2.5 dB drop, while Table 2 lists LiveHand at 32.32 dB and HandAvatar at 30.93 dB on the same split. Thus JGHand's margin over prior SOTA is entirely carried by the shadow module. If f is an arbitrary dataset-tuned depth comparison, the headline claim 'surpassing state-of-the-art methods' is not established for unseen data. The paper's ablation demonstrates that shadows help, but without the definition of f and the combination rule the contribution cannot be reproduced, tested for generalization, or compared fairly. This is a completeness gap in the central argument rather than an internal inconsistency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes JGHand, an animatable 3D Gaussian Splatting hand avatar driven solely by 3D joint keypoints. It introduces a differentiable skeleton transformation from a canonical pose to arbitrary poses and bone lengths, a per-pixel depth-based shadow simulation for finger self-occlusion, and a triplane/MLP-based Gaussian property estimator conditioned on identity and joint angles. Experiments on InterHand2.6M (three sequences) and HandCo (one sequence) report higher PSNR/SSIM and lower LPIPS than HTML, 3D-PSHR, HandAvatar, and LiveHand, with an average inference time of 0.040s per frame. Ablations demonstrate contributions from the skeleton transformation, the shadow module, and the isotropic Gaussian constraint.","tokens_in":12845,"tokens_out":5980,"duration_ms":54445,"significance":"If the method is fully specified and reproducible, it is a meaningful step toward practical keypoint-driven hand avatars: it removes the need for MANO pose/shape parameters at inference, achieves real-time rendering, and outperforms existing neural hand avatar methods on public benchmarks. The use of public datasets, re-trained baselines for most competitors, and systematic ablations are strengths. The central concerns are that the shadow simulation function is under-defined and the 'zero-error' transformation is a construction property rather than an empirical result; both need to be clarified before the reported gains can be independently verified.","major_comments":[{"comment":"The shadow mapping function f in Eq. (12) is never defined; the text only states that it is a differentiable mapping returning values between 0 and 1. The subsequent combination of the shadow mask with the rendered RGB image is described only as \"combined pixel by pixel\" without an explicit formula. This is load-bearing: Table 4 shows that removing the shadow module drops PSNR from 33.44 to 30.92 on test/capture0, which is below the 32.32 dB of LiveHand in Table 2, so the claimed margin over prior state-of-the-art is carried almost entirely by this unspecified component. Please provide the exact form of f (or the depth-difference threshold and kernel), the sampling radius, and the combination rule, and clarify whether any parameters of f are learned or fixed.","section":"Sec. 3.3, Eq. (12)"},{"comment":"The reported MPJPE of 0 for the proposed transformation is a tautology, because the transformation matrix B in Eq. (1) is computed from the target joints J precisely to satisfy J = B J^c. This does not constitute an experimental validation; it is a property of the construction. The meaningful experimental result in Table 1 is the Chamfer distance (0.92 vs 5.14), which measures the LBS-deformed mesh error. Please present the zero joint error as an analytical guarantee rather than an empirical advantage, and keep the mesh-based error as the only experimental metric in this comparison.","section":"Sec. 4.1, Table 1"}],"minor_comments":[{"comment":"The word \"Avater\" should be \"Avatar\" in the title and elsewhere.","section":"Title and throughout"},{"comment":"The depth blending formula in Eq. (11) uses d_j and o_j; please clarify whether d_j is the depth of the Gaussian center or the splatted depth, and how this relates to the standard 3DGS alpha blending.","section":"Sec. 3.3, Eq. (11)"},{"comment":"The reference \"HALO [17]\" in Table 1 should be \"HALO [18]\" to match the bibliography entry for Karunratanakul et al.","section":"Sec. 4.1, Table 1"},{"comment":"The statement that \"we re-trained them on the aforementioned datasets\" is inconsistent with the subsequent note that 3D-PSHR results are referenced from the original paper because code is unavailable; please clarify which methods were re-trained and which were taken from published results.","section":"Sec. 4.2"},{"comment":"The interpolation of poses is described only in the caption; please specify in the text how the interpolation of joint angles is performed (e.g., linear interpolation on angles or spherical linear interpolation on quaternions).","section":"Sec. 4.4 / Fig. 11"},{"comment":"The acknowledged limitation that training requires complete hand texture is important and should be stated in the Introduction or Abstract to set expectations for generalization.","section":"Sec. 5, Limitations"},{"comment":"Typographical errors should be corrected: \"Canoncial Gaussian transformation\", \"the the 3D Gaussian\", \"learning rete\", and \"HandAvater\".","section":"Secs. 3.2, 4.2, Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially interesting for the 3DGS and hand-animation community, but the under-specified shadow module and the tautological zero-error claim need to be addressed before the reported gains can be accepted. I recommend major revision. The comparison to prior work is fair in most respects, but the lack of definition of f makes independent verification impossible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on JGHand. It's a real, working hand-avatar system built from known pieces—3DGS, HALO-style canonicalization, LBS, triplane appearance, SSAO-style shadows—with one genuinely new twist: a differentiable skeleton transformation that handles bone-length scaling and reports zero joint error. The experiments are on public benchmarks, ablations are present, and the speed (0.040s/frame) is believable. The reader's conditional verdict is about right.\n\nWhat's new: the bone-length-aware transformation in Sec 3.1, and the depth-based shadow simulation as a cheap way to model finger self-occlusion. The ablation shows each component helps, especially shadows: removing them drops PSNR from 33.44 to 30.92, more than the gap to LiveHand (32.32). So the shadow module is doing the heavy lifting.\n\nSoft spots, in proportion. The stress-test note is accurate and not a nitpick: Eq. (12) never specifies f(a,b), the kernel radius, or how the shadow mask is combined with the RGB image. The ablation's ~2.5 dB gain rests on an undefined function. That's a completeness gap in the central visual-quality claim, though it's fixable with a definition and maybe a small generalization experiment.\n\nThe zero-error claim is softer than it looks. B is constructed from J so that J = B J^c by design; MPJPE between constructed joints and input joints is near-tautological. The more meaningful numbers are the Chamfer distances (0.92/0.97) and the 83% reduction in 3D LBS error, which suggest the canonicalization is reasonable. I'd ask the authors to present zero-error as a property of the construction rather than an empirical victory, and to report MPJPE on held-out joint positions or mesh vertices.\n\nMinor: the title has a typo (\"Avater\"), the MANUS reference is incomplete, and \"first joint-driven animatable 3DGS hand model\" is likely over-claimed. Implementation details are thin in places (MLP sizes, triplane resolution), but not alarmingly so. To their credit, the limitations paragraph honestly notes the requirement for complete hand texture in training data.\n\nOn the reader's report: I'd nudge significance up slightly—if f is specified and works, this is a useful reference for the hand-avatar subfield. The circularity burden at 4.0 is fair; the invented-entities check is clean.\n\nRecommendation: send it to peer review, but require the authors to define f and the combination rule, release code or detailed hyperparameters, and soften the zero-error wording. Without those, the SOTA claim should not be accepted on the current text.","headline":"A credible, well-ablated hand-avatar pipeline with one genuinely new skeleton transformation, but the shadow module's undefined f carries the entire SOTA margin and needs to be specified before the claim is accepted.","tokens_in":13372,"tokens_out":2519,"would_cite":true,"duration_ms":24625,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45"],"pacs":[],"model":"deepseek-v4-flash","headline":"JGHand establishes that a 3D Gaussian Splatting hand avatar driven only by 3D joint keypoints can render photorealistic hand images in real time, with higher PSNR and SSIM and lower LPIPS than prior state-of-the-art methods.","keywords":["3D Gaussian Splatting","animatable hand avatar","keypoint-driven deformation","skeleton transformation","real-time rendering","self-occlusion shadow simulation","hand pose estimation","neural rendering"],"falsifier":"Take a trained JGHand model, replace the unspecified $f$ in the shadow mask with a fixed monotone function such as $f(a,b)=\\text{sigmoid}(a-b)$, and re-measure PSNR and LPIPS on a held-out hand dataset captured under different lighting; if the visual-quality gap over the no-shadow ablation disappears, the shadow gain depends on the tuned mapping rather than on physical self-occlusion.","tokens_in":12384,"feed_emoji":"🖐","tokens_out":4491,"duration_ms":39197,"temperature":0.7,"pith_summary":"This paper tries to prove that a hand avatar built on 3D Gaussian Splatting can be driven solely by 3D joint keypoints, with no morphable-model shape or pose parameters, while still rendering photorealistic hand images in real time. The authors claim that their zero-error skeleton transformation lets the canonical template deform to any pose and bone length, and that their depth-based shadow simulation captures the self-occlusion shadows fingers cast. If these claims hold, JGHand would give interactive applications an easy-to-drive hand avatar that renders in 0.040 seconds per frame and is quantitatively sharper than existing parametric and neural hand renderers on InterHand2.6M and HandCo.","feed_headline":"Joint keypoints drive a photorealistic hand at 25 fps","feed_subtitle":"JGHand replaces body-model parameters with keypoint-only control, beating four baselines on quality while rendering in 0.040 s.","key_machinery":"Two load-bearing mechanisms carry the argument. The first is the zero-error skeleton transformation that maps canonical joint positions to an arbitrary input skeleton, composed as $B = P K' F' R F K$, where $K$ normalizes bone vectors, $F$ moves them into local frames, $R$ applies hierarchically computed abduction and flexion angles, $F'$ maps back along the kinematic chain, $K'$ rescales to target bone lengths, and $P$ aligns palm planes. The second is per-pixel depth shadow simulation: the renderer produces a depth image by opacity blending, then applies a convolution kernel whose $N$ samples compare each pixel's depth to its neighbors through a differentiable function $f$ to form a shadow mask that is merged into the final image.","core_discovery":"The central claim is that a 3D Gaussian Splatting hand avatar can be driven entirely by 3D joint keypoints, replacing the shape and pose parameters of morphable models, without sacrificing rendering quality or speed. Concretely, JGHand builds a canonical Gaussian template from the MANO mean model, deforms it to the target pose with a zero-error skeleton transformation $B = P K' F' R F K$, skins the Gaussians with Fast-SNARF weights, predicts per-Gaussian appearance from a uvd-coordinate triplane plus pose-aware angular features, and adds a depth-based shadow mask to capture finger self-occlusion. On InterHand2.6M and HandCo it reports higher PSNR and SSIM, lower LPIPS, and faster inference than HTML, 3D-PSHR, HandAvatar, and LiveHand.","pith_inferences":["Because control is keypoint-only, JGHand could tap off-the-shelf 2D and 3D hand keypoint estimators from monocular RGB, removing the need for MANO fitting in avatar pipelines.","The depth-convolution shadow mask is shape-agnostic, so the same idea could be tested on other articulated objects such as feet, ears, or animal paws where self-occlusion shadows matter.","A testable next step the paper leaves implicit is replacing the hand-tuned shadow sampling radius with a learned kernel or conditioning it on camera distance, which could improve generalization across datasets.","The isotropic-Gaussian choice trades a small numerical loss for pose stability, suggesting that a curriculum starting isotropic and switching to anisotropic late in training might recover the gap while keeping stability."],"forward_implications":["A pose-estimation network can drive JGHand directly from keypoint outputs, making end-to-end training with pose estimation a natural next step.","At 0.040 seconds per frame, the avatar supports interactive and real-time applications on a single RTX 3090.","The zero-error transformation handles arbitrary bone lengths, so the same model can be retargeted to hands of different sizes without re-fitting parameters.","Interpolating joint rotation angles yields smooth novel poses, enabling keyframe-based animation from sparse skeletons."],"supporting_citations":[{"why":"Supplies the HALO skeleton-canonicalization method that the zero-error transformation builds on and corrects.","marker":"[18]"},{"why":"Provides 3D Gaussian Splatting, the rendering primitives and rasterization the avatar uses.","marker":"[10]"},{"why":"Supplies the MANO mesh that seeds the canonical Gaussian template and its uv coordinates.","marker":"[17]"},{"why":"Fast-SNARF provides the skinning weight field for smooth LBS deformation of Gaussians.","marker":"[28]"},{"why":"Baseline HandAvatar, a NeRF-based hand avatar, against which rendering quality and speed are compared.","marker":"[5]"},{"why":"Baseline LiveHand, an implicit mesh-based hand renderer, against which speed and quality are compared.","marker":"[8]"},{"why":"InterHand2.6M is the main real multi-view training and evaluation dataset.","marker":"[34]"},{"why":"HandCo supplies the synthetic dataset used for cross-dataset evaluation.","marker":"[35]"},{"why":"SSAO inspires the differentiable per-pixel depth-based shadow simulation.","marker":"[31]"}],"fun_headline_variants":["Keypoint-only control drives photoreal hand at 25 fps","JGHand: joint-driven 3DGS hand beats SOTA in real time","Hand avatars from keypoints, not shape parameters","Joint keypoints replace morphable models in hand rendering","Real-time photoreal hands via keypoint-driven 3D Gaussians"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The depth-based shadow module relies on an unspecified differentiable function $f$ that maps pairs of depth values to shadow weights, so the roughly 2.5 dB PSNR gain it brings in the ablation could be tuned to the training data and may not transfer to other capture conditions.","fun_headline_variants_meta":{"raw":{"variants":["Keypoint-only control drives photoreal hand at 25 fps","JGHand: joint-driven 3DGS hand beats SOTA in real time","Hand avatars from keypoints, not shape parameters","Joint keypoints replace morphable models in hand rendering","Real-time photoreal hands via keypoint-driven 3D Gaussians"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000266,"raw_usage":{"total_tokens":1612,"prompt_tokens":947,"completion_tokens":665,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":577}},"tokens_in":563,"tokens_out":665,"duration_ms":6604,"temperature":1.0,"reasoning_tokens":577,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T21:22:28.090248+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained JGHand model, replace the unspecified $f$ in the shadow mask with a fixed monotone function such as $f(a,b)=\\text{sigmoid}(a-b)$, and re-measure PSNR and LPIPS on a held-out hand dataset captured under different lighting; if the visual-quality gap over the no-shadow ablation disappears, the shadow gain depends on the tuned mapping rather than on physical self-occlusion.","supporting_citations":[{"cited_title":"Screen space ambient occlusion,","cited_arxiv_id":null,"evidence_quote":"SSAO inspires the differentiable per-pixel depth-based shadow simulation."},{"cited_title":"A skeleton-driven neural occupancy representation for articulated hands,","cited_arxiv_id":null,"evidence_quote":"Supplies the HALO skeleton-canonicalization method that the zero-error transformation builds on and corrects."},{"cited_title":"3d gaus- sian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Provides 3D Gaussian Splatting, the rendering primitives and rasterization the avatar uses."},{"cited_title":"Fast-snarf: A fast deformer for articulated neural fields,","cited_arxiv_id":null,"evidence_quote":"Fast-SNARF provides the skinning weight field for smooth LBS deformation of Gaussians."},{"cited_title":"Hand avatar: Free-pose hand animation and rendering from monocular video,","cited_arxiv_id":null,"evidence_quote":"Baseline HandAvatar, a NeRF-based hand avatar, against which rendering quality and speed are compared."},{"cited_title":"Livehand: Real-time and photorealistic neural hand ren- dering,","cited_arxiv_id":null,"evidence_quote":"Baseline LiveHand, an implicit mesh-based hand renderer, against which speed and quality are compared."},{"cited_title":"Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image,","cited_arxiv_id":null,"evidence_quote":"InterHand2.6M is the main real multi-view training and evaluation dataset."},{"cited_title":"Contrastive represen- tation learning for hand shape estimation,","cited_arxiv_id":null,"evidence_quote":"HandCo supplies the synthetic dataset used for cross-dataset evaluation."}],"review_version":1}