{"id":"46ff4299-e09c-49bb-91c5-e1a2dd9457cc","arxiv_id":"2505.16535","paper_version":1,"verdict":"REJECT","confidence":"LOW","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"SHaDe combines explicit tri-plane deformation, SH attention rendering, and latent diffusion refinement to reconstruct dynamic 3D scenes from sparse multi-view images and reports state-of-the-art results on D-NeRF.","lead":"This paper proposes SHaDe, a dynamic 3D scene reconstruction method that uses tri-plane features, a spherical harmonics attention decoder, and a latent diffusion module, and reports better rendering quality than HexPlane and 4D Gaussian Splatting on synthetic benchmarks. A generalist reader might care because dynamic scene reconstruction is central to AR/VR, robotics, and digital twins, and sparse-view robustness would make capture cheaper.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim of a fully explicit, MLP-free deformation field is contradicted by the diffusion refinement path: Sec. 3.3's decoder D(z) outputs the refined offsets used for final rendering, making the motion model a learned network rather than Eq. (2).","rationale":"My check identifies a load-bearing concern different from the reader's weakest_assumption but in the same neighborhood. The reader worried that a fixed W in Eq. (2) cannot adapt; that specific worry is not valid because the tri-plane features fxy, fyz, fxz are learned functions of (x,t) and can carry scene- and time-specific motion information, so even a fixed linear readout can express a rich deformation field. The sharper issue is that the paper's final rendering uses refined offsets produced by the diffusion decoder D(z) in Eq. (6), not the explicit Δ from Eq. (2). This makes the central novelty claim false under the paper's own text: motion modeling is performed by a learned transformer/decoder pipeline, not by the 'fixed, non-learned' projection. It also makes the ablation 'w/o Deformation' difficult to interpret, since the diffusion decoder could learn motion. This internal inconsistency is a correctness risk that is not resolved by filling in W and b. Since the SOTA claim rests on this architecture, and the manuscript provides no code or detailed experimental protocol, the verdict remains REJECT (low confidence). The reader's REJECT is supported, though for a different primary reason.","tokens_in":8658,"tokens_out":6640,"duration_ms":51342,"concrete_test":"Trace the inference computation graph from the released code (or, if the project page provides none, from a precise textual specification): determine whether the volume rendering in Eqs. (3)-(4) receives Δ from Eq. (2) or the refined offsets from Eq. (6). If it uses the refined offsets, inspect D: count its trainable parameters and check for nonlinear activations. If D has any learned nonlinear layers, the deformation field is not explicit/MLP-free and the 'explicit tri-plane deformation' claim is false. Additionally, run an ablation replacing D with the identity and re-evaluate Table 1's T-Rex PSNR; if performance drops, the learned decoder is the effective motion model.","verdict_should_be":"REJECT","load_bearing_attack":"The central architectural claim (Abstract, Sec. 1, Sec. 3.1) is that deformation is computed by a fixed linear projection of learned tri-plane features, Eq. (2), with no MLPs. However, Sec. 3.3 introduces a transformer encoder T that maps (F, Δ, t) to a latent z and a decoder D that produces refined features and offsets (F_hat, Δ_hat) = D(z). The text then states: 'The resulting refined features F_hat, Δ_hat are then used for differentiable volume rendering.' Thus at inference the deformation offsets actually used in rendering are outputs of D, not the explicit Δ of Eq. (2). The architecture of D is unspecified but is a learned decoder (likely containing nonlinear layers), which directly contradicts the 'fully explicit, no MLPs' claim. Even if W and b were fully specified, the deformation field used in the final render is not the fixed projection unless D is the identity, which is not stated. This is an internal inconsistency, not a matter of missing hyperparameters: the method's core contribution is undefined. Additionally, Fig. 2's rendering path uses the original Δ, while Sec. 3.3 says refined offsets are used, so the paper is ambiguous about which deformation is rendered. The reported SOTA conclusions cannot be attributed to the claimed explicit deformation mechanism until this is resolved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SHaDe, a framework for dynamic 3D reconstruction that combines three components: an explicit tri-plane deformation field, a canonical radiance field with spherical-harmonics (SH) attention, and a latent diffusion refinement module. The method is evaluated on synthetic D-NeRF scenes against HexPlane and 4D Gaussian Splatting, with additional ablations on the T-Rex scene. The central claims are that deformation is computed by a fixed, non-learned linear projection of tri-plane features, that SH attention replaces MLP decoders, and that a transformer-guided latent diffusion prior improves robustness under sparse views and out-of-distribution motion. The paper reports state-of-the-art PSNR/SSIM/LPIPS numbers and efficiency measurements. As written, however, several load-bearing architectural statements are internally inconsistent or underspecified, particularly the relationship between the explicit deformation of Eq. (2) and the learned decoder output of Eq. (6), the status of W and b, and the sparse-view evaluation protocol.","tokens_in":8988,"tokens_out":3987,"duration_ms":37186,"significance":"If the claims were fully substantiated, the work would offer a useful direction: compact explicit deformation via tri-planes, an interpretable SH attention radiance head, and diffusion-based latent refinement are all relevant to the dynamic reconstruction community. The reported parameter and memory counts (Table 2) are concrete positives, and the modular ablation structure is sensible. However, the paper provides no code, no error bars, and no reproducible specification of the core projection in Eq. (2), and the final rendering path appears to use the learned decoder output of Sec. 3.3 rather than the explicit offset. These issues prevent verification of the state-of-the-art claim and of the claimed MLP-free deformation mechanism. The significance is therefore conditional on a major clarification and on additional experimental detail.","major_comments":[{"comment":"The deformation offset is defined as Delta x = W(fxy + fyz + fxz) + b with W in R^{3x32}, b in R^3, and the text says W and b are a 'fixed, non-learned linear projection.' The values or initialization of W and b are never given, so the model is not reproducible. Moreover, because the tri-plane features fxy, fyz, and fxz are learned, the resulting offset is a learned function of (x, t) even if W and b are fixed; the statement in Sec. 3.1 that 'the deformation is computed without any learnable components' is therefore misleading and should be corrected or justified.","section":"Sec. 3.1, Eq. (2)"},{"comment":"The paper states at the end of Sec. 3.3 that 'the resulting refined features F_hat, Delta_hat are then used for differentiable volume rendering,' and Eq. (6) defines (F_hat, Delta_hat) = D(z) with a learned decoder D. The rendering path in Fig. 2, however, uses the original offset Delta x from Eq. (2). Unless D is the identity (which is not stated), the deformation actually used for rendering is the output of a learned network, directly contradicting the central claim in the Abstract and Sec. 3.1 that motion is modeled by a fully explicit, MLP-free deformation field. The paper must specify which offset is rendered, give the architecture and training details of D, and reconcile this with the explicit-deformation claim.","section":"Sec. 3.3, Eq. (6) and Fig. 2"},{"comment":"The mapping from tri-plane features to the SH coefficients {c_lm} is not defined. The text says SH coefficients are 'stored in the tri-plane grid,' but it does not specify how they are retrieved, interpolated, or associated with canonical points, nor how the per-band attention weights alpha_lm(d, t) are parameterized beyond saying they come from a 'lightweight MLP.' This MLP also conflicts with the paper's repeated claim of avoiding MLPs; the MLP-free statement should be scoped to the deformation field only, and Eq. (3) should be made fully concrete.","section":"Sec. 3.2, Eq. (3)"},{"comment":"The sparse-view evaluation is not described as a protocol. The paper reports results for 3, 5, 10, and 20 views in Fig. 3 but does not state how those views were selected, whether HexPlane and 4D-GS were retrained under the same sparse-view conditions, or whether the D-NeRF baselines used their default dense-view training. Tables 1-3 also report no standard deviations or number of seeds, so the claimed superiority over HexPlane and 4D-GS is not statistically supported. These details are necessary to evaluate the central claim of robustness to sparse dynamic inputs.","section":"Sec. 4, Fig. 3 and Tables 1-3"},{"comment":"The latent diffusion prior is trained 'using only the same synthetic multi-view data (e.g., D-NeRF)' and evaluated on D-NeRF, so the claimed generalization to out-of-distribution motion is not demonstrated by any held-out OOD experiment. Furthermore, the inference procedure injects controlled Gaussian noise into the initial latent z0 and applies DDIM denoising, but the relationship between this noisy test-time latent and the training-time diffusion objective in Eq. (7) is never analyzed. The paper should either provide OOD experiments or remove the OOD robustness claim.","section":"Sec. 3.3 and Sec. 4"}],"minor_comments":[{"comment":"The method 'Planeoxel' in Sec. 2 should be 'Plenoxel' to match reference [25] and standard terminology.","section":"Related Work and Sec. 3.2"},{"comment":"The phrase 'without any MLPs' in the module listing and 'without any learnable components' in Sec. 3.1 are inconsistent with the MLP for SH attention in Sec. 3.2 and with the learned tri-plane features; the claims should be reworded to state precisely which components contain no MLPs.","section":"Sec. 3.1"},{"comment":"The table reports results on four scenes but does not say which scenes from D-NeRF are included or whether they are the standard test split; this should be stated for reproducibility.","section":"Sec. 4, Table 1"},{"comment":"The ablation 'w/o Deformation' removes the deformation field while retaining a dynamic model, but the paper does not describe how motion is represented in that variant; a one-sentence description of each ablation's architecture would improve interpretability.","section":"Sec. 5, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper has a single author and no code release, and the central architectural inconsistency between Eq. (2) and Eq. (6) is serious enough that the current manuscript cannot be accepted as-is. I do not recommend outright rejection because the inconsistency is potentially fixable by clarifying the role of D, specifying W and b, and adding the missing experimental protocol; however, the authors must also justify why the diffusion prior is not simply memorizing the D-NeRF training set given that both training and evaluation use the same data."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a mixed bag. The specific combination of tri-plane deformation, SH attention, and latent diffusion refinement is new, and the related work is honestly surveyed. The ablation study, if we take the numbers at face value, shows each module contributes, which is a good sign. But I can't take the numbers at face value, and the core method description doesn't hold together.\n\nThe load-bearing problem is the deformation field. Eq. (2) defines the offset as a fixed, non-learned linear projection with W and b. Those constants are never given, and it's unclear whether they are hand-set or learned. If they are fixed, the deformation cannot adapt to scene-specific motion; if they are learned, the claim of a fully explicit, MLP-free motion model is false. Worse, Sec. 3.3 says the decoder D outputs refined offsets that are used for the final volume rendering. That means the offsets actually used at test time are outputs of a learned network, not the fixed projection of Eq. (2). The paper never reconciles this, and Fig. 2 even shows the rendering path using the original Δ while the text says the refined ones are rendered. This is an internal contradiction, not a missing hyperparameter.\n\nThe SH attention module also quietly reintroduces an MLP: the attention weights α_lm are predicted by a lightweight MLP, so the claim that this replaces MLP decoders is overstated. And Eq. (3) is incomplete: it says SH coefficients are stored in the tri-plane grid, but the mapping from interpolated features to coefficients is never specified.\n\nThe experimental section is thin. Tables have no error bars, no seeds, and no protocol for the sparse-view comparison. The diffusion prior is trained on the same D-NeRF data used for evaluation, so the OOD robustness claim is not independently evidenced. No code or data is released, so there is no way to verify the numbers.\n\nThat said, the paper is not nonsense. The modular design is plausible, and the ablations, if reproducible, would support the usefulness of each component. But as written, the method cannot be implemented from the text, and the central architectural claim is contradicted internally.\n\nI would not send this to referees in its current form. The right call is desk reject with an invitation to resubmit after a full rewrite: specify W and b or drop the fixed-projection claim, clarify whether D is identity or learned, define the SH coefficient lookup, release code and data, and provide a real sparse-view protocol. A serious referee could not review this fairly without those basics.","headline":"A coherent three-module architecture with an honest related work section, but the paper's central claim—a fully explicit, MLP-free deformation field—is contradicted by its own inference path, and too many load-bearing details are unspecified to support the SOTA claim.","tokens_in":9481,"tokens_out":2018,"would_cite":false,"duration_ms":19354,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that dynamic 3D reconstruction from sparse multi-view video can be made compact, explicit, and temporally consistent by replacing motion MLPs with a tri-plane offset field, decoding appearance with spherical-harmonics…","keywords":["dynamic 3D reconstruction","tri-plane deformation field","spherical harmonics attention","latent diffusion","sparse-view reconstruction","temporal consistency","4D scene representation","volume rendering"],"falsifier":"Inspect the released code or weights for Eq. (2). If $W$ and $b$ change during training, the 'fixed, non-learned' claim is false; if they are frozen, then a controlled experiment that randomly re-initializes and freezes $W$ and $b$ and retrains the rest of the model on the T-Rex scene should reproduce the reported 35.8 dB PSNR, and if it does not, the deformation field is not a fixed parameter-free component as described.","tokens_in":8453,"feed_emoji":"🎥","tokens_out":13176,"duration_ms":94332,"temperature":0.7,"pith_summary":"The paper attempts to show that dynamic 3D reconstruction from sparse multi-view video does not need learned deformation MLPs or large MLP radiance decoders. Its recipe is to store time-varying features on three orthogonal 2D planes, warp each query point into canonical space with a linear projection of summed plane features, and decode color with a spherical-harmonics attention head, while a latent diffusion module refines the tri-plane latents. If this is right, 4D reconstruction becomes more compact (27.1M parameters, 6.4 GB peak memory) and more robust to sparse inputs, beating HexPlane and 4D Gaussian Splatting on the D-NeRF benchmark in PSNR, SSIM, and LPIPS. The reason to care is that the same components could make high-quality dynamic reconstruction practical with only a handful of cameras.","feed_headline":"Explicit tri-plane warp beats HexPlane and 4D-GS from few views","feed_subtitle":"State-of-the-art D-NeRF quality from 3-5 views, with fewer parameters and lower memory.","key_machinery":"Three mechanisms carry the argument. The explicit tri-plane deformation field: three 256x256x32 time-conditioned planes $F_{xy}$, $F_{yz}$, $F_{xz}$; query features are interpolated, summed, and linearly projected by $\\Delta x = W(f_{xy}+f_{yz}+f_{xz})+b$ and then $x_c = x + \\Delta x$, with no MLP or nonlinearity in the motion path. The SH-attention radiance head: stored SH coefficients up to order $L=4$ are reweighted by learned view- and time-dependent attention weights $\\alpha_{lm}(d,t)$ before composing color as $c(d,t)=\\sum_{l,m} \\alpha_{lm}(d,t)\\, c_{lm}\\, Y_{lm}(d)$, with density coming from a separate tri-plane. The latent diffusion refinement path: tri-plane patches become 768 tokens, a transformer pools them to a 512-dimensional latent, a time-conditioned 3D U-Net denoises that latent with a DDPM objective, a decoder returns refined features and offsets, and a temporal-consistency loss encourages smooth latent evolution.","core_discovery":"The paper's central discovery claim is that a 4D scene can be carried by three time-conditioned 2D feature planes whose summed interpolated features map, through the fixed linear projection of Eq. (2), to a per-point deformation offset, so all motion is an explicit warp into canonical space with no MLP in the motion path. In canonical space, a light SH-attention head computes color as a sum over SH bands with learned view- and time-dependent weights on stored coefficients, and a separate tri-plane gives density. A transformer pools plane patches into a 512-dimensional latent, a DDPM-style diffusion process trained with a DDIM sampler and a temporal-consistency loss refines that latent, and the refined features and offsets feed differentiable volume rendering. On four D-NeRF scenes the paper reports higher PSNR and SSIM and lower LPIPS than HexPlane and 4D-GS, with 27.1M parameters and 6.4 GB peak memory, and retains quality when input views drop to 3-5.","pith_inferences":["The 'no MLP' claim is scoped to the deformation path: the SH attention weights, the transformer, and the diffusion denoiser are themselves learned networks, so the whole system is not MLP-free.","Because $W$ and $b$ in Eq. (2) are never specified, a reader cannot verify that the deformation field is fixed; if these parameters are actually optimized, the central architectural claim is overstated.","The paper's gains are shown only on synthetic D-NeRF scenes, so the sparse-view robustness claim has not yet been tested on real-world video with occlusions; the modular design makes that extension a natural next experiment.","Varying the number of DDIM steps at test time should interpolate between fidelity to input views and smoothness of the learned prior, giving a tunable robustness knob the paper does not quantify."],"forward_implications":["Dynamic reconstruction becomes an explicit, inspectable warp: the motion path is a linear projection of plane features rather than a trained network, which cuts memory and inference cost.","Sparse capture becomes viable: the paper reports strong results from 3-5 views on D-NeRF, where the original D-NeRF method uses 100 or more views.","The SH-attention color head makes appearance interpretable: which SH frequency bands dominate a given view and time can be read directly from the attention weights.","The latent diffusion prior should stabilize fast or ambiguous motion, turning underdetermined tri-plane representations into coherent temporal sequences rather than letting the renderer hallucinate arbitrary geometry."],"supporting_citations":[{"why":"Supplies the tri-plane dynamic-scene representation and serves as the main factorized baseline the paper compares against.","marker":"[1]"},{"why":"Introduces the tri-plane feature representation this work uses for spatial and temporal encoding.","marker":"[2]"},{"why":"Extends tri-planes to space-time and is the reference for time-conditioned plane features.","marker":"[4]"},{"why":"Provides the DDPM denoising objective used to train the latent diffusion module in Eq. (7).","marker":"[6]"},{"why":"Defines the D-NeRF benchmark and the dense-view training regime the paper contrasts with 3-5 view inputs.","marker":"[17]"},{"why":"Gives the DDIM sampler used for test-time denoising of the refined latent.","marker":"[21]"},{"why":"Is the point-based dynamic rendering baseline whose sparse-view limitations the paper claims to overcome.","marker":"[24]"},{"why":"Establishes the practice of storing SH coefficients on voxel grids that the SH-attention head extends to a dynamic, view- and time-conditioned form.","marker":"[25]"}],"fun_headline_variants":["Explicit tri-plane warp beats HexPlane and 4D-GS","No MLP motion: explicit warp for dynamic scenes","Latent diffusion refines deformed tri-planes for 4D","SH attention head and diffusion prior make compact 4D","From 3-5 views: explicit warp and diffusion for 4D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fixed linear projection in Eq. (2) can actually express a scene's motion; the paper never says how the matrix $W$ and bias $b$ are set, and if they are truly fixed constants the deformation field cannot adapt to different scenes, while if they are learned the 'fully explicit, non-learned' claim is false.","fun_headline_variants_meta":{"raw":{"variants":["Explicit tri-plane warp beats HexPlane and 4D-GS","No MLP motion: explicit warp for dynamic scenes","Latent diffusion refines deformed tri-planes for 4D","SH attention head and diffusion prior make compact 4D","From 3-5 views: explicit warp and diffusion for 4D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000833,"raw_usage":{"total_tokens":3677,"prompt_tokens":1027,"completion_tokens":2650,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":2560}},"tokens_in":643,"tokens_out":2650,"duration_ms":16816,"temperature":1.0,"reasoning_tokens":2560,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:58:35.256812+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the released code or weights for Eq. (2). If $W$ and $b$ change during training, the 'fixed, non-learned' claim is false; if they are frozen, then a controlled experiment that randomly re-initializes and freezes $W$ and $b$ and retrains the rest of the model on the T-Rex scene should reproduce the reported 35.8 dB PSNR, and if it does not, the deformation field is not a fixed parameter-free component as described.","supporting_citations":[{"cited_title":"Hexplane: A fast representa- tion for dynamic scenes","cited_arxiv_id":null,"evidence_quote":"Supplies the tri-plane dynamic-scene representation and serves as the main factorized baseline the paper compares against."},{"cited_title":"Chan, Connor Z","cited_arxiv_id":null,"evidence_quote":"Introduces the tri-plane feature representation this work uses for spatial and temporal encoding."},{"cited_title":"K-planes: Explicit radiance fields in space, time, and appearance","cited_arxiv_id":null,"evidence_quote":"Extends tri-planes to space-time and is the reference for time-conditioned plane features."},{"cited_title":"D-nerf: Neural radiance fields for dynamic scenes","cited_arxiv_id":null,"evidence_quote":"Defines the D-NeRF benchmark and the dense-view training regime the paper contrasts with 3-5 view inputs."},{"cited_title":"Denois- ing diffusion implicit models, 2022","cited_arxiv_id":null,"evidence_quote":"Gives the DDIM sampler used for test-time denoising of the refined latent."},{"cited_title":"4d gaussian splatting for real-time dynamic scene render- ing","cited_arxiv_id":null,"evidence_quote":"Is the point-based dynamic rendering baseline whose sparse-view limitations the paper claims to overcome."},{"cited_title":"Plenoxels: Radiance fields without neural networks, 2021","cited_arxiv_id":null,"evidence_quote":"Establishes the practice of storing SH coefficients on voxel grids that the SH-attention head extends to a dynamic, view- and time-conditioned form."}],"review_version":1}