{"id":"6c3247ab-3b6b-47b7-8324-d434328150c7","arxiv_id":"2507.19141","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DASH combines self-supervised static/dynamic decomposition with a multiresolution 4D hash encoder to render dynamic scenes in real time with higher PSNR and LPIPS than previous Gaussian splatting baselines.","lead":"The paper presents DASH, a system that renders dynamic video scenes from novel camera viewpoints in real time by splitting a scene into static and moving parts and encoding the moving parts with a 4D hash representation. It reports higher image quality than prior Gaussian-based dynamic renderers (32.22 dB PSNR on Neural 3D Video, 33.94 dB on Technicolor) at 264 FPS on one 4090 GPU, so it matters for applications that need fast free-viewpoint video.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Decomposition threshold and motion-magnitude reliability are under-specified and could mask failure on complex motions; single-run comparisons leave the claimed gains fragile.","rationale":"The reader correctly identified the decomposition reliability as the weakest assumption. My analysis agrees, and I add that the threshold k and its update schedule are never reported, making the method under-specified for reproduction and for assessing sensitivity. The paper's own ablation (Table 4) shows that removing decomposition ('w/o dec') drops PSNR by 0.57 dB, but this does not test the threshold's robustness; it only tests whether applying 4D hash encoding to the whole scene is worse. The manuscript itself explicitly says 'threshold τ is set at the top k% percentile' without giving k, which is a missing support flagged by my review rule. No internal inconsistency is present, but the central claim's accuracy rests on an unverified hyperparameter. The correct verdict is CONDITIONAL: the paper should be accepted only after the authors release the missing hyperparameters, demonstrate robustness with a sensitivity study (multiple seeds and k values), and provide code that reproduces the reported numbers. I do not find grounds to reject, since the architecture and results are plausible and the code is promised.","tokens_in":13138,"tokens_out":1950,"duration_ms":15957,"concrete_test":"Re-run the full optimization on a scene with mixed slow and fast motion (e.g., Neural 3D Video 'flame_steak' and Technicolor 'Train' or 'Fabien'), varying k across a plausible range (1%, 5%, 10%, 20%) and recomputing the threshold every N iterations, while reporting PSNR/SSIM/LPIPS and the static/dynamic split ratio. If the mean PSNR varies by more than ~0.3 dB across settings or the decomposition ratio changes dramatically, the claim that the method is robust to the threshold is not supported. Additionally, inspect the static/dynamic assignment for slow-moving regions: if a significant fraction of ground-truth dynamic pixels is classified static and remains under-rendered, the decomposition assumption fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that DASH achieves state-of-the-art quality at 264 FPS, resting on the self-supervised decomposition that assigns Gaussians to static or dynamic sets based on a motion-magnitude threshold τ at the top k% percentile of ||Δp|| (Section 4.1). This is load-bearing because all downstream claims—4D hash encoding only on dynamic Gaussians, reduced hash collisions, and the reported gains—depend on correct assignment. The paper never reports k or τ, and the only validation is two qualitative visualizations (Fig. 7), one convergence curve for Ls (Fig. 6), and an indirect ablation (Table 4). The mechanism is fragile in two specific ways. First, the method assumes that the linear-motion magnitude predicted during optimization is a trustworthy proxy for true motion, and that a fixed percentile cleanly separates static from dynamic content. If a slowly moving region has small ||Δp|| early in training, it will be classified as static and locked out of the 4D hash encoder, permanently losing temporal capacity. Conversely, noisy static regions with spurious deformation may be classified dynamic, wasting hash capacity. Second, the training process itself changes the magnitude distribution; the paper does not specify when the threshold is computed or whether it is recomputed, so the effective behavior is under-determined. The reported gains are small (0.5–1 dB over Grid4D and Grid4D+dec) and single-run. Because the decomposition is the key novelty, an unverified threshold directly undermines the central claim for scenes with more complex or heterogeneous motion than the two datasets shown.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes DASH, a dynamic scene rendering method based on 3D Gaussian Splatting. The main novelty is a self-supervised static/dynamic decomposition: during optimization, each Gaussian is assigned to the dynamic set if its predicted linear deformation magnitude ||Δp|| exceeds a top-k-percentile threshold τ, and static Gaussians are constrained to have zero deformation. Dynamics are then encoded by a multiresolution 4D hash encoder followed by a multi-head deformation decoder, and a feature-space smoothness regularizer is added to stabilize deformation. Experiments on Neural 3D Video and Technicolor Light Field report state-of-the-art PSNR/SSIM/LPIPS (32.22 dB mean PSNR on Neural 3D Video, 33.94 dB on Technicolor) and real-time rendering at 264 FPS on an RTX 4090. The paper also applies the decomposition to Grid4D (Grid4D+dec) and includes ablations of decomposition, density control, and smoothness regularization.","tokens_in":13497,"tokens_out":5481,"duration_ms":54298,"significance":"If the reported results are reproducible, DASH would be a meaningful advance in real-time dynamic scene rendering: it offers a simple remedy to the feature-overlap and low-rank limitations of plane- and grid-based deformation encoders, and it demonstrates a credible speed/quality tradeoff against recent Gaussian baselines on two standard real-world benchmarks. Strengths include ablations that support the contribution of each component, the transfer of the decomposition idea to Grid4D as an external sanity check, comparison with official codes of baseline methods, and a public code release. However, the central decomposition mechanism is under-specified and the empirical margins over the strongest baselines are small and reported without variance, so the state-of-the-art claim should be treated as conditional until these points are addressed.","major_comments":[{"comment":"The decomposition threshold τ (the top k% percentile of ||Δp||) is a load-bearing free parameter, but the value of k is never reported and the text does not state when the percentile is computed or whether it is recomputed during optimization. Since the entire accuracy argument depends on which Gaussians are assigned to the 4D hash encoder, the method as described is not fully reproducible. Please report k (and any update schedule), the values of λs and λr, and add a sensitivity study over k (e.g., PSNR versus k, and the resulting fraction of dynamic Gaussians) for at least one scene.","section":"Section 4.1, Fig. 3, Table 4"},{"comment":"The decomposition's correctness is not quantitatively established. The only evidence is the convergence curve of Ls and two qualitative visualizations. Because thresholding uses the model's own deformation predictions, a Gaussian whose motion is underestimated early in training is permanently assigned to the static set and excluded from the 4D hash encoder; this could break on scenes with complex or slow motion. Please add a quantitative robustness study: vary k over a range and report PSNR and the fraction of dynamic Gaussians, and evaluate on a scene (real or synthetic) with known motion or ground-truth masks to measure assignment accuracy.","section":"Section 4.1, Fig. 6, Fig. 7"},{"comment":"The headline gains are small relative to the closest baselines (e.g., 32.22 vs. 31.74 dB on Neural 3D Video; 33.94 vs. 33.04 dB on Technicolor Light Field) and all metrics are single-run, with no standard deviations or repeated training runs. Without variance estimates, it is unclear whether the reported margins exceed optimization noise. Please provide mean ± std over at least three seeds or initializations for the main tables and the ablation, or justify why single runs are sufficient for the state-of-the-art claim.","section":"Section 5.2, Tables 1, 2, and 4"}],"minor_comments":[{"comment":"The last column header reads 'Ours [13]'; the citation [13] belongs to Grid4D and should be removed so the column is simply 'Ours'.","section":"Table 1"},{"comment":"The sentence 'our use COLMAP point clouds from the first frame for initialization' contains a grammatical error; it should be 'we use COLMAP point clouds from the first frame for initialization'.","section":"Section 5.1"},{"comment":"The bit-wise XOR operator is not defined, and the symbol L is reused for the number of resolution levels and for the total loss; please disambiguate the notation.","section":"Equation (8)"},{"comment":"The caption contains a duplicated word: 'and and the horizontal axis' should be 'and the horizontal axis'.","section":"Figure 2 caption"},{"comment":"The implementation details do not report the hash table sizes T_l, the number of levels L, the feature dimension F, the perturbation scale ε, or the values of λc, λs, and λr; these should be listed in the paper or the supplementary material for reproducibility.","section":"Section 5.1, Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"I would ask the authors to provide the exact threshold parameter k and its update schedule, and to add seed variance for the main tables, before the paper is accepted. The manuscript is within scope and the central idea is promising, but the decomposition mechanism needs to be specified and empirically robust in the ways described in the major comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: DASH is a solid systems paper that combines deformable 3D Gaussian splatting, multiresolution 4D hash encoding, a self-supervised static/dynamic decomposition, and a feature-space smoothness regularizer. The genuinely new part is applying a full 4D hash encoder only to the dynamic Gaussians, after splitting the scene by predicted motion magnitude, rather than to the whole scene. On Neural 3D Video and Technicolor Light Field it reports the best PSNR/LPIPS among the compared Gaussian-based methods, at 264 FPS on a single 4090. That is a real, if incremental, subfield-level improvement.\n\nWhat the paper does well: the experiments are mostly run against official baseline codes, the ablations in Table 4 show each component contributes, and the authors validate their decomposition mechanism by plugging it into Grid4D and showing gains (Grid4D+dec). The qualitative figures are consistent with the numbers, the related work is fairly attributed, and the writing is clear. The convergence curve for the static constraint loss and the decomposition visualizations give some evidence that the mechanism behaves as intended.\n\nSoft spots, in proportion: the stress-test concern about the decomposition threshold is partly valid. The paper never reports k or τ, and the split is computed from the model's own predicted deformation magnitudes, so it is not an externally validated ground truth. The text also does not specify when the percentile is computed or whether it is recomputed during training. A slowly moving region that is under-predicted early could be locked into the static set and never receive 4D encoding. That is a genuine gap in the description. However, it is not fatal: the w/o dec ablation shows decomposition helps by about 0.57 dB, and the Grid4D+dec transfer suggests the mechanism has some generality beyond the specific architecture. The single-run comparisons are a minor concern; the margins over Grid4D+dec are 0.48 and 0.90 dB, which could shrink with multiple seeds. Missing comparisons with STG and Masked Space-Time Hash Encoding are a moderate omission, especially since Masked Space-Time Hash Encoding is cited but not evaluated.\n\nWho this is for: people working in dynamic novel view synthesis, 3D Gaussian splatting, or real-time rendering. The paper does not reshape the field conceptually, but it offers a useful, believable accuracy/speed point and a decomposition idea that could transfer to other renderers.\n\nRecommendation: this deserves a serious referee. I would send it to peer review and request revisions: report k, τ, λs, λr, state when the threshold is applied/recomputed, add sensitivity analysis and multiple seeds, and ideally add the missing baselines. The central argument holds up, and I would cite it if the public code reproduces the stated numbers.","headline":"A well-engineered combination of known pieces with state-of-the-art benchmark numbers on two real-world dynamic datasets; the decomposition threshold is under-specified but not a load-bearing flaw, and the paper deserves serious peer review.","tokens_in":14064,"tokens_out":1954,"would_cite":true,"duration_ms":20442,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Self-supervised split lets 4D hash render dynamic scenes at 264 FPS.","keywords":["dynamic scene rendering","4D hash encoding","3D Gaussian splatting","static-dynamic decomposition","self-supervised decomposition","novel view synthesis","real-time rendering","spatio-temporal smoothness regularization"],"falsifier":"Use a synthetic dynamic scene with a known ground-truth mask of which Gaussians truly move. If any Gaussian with nonzero true velocity has $\\|\\Delta p\\|$ below the chosen top-$k$ threshold throughout training, DASH will classify it static and render it frozen, producing a measurable error at that region; observing such misclassification directly disproves the decomposition's reliability.","tokens_in":12942,"feed_emoji":"🎥","tokens_out":8283,"duration_ms":75403,"temperature":0.7,"pith_summary":"The paper sets out to show that real-time rendering of dynamic scenes does not require low-rank factorization of space-time features, and that a full 4D hash encoding can be made practical by applying it only where motion actually happens. It introduces DASH, which first separates each Gaussian into static or dynamic based on the magnitude of its predicted linear deformation, then encodes only the dynamic ones with a multiresolution 4D hash table and decodes position, rotation, and scale changes through three small MLP heads. A spatio-temporal smoothness loss stabilizes the deformation field. On two real-world multi-view video datasets, the method reports mean PSNR of 32.22 dB on Neural 3D Video and 33.94 dB on Technicolor Light Field, at 264 FPS on a single 4090 GPU.","feed_headline":"Self-supervised split lets 4D hash render dynamic scenes at 264 FPS","feed_subtitle":"Static Gaussians skip the 4D hash table, curbing collisions and lifting PSNR to 32.22 dB on Neural 3D Video.","key_machinery":"The load-bearing object is the multiresolution 4D hash encoder, a hash grid that maps a 4D coordinate $(p,t)$ to a feature vector through hashing and quadrilinear interpolation, with geometric progression of resolutions and per-dimension prime factors. Its enabling partner is the self-supervised decomposition: a linear motion model $p(t)=p+t\\Delta p$ whose predicted displacement magnitude $\\|\\Delta p\\|$ is thresholded at the top $k$ percent to label Gaussians dynamic or static, plus the static constraint $L_s=\\frac{1}{N_s}\\sum_{p\\in G_s}\\|\\Delta p\\|$ that pushes static displacements to zero. The third piece is the spatio-temporal smoothness regularization, which penalizes feature differences under small coordinate perturbations so the explicit hash representation does not produce erratic deformations.","core_discovery":"The central discovery is that a dynamic scene can be decomposed into static and dynamic Gaussians without any annotation, by reading each Gaussian's predicted per-unit-time displacement $\\Delta p$ from a 3D hash encoder plus a small MLP, and thresholding at the top $k$ percent of displacement magnitudes. Static Gaussians are then frozen and rendered by standard 3D Gaussian splatting, while dynamic Gaussians are deformed by a 4D multiresolution hash encoder that avoids both the low-rank assumption of plane-based methods and the feature overlap of grid-decomposed variants. The method reports that this split, together with decoupled density control and a feature-space smoothness regularization $L_r = \\|G_{4D}(x) - G_{4D}(x+\\epsilon_x)\\|_2^2$, yields state-of-the-art rendering quality and 264 FPS at $1352\\times 1014$ resolution on a single 4090 GPU.","pith_inferences":["A stress test the paper does not run: scenes dominated by fast, large-displacement motion, where the top-$k$ threshold could lock slow-moving content into the static set at initialization, after which it never receives 4D encoding and stays frozen; quality would drop most on exactly the regions that matter.","The linear-motion descriptor $\\Delta p$ is only a first-order average speed, so trajectories with reversals or high curvature may be under-modeled; extending the model with a second-order term or per-Gaussian confidence would be a natural test.","The decomposition also reads as a compression principle: if static Gaussians require no per-frame deformation state, mostly-static scenes should compress well below the reported storage, and the same split could be reused for editing or animation by moving only the dynamic subgraph.","One could test the decomposition's necessity by measuring the hash-collision rate with and without the split; the paper argues collisions drop because the 4D table covers fewer Gaussians, and that is checkable directly from table statistics."],"forward_implications":["DASH's results imply that explicit 4D hash encoding is a workable alternative to plane-based factorization for dynamic Gaussian splatting, removing the need for the low-rank assumption that causes feature overlap.","Because only dynamic Gaussians consume 4D hash capacity, scenes with large static backgrounds need far fewer hash entries and suffer fewer collisions, which is why the full model runs faster than plane-based baselines while storing 146 MB.","The self-supervised decomposition is a portable component: inserting it into a grid-decomposed 4D encoder (Grid4D+dec) raises that model's PSNR on both datasets, suggesting the static/dynamic split itself, not just the encoder, drives part of the gain.","Real-time 264 FPS rendering at 1352×1014 with about 106.6k Gaussians indicates that dynamic-view synthesis can be fast enough for interactive applications such as VR and film previz."],"supporting_citations":[{"why":"It supplies the base 3D Gaussian representation and optimization, including density control, that DASH modifies.","marker":"[14]"},{"why":"It provides the multiresolution hash encoding scheme that DASH extends from 3D to 4D.","marker":"[25]"},{"why":"It is the plane-based dynamic Gaussian method whose low-rank assumption and feature overlap DASH targets as the main comparison.","marker":"[41]"},{"why":"It is the grid-decomposed 4D hash method that DASH compares against and also augments with its decomposition as Grid4D+dec.","marker":"[13]"},{"why":"It provides the first main real-world multi-view video dataset and evaluation protocol.","marker":"[15]"},{"why":"It provides the second main real-world light-field dataset on which DASH reports 33.94 dB PSNR.","marker":"[29]"}],"fun_headline_variants":["Self-supervised 4D hash split renders dynamic scenes at 264 FPS","No masks needed: 4D hash decomposes scenes for real-time rendering","DASH: self-supervised 4D hash hits 264 FPS for dynamic scenes","4D hash with auto static-dynamic split reaches 264 FPS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the magnitude of the deformation predicted during optimization is a trustworthy measure of true motion, and that one fixed percentile threshold cleanly separates static from dynamic Gaussians; if a moving region is under-predicted at the start, it is frozen as static and never gets the 4D encoding that could fix it.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised 4D hash split renders dynamic scenes at 264 FPS","No masks needed: 4D hash decomposes scenes for real-time rendering","DASH: self-supervised 4D hash hits 264 FPS for dynamic scenes","4D hash with auto static-dynamic split reaches 264 FPS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001072,"raw_usage":{"total_tokens":4494,"prompt_tokens":957,"completion_tokens":3537,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":3451}},"tokens_in":573,"tokens_out":3537,"duration_ms":25280,"temperature":1.0,"reasoning_tokens":3451,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:59:37.859566+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use a synthetic dynamic scene with a known ground-truth mask of which Gaussians truly move. If any Gaussian with nonzero true velocity has $\\|\\Delta p\\|$ below the chosen top-$k$ threshold throughout training, DASH will classify it static and render it frozen, producing a measurable error at that region; observing such misclassification directly disproves the decomposition's reliability.","supporting_citations":[{"cited_title":"3D Gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"It supplies the base 3D Gaussian representation and optimization, including density control, that DASH modifies."},{"cited_title":"Instant neural graphics primitives with a mul- tiresolution hash encoding","cited_arxiv_id":null,"evidence_quote":"It provides the multiresolution hash encoding scheme that DASH extends from 3D to 4D."},{"cited_title":"4d gaussian splatting for real-time dynamic scene rendering","cited_arxiv_id":null,"evidence_quote":"It is the plane-based dynamic Gaussian method whose low-rank assumption and feature overlap DASH targets as the main comparison."},{"cited_title":"Grid4D: 4D decomposed hash encoding for high-fidelity dynamic scene rendering","cited_arxiv_id":null,"evidence_quote":"It is the grid-decomposed 4D hash method that DASH compares against and also augments with its decomposition as Grid4D+dec."},{"cited_title":"Neural 3D video synthesis from multi-view video","cited_arxiv_id":null,"evidence_quote":"It provides the first main real-world multi-view video dataset and evaluation protocol."},{"cited_title":"Dataset and pipeline for multi-view light-field video","cited_arxiv_id":null,"evidence_quote":"It provides the second main real-world light-field dataset on which DASH reports 33.94 dB PSNR."}],"review_version":2}