{"id":"6c7c0cca-c0eb-49ef-80dd-919c2af980d5","arxiv_id":"2505.13215","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A dynamic scene rendering method that tags Gaussians as static or dynamic by their temporal scale, converting static ones to 3D to cut training time.","lead":"This paper introduces a hybrid 3D-4D Gaussian splatting method that converts static scene regions from expensive 4D Gaussians to cheaper 3D Gaussians during training. It reports 3-5x faster training than 4D Gaussian Splatting baselines while keeping rendering quality about the same.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Speedup attribution is confounded: the 4DGS baseline lacks the Taming-3DGS backward pass, no-reset schedule, and 6k-iteration budget; the 5.5h-to-12min gain is not shown to come from static-to-3D conversion.","rationale":"Good-faith reading: the paper is honest about the threshold heuristic and includes a tau ablation; that part of the reader's weakest assumption is real but modest (tau=3 vs tau=3.5 changes PSNR from 32.25 to 31.98) and is acknowledged as a limitation. My concern is the causal claim about speed. The central promise is not just a fast dynamic Gaussian-splatting system, but that removing redundant temporal parameters causes the speedup. Table 1 compares a pipeline that also uses Taming-3DGS, removes opacity resets, and stops at 6K iterations against an original 4DGS baseline with 20K-30K iterations. A reader cannot tell whether the hybrid representation is responsible for the speedup. The magnitude of the wall-clock gap (about 28x) also exceeds what one would predict from the roughly 3.1x Gaussian-count reduction in Table 4 and the 3-5x iteration reduction alone, so the per-iteration cost reduction must come from somewhere; Taming-3DGS is the obvious candidate and is not controlled. Because this is an empirical systems claim, the fix is straightforward: run the baseline with the same optimization stack and schedule, with and without conversion. No code is released, so this control is not available to readers. I would keep the CONDITIONAL verdict, with the condition being the missing speed-attribution control rather than only the tau sensitivity.","tokens_in":15619,"tokens_out":8152,"duration_ms":75308,"concrete_test":"Run the paper's own pipeline on coffee martini with conversion disabled (tau = infinity, so every Gaussian remains 4D) while holding fixed the Taming-3DGS backward pass, the no-opacity-reset policy, the 6,000-iteration budget, and the hardware. Compare wall-clock time and PSNR against both Ours (tau=3) and the original 4DGS baseline. If the tau=infinity run finishes in roughly 12-30 minutes with PSNR near 32, the static-to-3D conversion is not the main driver of the advertised speedup; if it takes multiple hours, the conversion claim is supported. Also report a tau=3 run under the same conditions so the only changed variable is the conversion mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core claim is that converting temporally invariant Gaussians to 3D is what enables 3-5x faster training (Section 1 bullets) and the Table 1 wall-clock gain (11m53s vs 5.5h). The comparison does not isolate this mechanism. Section 5.2 says the implementation 'further leverages the efficient backward pass from Taming-3DGS [33]' and also drops opacity resets; the paper trains 6,000 iterations for 10-second scenes while stating that 4DGS needs 20,000-30,000. The Table 1 baseline 4DGS time is 'measured on our machine,' but there is no indication that it received the Taming-3DGS backward pass, the reset-free schedule, or the shorter budget. Thus Ours and the baseline differ in at least four variables at once: representation conversion, backward-pass implementation, opacity-reset policy, and iteration count. Without a control, the observed speedup cannot be assigned to the hybrid 3D-4D representation. Table 4 ablate tau and opacity resets but do not report training time, so they cannot repair the attribution. This is load-bearing because the strongest advertised advantage is speed.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 3D-4DGS, a hybrid Gaussian splatting representation for dynamic scenes. Training starts with a fully 4D Gaussian model; at each densification stage, Gaussians whose temporal scale exp(s_t) exceeds a threshold tau are classified as static and converted to 3D Gaussians by discarding the temporal mean and rotation components. The rendering pipeline projects both 3D and 4D Gaussians into screen space in a unified CUDA rasterizer. Experiments on the N3V dataset (10-second clips and one 40-second sequence) and the Technicolor dataset report PSNR/SSIM/LPIPS comparable to or better than 4DGS, with substantially reduced training times (e.g., 11m53s versus 5.5h on the 10-second N3V clips). The paper also ablates the threshold tau and the use of opacity resets, and it includes a short Limitations section.","tokens_in":15949,"tokens_out":7947,"duration_ms":74587,"significance":"If the speed advantage were cleanly established, the hybrid representation would be a practically useful contribution to dynamic Gaussian splatting, and the paper has real strengths: it evaluates on standard multi-view dynamic benchmarks, reports consistent metrics including LPIPS, provides ablations and qualitative visualizations, and explicitly acknowledges limitations. However, the headline speed advantage is currently not attributable to the proposed 3D-4D conversion because the comparison changes several implementation variables at once, and the key threshold tau is selected after inspecting trained models on the same datasets on which the method is evaluated. With controlled ablations and a more principled threshold analysis, the contribution could be substantial; in its current form the paper supports an engineering result whose main advertised advantage is under-supported.","major_comments":[{"comment":"The central speed claim is not isolated from other implementation changes. The proposed method is trained for 6,000 iterations on 10-second N3V clips and uses the Taming-3DGS backward pass, while the 4DGS baseline is compared at 20,000-30,000 iterations and is not stated to receive that backward pass or the reset-free schedule. The paper also removes opacity resets in the proposed pipeline. Table 4 ablates tau and opacity resets but does not report training time, so it cannot attribute the 5.5h-to-11m53s gain to the 3D-4D conversion. Please add controlled runs: (a) 4DGS with the Taming-3DGS backward pass, no opacity resets, and 6,000 iterations; (b) the proposed pipeline with conversion disabled, holding all other choices fixed; and report wall-clock times for all Table 4 rows.","section":"§5.2 and Table 1"},{"comment":"The temporal-scale threshold tau is load-bearing and is selected post hoc. The text states that tau was empirically determined based on the distribution of temporal scales in fully trained 4DGS and the characteristics of the target datasets, i.e., after inspecting trained models on the same datasets used for evaluation. The ablation in Table 4 tests only tau=2.5, 3.0, and 3.5 on N3V and does not include Technicolor or the 40-second sequence. Because a wrong tau produces a clear failure mode (e.g., flame steak collapse in Fig. 8), the paper should either provide a principled selection rule or report sensitivity across all datasets and scenes; otherwise the reported averages reflect a hyperparameter tuned to the test sets.","section":"§4.1 and Table 4"},{"comment":"The quaternion conversion is not numerically robust as written. Eq. (9) computes w = 0.5*sqrt(1+tr(R3D)) and then x, y, z with 4w in the denominator; when R3D is a rotation by pi, the trace is -1, w=0, and the expression divides by zero. A standard branch based on the largest quaternion component is needed. Since this conversion is the core operation of the method, the manuscript should specify the robust version or explicitly state an assumption that such rotations do not occur.","section":"§4.2, Eq. (9)"}],"minor_comments":[{"comment":"Typo: 'piplines' should be 'pipelines'.","section":"§4.3"},{"comment":"The ** marker on 4DGS** and 4K4D** is not explained in the caption; clarify which training protocol (e.g., all 300 frames split for training, sparse COLMAP initialization) applies to each baseline and whether the proposed method uses the same protocol.","section":"Table 2"},{"comment":"Training times in Table 3 are measured on an RTX 3090 while Table 1 uses an RTX 4090; the text should state explicitly that wall-clock values across the two tables are not directly comparable.","section":"Tables 1 and 3"},{"comment":"The histogram would be more informative with axis labels and a marked position of tau; as shown, the claimed valley between dynamic and static temporal scales cannot be verified by the reader.","section":"Fig. 2"},{"comment":"There are typographical issues in the related work text, e.g., 'V olumes' for 'Volumes'.","section":"§2.2"},{"comment":"The 4D-to-3D conversion is irreversible: a converted 3D Gaussian is never re-promoted to 4D, even though the iterative classification suggests adaptivity. This design choice is not discussed in the Limitations section and should be acknowledged.","section":"§4.1 and Limitations"}],"recommendation":"major_revision","confidential_remarks":"The main gate for revision is the controlled speed ablation. I would also ask for wall-clock training times in every row of Table 4 and for a clear statement of which codebase and backward pass the 4DGS baseline uses. If the authors can release the modified rasterizer code, the reproducibility of the speed measurements would improve substantially. I do not see grounds for rejection; the direction is reasonable and the quality results are competitive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a genuinely new mechanism—iteratively flipping temporally invariant 4D Gaussians to 3D based on a scale threshold—and it shows real quality gains on N3V and Technicolor. But the headline speedup number is not cleanly attributable to that mechanism, and the main hyperparameter is tuned on the evaluation datasets. With controls, it could be a solid contribution; as written, the central claim is overreached.\n\nWhat is new: the static/dynamic split via temporal scale, with separate densification, is a clean idea and not present in cited 4DGS or Ex4DGS. The paper documents a large reduction in Gaussian count (3.3M to ~1.07M) and consistent PSNR/SSIM/LPIPS improvements over 4DGS. The ablation around tau (Table 4) is honest about the trade-off, and the per-scene curves in the supplement show the threshold sensitivity. The writing is clear and the implementation seems careful.\n\nSoft spots: The speedup claim is load-bearing and confounded. Section 5.2 says the implementation borrows the Taming-3DGS backward pass, drops opacity resets, and trains 6k iterations vs 20-30k for 4DGS. Table 1 compares against a 4DGS baseline that presumably has none of those advantages. So the 11m53s vs 5.5h gap cannot be assigned to the hybrid representation; it could mostly come from the cheaper backward pass, the shorter schedule, or the reset-free policy. Table 4 ablates tau and resets but does not report training time, so it cannot repair the attribution. A control that keeps everything else fixed and toggles only the 4D-to-3D conversion is needed. This is the kind of thing a sharp reviewer will catch immediately.\n\nSecond, tau is selected after inspecting fully trained 4DGS on the same datasets (Sec 4.1). That is hyperparameter fitting to test sets, even if it's a single scalar. The paper should say how tau would be chosen on a new scene without a trained reference.\n\nThird, 4D Scaffold GS [10] is cited but not compared; it also separates static and dynamic content, so at least a discussion or a number is warranted. No code is released.\n\nBottom line: the mechanism is plausible and the quality numbers hold up, but the speed advantage—the main selling point—needs isolation. I'd send this to a serious referee and ask for the control experiment and a test-set-independent rule for tau. It's not a reject-on-sight; it's a revise.","headline":"A genuinely new hybrid 3D-4D Gaussian mechanism with solid quality results, but the headline speedup is confounded by unequal baselines and the tau threshold is tuned on test data.","tokens_in":16423,"tokens_out":3296,"would_cite":true,"duration_ms":29089,"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":"The paper argues that representing static regions with 3D Gaussians and reserving 4D Gaussians for moving content makes dynamic scene training 3-5x faster without losing quality.","keywords":["dynamic scene representation","4D Gaussian splatting","3D Gaussian splatting","static-dynamic classification","temporal scale threshold","novel view synthesis","hybrid representation","training acceleration"],"falsifier":"Train the method on a synthetic scene with one known static background and one known moving object, sweeping $\\tau$ over a fine grid; if no threshold simultaneously keeps the background in 3D and the moving object in 4D while matching 4DGS quality, then the scalar-threshold separation premise fails. A second check: after training, take every converted 3D Gaussian and test whether any region it covers later moves; re-enabling 4D for those regions should recover quality if the one-way conversion is lossy.","tokens_in":15470,"feed_emoji":"🎬","tokens_out":5709,"duration_ms":49890,"temperature":0.7,"pith_summary":"The paper claims that most of the cost of 4D Gaussian Splatting for dynamic scenes is wasted on static regions, which do not need time-varying parameters. Its proposal is a hybrid representation: start with all 4D Gaussians, then repeatedly convert any Gaussian whose time-axis scale exceeds a threshold into a pure 3D Gaussian, so that temporal modeling is spent only on genuinely moving content. The claimed payoffs are concrete: training on the N3V 10-second clips takes about 12 minutes instead of 5.5 hours for the 4DGS baseline, average PSNR rises slightly from 32.01 to 32.25 dB, and storage drops from 2.1 GB to 273 MB. If true, this makes dynamic scene reconstruction practical for interactive and production workflows that currently wait hours per sequence.","feed_headline":"Hybrid 3D-4D Gaussians train 3-5x faster than 4DGS","feed_subtitle":"Keeping 4D parameters only for moving content matches or beats 4DGS quality on dynamic scenes.","key_machinery":"The central object is the 4D Gaussian's time-axis scale $\\exp(s_t)$, a scalar that measures how far the Gaussian stretches along the temporal dimension; a large value means the Gaussian covers the whole sequence and is effectively static. The threshold $\\tau$ (set by hand per dataset: 3 for 10-second N3V clips, 6 for the 40-second sequence, 1 for Technicolor) turns this scale into a binary static/dynamic classifier, applied at every densification stage. The conversion operation discards the temporal components of the mean and rotation, producing a standard 3D Gaussian, and a unified CUDA rasterizer projects both 3D and 4D Gaussians into one screen-space list for compositing. The removal of periodic opacity resets is a second mechanism: keeping opacities continuous avoids erasing learned motion cues during the shortened training schedule.","core_discovery":"On the paper's own terms, the discovery is that temporally invariant Gaussians carry a reliable signature—their time-axis scale $\\exp(s_t)$—and can be collapsed to 3D without harming motion fidelity. The method trains a full 4DGS model for about 500 iterations to stabilize the Gaussians, then classifies at each densification stage: Gaussians with $\\exp(s_t) > \\tau$ are converted to 3D by discarding the temporal mean and the temporal off-diagonal rotation, keeping the spatial mean, the $3\\times3$ rotation, scales, opacity, and spherical-harmonic colors. Dynamic Gaussians keep their full 4D parameters and are sliced at the query timestamp during rendering. Because static 3D Gaussians are updated in every training iteration instead of being culled like many 4D Gaussians, convergence accelerates to roughly 6,000 iterations for 10-second scenes versus 20,000–30,000 for 4DGS, and the reduced parameter count yields the reported speed, memory, and quality results.","pith_inferences":["Because the threshold is hand-set per dataset, a learned or per-scene adaptive classifier could remove the main tuning knob and likely widen the gap over 4DGS; this is testable by optimizing $\\tau$ per scene and comparing average PSNR.","The one-way 4D-to-3D conversion assumes static regions never become dynamic; scenes with moving cameras or newly appearing objects could break that assumption, so a mechanism allowing reclassification back to 4D is a natural extension.","The speedup partly comes from updating static 3D Gaussians every iteration while 4DGS culls many Gaussians; this suggests that smarter per-frame update scheduling could accelerate even full-4D training without any conversion.","Hybrid 3D-4D representations also invite compression: static 3D Gaussians could be encoded with static-scene codecs and dynamic ones with video codecs, potentially cutting storage well below the reported 273 MB."],"forward_implications":["10-second N3V scenes train in about 12 minutes on an RTX 4090, roughly 3–5x faster than the 4DGS baseline, making per-scene tuning practical.","Static regions shed redundant parameters: a typical N3V scene uses 843k 4D Gaussians plus 230k 3D Gaussians instead of 3.3M 4D Gaussians, and storage drops from 2.1 GB to 273 MB.","Rendering quality does not drop with the smaller model: N3V average PSNR is 32.25 dB versus 32.01 dB for 4DGS, with SSIM and LPIPS essentially tied.","The 40-second flame salmon sequence trains in 52 minutes and achieves the lowest LPIPS among compared methods, suggesting the speedup carries to longer capture.","Eliminating opacity resets stabilizes dynamic optimization, avoiding flicker that periodic reinitialization introduces in time-limited training."],"supporting_citations":[{"why":"the 4D Gaussian Splatting baseline that defines the full 4D representation and supplies the codebase and training schedule the paper modifies","marker":"[60]"},{"why":"3D Gaussian Splatting, whose rasterization, densification, and pruning routines are extended to handle hybrid 3D/4D Gaussians","marker":"[18]"},{"why":"the Neural 3D Video dataset that provides the 10-second and 40-second multi-view sequences used for the main comparisons","marker":"[25]"},{"why":"the Technicolor light-field dataset used to validate the method on short, complex multi-camera scenes","marker":"[47]"},{"why":"the fully explicit dynamic Gaussian baseline (Ex4DGS) whose training time and quality are compared against in the main tables","marker":"[22]"},{"why":"Taming-3DGS, whose efficient backward pass is reused to accelerate optimization","marker":"[33]"}],"fun_headline_variants":["3D-4D Gaussians slash dynamic scene training time","Hybrid Gaussians: static parts go 3D, motion stays 4D","Fast dynamic scenes with hybrid 3D-4D Gaussians","Speed up 4DGS by retiring static Gaussians to 3D","3D-4D split makes dynamic Gaussian training 3-5x faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single hand-set number—the threshold on the time-axis scale $\\exp(s_t)$—cleanly separates static from dynamic Gaussians, and that a Gaussian classified as static never needs to become dynamic again; if the threshold is too low, moving content gets folded into the 3D representation and motion collapses, and if it is too high the speedup disappears.","fun_headline_variants_meta":{"raw":{"variants":["3D-4D Gaussians slash dynamic scene training time","Hybrid Gaussians: static parts go 3D, motion stays 4D","Fast dynamic scenes with hybrid 3D-4D Gaussians","Speed up 4DGS by retiring static Gaussians to 3D","3D-4D split makes dynamic Gaussian training 3-5x faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1463,"prompt_tokens":975,"completion_tokens":488,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":389}},"tokens_in":591,"tokens_out":488,"duration_ms":4441,"temperature":1.0,"reasoning_tokens":389,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:16:14.141650+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the method on a synthetic scene with one known static background and one known moving object, sweeping $\\tau$ over a fine grid; if no threshold simultaneously keeps the background in 3D and the moving object in 4D while matching 4DGS quality, then the scalar-threshold separation premise fails. A second check: after training, take every converted 3D Gaussian and test whether any region it covers later moves; re-enabling 4D for those regions should recover quality if the one-way conversion is lossy.","supporting_citations":[{"cited_title":"Neural 3d video synthesis from multi-view video","cited_arxiv_id":null,"evidence_quote":"the Neural 3D Video dataset that provides the 10-second and 40-second multi-view sequences used for the main comparisons"},{"cited_title":"Dataset and pipeline for multi-view light-field video","cited_arxiv_id":null,"evidence_quote":"the Technicolor light-field dataset used to validate the method on short, complex multi-camera scenes"},{"cited_title":"Fully explicit dynamic gaussian splat- ting","cited_arxiv_id":null,"evidence_quote":"the fully explicit dynamic Gaussian baseline (Ex4DGS) whose training time and quality are compared against in the main tables"},{"cited_title":"Taming 3dgs: High-quality radiance fields with limited resources","cited_arxiv_id":null,"evidence_quote":"Taming-3DGS, whose efficient backward pass is reused to accelerate optimization"}],"review_version":1}