{"id":"c4a6fdb9-b32f-4cde-aab8-470c94eda972","arxiv_id":"2505.08196","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ADC-GS uses anchor-driven coarse-to-fine deformation and a rate-distortion optimized entropy model to cut 4DGS storage to roughly 4 to 6 MB while rendering faster than per-Gaussian deformation baselines.","lead":"ADC-GS compresses 4D dynamic scenes by grouping Gaussian primitives around anchor points and deforming only anchors per frame. It reports models of about 4 to 6 MB rendering at over 100 FPS with quality close to prior 4DGS methods, which matters for streaming and real-time 3D video.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 300%-800% FPS advantage rests on baseline numbers that are not tied to a common GPU or resolution; the paper's own LPIPS values also undercut 'without compromising quality'.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: the FPS comparison has no controlled hardware/resolution baseline. The paper is otherwise coherent: the anchor representation, coarse-to-fine deformation, and entropy-based rate-distortion optimization are all described, ablations show monotone improvements, and the storage numbers (4-6 MB) are large relative to baselines. Appendix C explicitly acknowledges a real tradeoff (ADC-GS training time is about 70 minutes vs 35 for E-D3DGS and 20 for 4DGaussian), and the per-scene tables show some quality loss on harder scenes such as HyperNeRF's Banana (27.35 vs 28.98 dB for E-D3DGS at the lowest bitrate). These facts do not invalidate the method, but they mean the central claim should be stated as 'comparable PSNR with much lower storage and potentially faster rendering on matched hardware,' not '300-800% faster without compromising quality.' The missing FPS control is exactly the kind of unverifiable quantitative claim that justifies a conditional verdict; if the authors re-benchmark baselines on the same RTX 3090 and the speed advantage persists, the strongest version of the claim would then be supported.","tokens_in":14712,"tokens_out":12903,"duration_ms":128811,"concrete_test":"Run the official released code of E-D3DGS, DN-4DGS, and 4DGaussian on the same RTX 3090, same preprocessed HyperNeRF/Neu3D data, same render resolution, and the same FPS measurement script used for ADC-GS, reporting per-scene FPS. If the aggregate speed advantage over E-D3DGS drops below roughly 3x on either dataset, the 300%-800% speed claim in the abstract should be withdrawn or qualified. As a secondary check, compute LPIPS at matched PSNR/bitrate to test the 'without compromising rendering quality' phrasing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's headline speed claim (300%-800% faster than per-Gaussian deformation approaches) is the paper's main quantitative novelty, but it depends entirely on the FPS columns in Tables 1 and 2. Section 4.1 specifies the authors' own hardware (single RTX 3090) and training setup, yet it never states the GPU, image resolution, or FPS measurement protocol used for the baselines. The baseline values (E-D3DGS 26 FPS on HyperNeRF and 42 FPS on Neu3D; DN-4DGS 20 and 15 FPS; 4DGaussian 22 and 34 FPS) appear to be taken from the original papers' setups, which used different GPUs and often different render resolutions. Since FPS scales directly with pixel count and GPU generation, the reported 3x-8x ratios could shrink substantially or invert when measured on the same hardware and resolution. This is a missing control, not an accusation of misconduct; without re-benchmarking, the speed part of the central claim is unverified. The claim of no quality compromise is also strained by the paper's own Tables 1 and 2: LPIPS is consistently worse than E-D3DGS (HyperNeRF 0.315/0.278/0.252 vs 0.231; Neu3D 0.066/0.065/0.061 vs 0.030) even where PSNR is comparable. The conditional verdict is appropriate, but the abstract should be narrowed to 'comparable PSNR at 10x smaller storage' until the speed comparison is controlled.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents ADC-GS, a 4D Gaussian splatting representation for dynamic scenes. Instead of deforming each Gaussian independently, ADC-GS organizes Gaussian primitives under a sparse set of anchors in a canonical space. Each anchor predicts K Gaussian primitives; a coarse-to-fine deformation pipeline first deforms the anchor's explicit attributes and then refines per-primitive opacity and color. A multi-dimension entropy model is used in a rate-distortion optimization to compress anchor features, and temporal-significance-based anchor growing and pruning is introduced. Experiments on HyperNeRF and Neu3D report model sizes of roughly 4–7 MB, rendering speeds of 101–135 FPS, and PSNR values within about 0.3–0.6 dB of the best deformation-based baseline (E-D3DGS). The paper includes ablations, per-scene results, and an analysis of bitstream composition.","tokens_in":15065,"tokens_out":5012,"duration_ms":47966,"significance":"If the reported figures are reproducible, ADC-GS would constitute a practically useful step for compact and fast dynamic scene rendering. The anchor-driven deformation idea is simple and plausible, and the component-wise ablation supports the claim that each proposed module contributes to the final trade-off. The code is released, which aids reproducibility. However, the two headline claims—the 300–800% rendering-speed advantage and the statement that quality is preserved without compromise—are not fully supported by the evidence as presented. The speed comparison lacks controlled baseline measurements on comparable hardware, and the LPIPS numbers are consistently worse than E-D3DGS on both datasets. These issues are load-bearing for the abstract and conclusions, so they need to be resolved before the central claims can be accepted.","major_comments":[{"comment":"The FPS comparison for baselines lacks hardware and measurement details. The paper specifies that the authors' model is trained on a single NVIDIA RTX 3090, but it does not state the GPU, rendering resolution, or measurement protocol used for the baseline FPS values (e.g., E-D3DGS 26 FPS on HyperNeRF and 42 FPS on Neu3D). If these numbers are taken from the original papers, which used different GPUs and often different render resolutions, the reported 300–800% speedup is not verifiable. Please re-benchmark all methods on the same GPU, the same resolution, and the same codebase/measurement protocol, or explicitly state the source and conditions of each FPS number and narrow the speed claim accordingly.","section":"Section 4.1, Tables 1 and 2"},{"comment":"The claim of achieving state-of-the-art storage efficiency 'without compromising rendering quality' is contradicted by the LPIPS results in the paper's own tables. On HyperNeRF, the three ADC-GS operating points have LPIPS 0.315, 0.278, and 0.252, versus 0.231 for E-D3DGS; on Neu3D, they are 0.066, 0.065, and 0.061, versus 0.030. This is a consistent and material perceptual-quality gap. Please revise the abstract and conclusions to say that PSNR and SSIM are comparable at reduced storage, and explicitly discuss the LPIPS trade-off.","section":"Abstract and Section 4.2, Tables 1 and 2"},{"comment":"The statement 'up to 200× storage reduction over existing 4DGS methods' is not tied to a specific baseline in the text. Against the deformation-based baselines in Table 1, the reduction is approximately 11–16× (e.g., 47 MB for E-D3DGS versus 4.02 MB for ours). Against Real-Time4DGS in Table 2, the reduction is about 250×. Please specify which baseline yields the 200× figure and report per-baseline size ratios, so the claim is unambiguous and reproducible.","section":"Section 4.2 and contribution list"}],"minor_comments":[{"comment":"The rate expression uses the notation 'MY' instead of a summation over the M chunks; please fix the typography and define all variables (e.g., the meaning of the product/sum over chunks).","section":"Section 3.4, Eq. (9)"},{"comment":"The temporal significance Ψ(k,t) is introduced in the main text, but its precise definition (the alpha-blending weight from Eq. (12) in Appendix A) appears only later. Please state the definition when the formula is first used.","section":"Section 3.5, Eq. (11)"},{"comment":"The per-scene tables report PSNR and size but not SSIM or LPIPS; consider adding these metrics so that the per-scene behavior of perceptual quality can be assessed.","section":"Section 4.2, Table 6 and Table 7"},{"comment":"The coding time is reported as '1.27 seconds and 0.83 seconds under λe = e−2' without specifying the scene, resolution, or bitrate; please provide the experimental context.","section":"Appendix C"},{"comment":"The heading 'Ablation studies on the proportion of Gaussian primitives' is misleading; the parameter K is the number of primitives per anchor, not a proportion. Please rephrase.","section":"Tables 4 and 5"},{"comment":"The text says 'λssim donates weighting coefficients'; 'donates' should be 'denotes'.","section":"Section 3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is well structured and the code release is appreciated. The central risk is the uncontrolled FPS comparison; please insist on a re-benchmark on matched hardware and resolution before acceptance. The LPIPS gap is not catastrophic, but it should be acknowledged honestly in the abstract and conclusions. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is real: instead of deforming every Gaussian separately, deform a sparse set of anchors and let the K associated primitives follow. That cuts deformation redundancy and, combined with a multi-dimension entropy model, shrinks storage to about 4-6 MB per scene while keeping PSNR within a few tenths of a dB of E-D3DGS. The ablations show each module earns its keep, and the code is released. This is a genuine step for 4DGS compression, and the authors deserve credit for building it and reporting the trade-offs.\n\nThe soft spots are mostly about the claims, not the method. The '300%-800% faster' headline rests on FPS columns in Tables 1 and 2, but Section 4.1 only specifies their own hardware (RTX 3090), not the baselines' GPUs or measurement protocol. FPS scales directly with render resolution and GPU generation, so those ratios could shrink a lot on a common testbed. That is a missing control, not a sign of fraud, but it means the speed part of the central claim is unverified until re-benchmarked.\n\nThe other overclaim is 'without compromising rendering quality.' Their own numbers contradict that: LPIPS is consistently worse than E-D3DGS on both datasets (e.g., 0.315 vs 0.231 on HyperNeRF, 0.066 vs 0.030 on Neu3D). PSNR and SSIM are roughly comparable, but the perceptual metric says otherwise. The abstract should say 'comparable PSNR at 10x smaller storage' and leave speed for after a controlled comparison.\n\nMinor issues: no variance reporting across runs, and the training time roughly doubles (70 min vs 35 min for E-D3DGS), which the appendix discloses honestly. The 'up to 200x' reduction in the contributions is not clearly mapped to the tables, which show 10-15x against deformation baselines.\n\nWho gets value from this: anyone working on 4DGS compression, streaming, or real-time dynamic scene rendering. It deserves a serious referee, but with a required revision: benchmark FPS on the same GPU and resolution for all methods, and tone down the abstract to match the actual numbers. I would engage with it, and I would cite it once the speed claim is under control.","headline":"Solid system paper on anchor-driven 4DGS compression with a genuine storage win, but the headline FPS claim needs controlled benchmarking before I'd trust the speed numbers.","tokens_in":15591,"tokens_out":2053,"would_cite":true,"duration_ms":22125,"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":"The paper claims that deforming a sparse set of anchors instead of every Gaussian yields 4D scenes at 4-6 MB and 100+ FPS with little quality loss.","keywords":["4D Gaussian splatting","dynamic scene reconstruction","anchor-based representation","coarse-to-fine deformation","rate-distortion optimization","entropy model","view synthesis","Gaussian compression"],"falsifier":"Run the per-Gaussian deformation baseline and ADC-GS on the same GPU, same evaluation frames, and same timing protocol, then check whether ADC-GS is still several times faster at matched model size and quality; if the gap disappears, the central speed claim is false.","tokens_in":14536,"feed_emoji":"🎥","tokens_out":5596,"duration_ms":51313,"temperature":0.7,"pith_summary":"ADC-GS claims that dynamic scenes can be reconstructed and rendered far more cheaply than existing 4D Gaussian splatting methods by replacing per-Gaussian deformation with anchor-driven deformation. The paper organizes Gaussian primitives under sparse anchors in a canonical space, deforms only the anchors at each frame, and refines per-primitive opacity and color in a fine stage. A multi-dimension entropy model estimates bitrate and trains the whole representation under a rate-distortion loss, shrinking models to about 4-6 MB while rendering at 101-135 FPS on the tested benchmarks. The payoff, if true, is that high-quality dynamic free-viewpoint video becomes practical to store, transmit, and play in real time.","feed_headline":"4D scenes fit in ~5 MB and render at 135 FPS","feed_subtitle":"Deforming sparse anchors instead of every Gaussian cuts storage 10x while keeping PSNR within ~0.6 dB.","key_machinery":"The central object is the anchor representation: a sparse set of anchors, each associated with K neural Gaussian primitives and stored as a compact residual form with a reference feature, a residual feature, and explicit Gaussian attributes. The coarse-to-fine deformation pipeline uses a tiny MLP to deform anchor attributes and a second MLP to refine per-primitive opacity and color, so the number of deformations per frame depends on the anchor count rather than the full Gaussian count. The multi-dimension entropy model estimates bitrate from adaptively quantized features using a hyperprior and contextual chunk-wise prediction, enabling rate-distortion optimization. The temporal-significance refinement grows and prunes anchors based on gradient-weighted rendering weights and accumulated opacity, addressing under- and over-reconstruction in dynamic scenes.","core_discovery":"ADC-GS establishes that the redundancy among neighboring Gaussian primitives in 4D Gaussian splatting is large enough that motion can be modeled by deforming a sparse set of anchors rather than every Gaussian. In the canonical space at time zero, each anchor stores a shared reference feature, per-primitive residual features, and explicit position, covariance, and color attributes, from which K Gaussian primitives are generated. At each target frame, a coarse stage deforms the anchors and thereby updates all associated primitives, while a fine stage adjusts opacity and color per primitive. Combined with temporal-significance-based anchor growing and pruning and rate-distortion optimization, this yields models of roughly 4-6 MB with PSNR within about 0.3-0.6 dB of the strongest deformable baseline while rendering 3-8 times faster.","pith_inferences":["If anchor-driven deformation generalizes, the same structure could be applied to monocular or sparse-view dynamic scenes, where per-Gaussian deformation cost is a known bottleneck.","The temporal-significance weighting used for anchor growing could be repurposed as a saliency prior for allocating bits to motion-heavy regions in dynamic scene compression.","Because anchors are shared across all frames, the method suggests model size may stay roughly constant as sequence length grows; a direct test on longer captures would confirm whether storage scales sublinearly."],"forward_implications":["Models shrink to roughly 4-6 MB per scene, a 10x or larger reduction versus deformation-based 4DGS, making dynamic scenes practical to store and transmit.","Rendering reaches 101-135 FPS on the tested multi-view datasets, fast enough for interactive playback on a single consumer GPU.","Varying the Lagrange multiplier traces a rate-distortion curve, so the same representation covers a range of storage and quality trade-offs.","Quality stays within roughly 0.3-0.6 dB PSNR of the strongest per-Gaussian deformation baseline, so the compression does not come at a large fidelity cost."],"supporting_citations":[{"why":"Supplies the Gaussian splatting renderer and the L1 and SSIM losses used for all rendering and training.","marker":"[Kerbl et al., 2023]"},{"why":"Defines the per-Gaussian embedding deformation baseline whose speed and storage ADC-GS is compared against, and provides the local-similarity observation motivating anchors.","marker":"[Bae et al., 2025]"},{"why":"Provides the anchor-to-primitive generation structure that ADC-GS adapts to build its canonical space.","marker":"[Lu et al., 2024b]"},{"why":"Provides the variational hyperprior and uniform-noise quantization approach underlying the multi-dimension entropy model's bitrate estimates.","marker":"[Ballé et al., 2018]"},{"why":"Supplies the 4DGaussian baseline and the data preprocessing protocol used for the evaluation datasets.","marker":"[Wu et al., 2024]"},{"why":"Provides the G-PCC codec used to compress anchor positions as part of the final bitstream.","marker":"[Liu et al., 2024a]"},{"why":"Supplies the HyperNeRF dataset and the per-frame deformation baseline context for dynamic scenes.","marker":"[Park et al., 2021b]"},{"why":"Supplies the Neu3D dataset used for evaluation.","marker":"[Li et al., 2022]"}],"fun_headline_variants":["Sparse anchors speed 4D Gaussian splatting 3–8x","Deform anchors, not Gaussians: 4D scenes in 5 MB","ADC-GS: 10x smaller 4D scenes, 3-8x faster rendering","Anchor-driven 4D splatting: ~5 MB, 135 FPS","Sparse anchor deformation cuts storage 10x, ups speed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The speed advantage rests on comparing ADC-GS's frame rates against baseline frame rates that may have been measured on different hardware or with different evaluation settings, since the paper does not report the baselines' GPUs.","fun_headline_variants_meta":{"raw":{"variants":["Sparse anchors speed 4D Gaussian splatting 3–8x","Deform anchors, not Gaussians: 4D scenes in 5 MB","ADC-GS: 10x smaller 4D scenes, 3-8x faster rendering","Anchor-driven 4D splatting: ~5 MB, 135 FPS","Sparse anchor deformation cuts storage 10x, ups speed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000664,"raw_usage":{"total_tokens":3010,"prompt_tokens":898,"completion_tokens":2112,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":2005}},"tokens_in":514,"tokens_out":2112,"duration_ms":14177,"temperature":1.0,"reasoning_tokens":2005,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:00:44.260349+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the per-Gaussian deformation baseline and ADC-GS on the same GPU, same evaluation frames, and same timing protocol, then check whether ADC-GS is still several times faster at matched model size and quality; if the gap disappears, the central speed claim is false.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian splatting renderer and the L1 and SSIM losses used for all rendering and training."},{"cited_title":"Per-gaussian embedding-based deformation for deformable 3d gaussian splatting","cited_arxiv_id":null,"evidence_quote":"Defines the per-Gaussian embedding deformation baseline whose speed and storage ADC-GS is compared against, and provides the local-similarity observation motivating anchors."}],"review_version":1}