{"id":"304018b3-492d-4646-a6f2-cbf6513ef236","arxiv_id":"2607.08250","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Two MoE integration strategies (joint canonical MoDE vs. independent-then-route MoE-GS) improve dynamic Gaussian Splatting by composing complementary deformation priors.","lead":"The paper designs two Mixture-of-Experts schemes that combine specialized motion models inside dynamic 3D Gaussian Splatting. One jointly optimizes experts on a shared canonical representation; the other trains them independently then routes their renders, clarifying trade-offs for more robust dynamic novel-view synthesis.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The strongest claim is carefully scoped as a comparative design study under two integration constraints, supported by quantitative tables, qualitative routing visualizations, efficiency ablations, and public code. The reader's identified weakest assumption is real but already acknowledged and partially quantified by the authors (static/dynamic trade-off in MoDE, expert-candidate ablation, stability across seeds). No load-bearing technical flaw (e.g., circular evaluation, missing baseline, or unstated assumption that collapses the argument) emerges on a second pass. Therefore the ACCEPT verdict with high confidence stands; the concrete test above is a useful but non-decisive verification rather than a required fix.","tokens_in":36196,"tokens_out":428,"duration_ms":4103,"concrete_test":"Re-run the N=3 MoE-GS configuration of Table III after replacing one expert with a deliberately redundant copy of the strongest expert (e.g., two copies of E-D3DGS + STG); if average PSNR still exceeds the single best expert by >0.3 dB the complementarity claim is robust, otherwise the gains partly reflect ensemble averaging rather than inductive-bias diversity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is a design-space analysis of two integration constraints for multi-deformation modeling (joint canonical composition in MoDE vs. independent experts + volume-aware image-space routing in MoE-GS), not a universal superiority claim. The reader's weakest assumption (non-interfering complementary biases) is already partially stress-tested by the paper itself: MoDE shows static-region degradation when the baseline is already strong (Fig. 6, E-D3DGS variants), Table X shows gains across different third experts, and MoE-GS consistently exceeds the best single expert under repeated runs (Table IX) and large-motion benchmarks (Appendix Tables XVI–XVII). No internal inconsistency or unsupported leap is present; the trade-offs are explicitly tabulated (Table I, XI) and the code is released.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper studies multi-deformation modeling for dynamic 3D Gaussian Splatting under two integration constraints, framed as Mixture-of-Experts. MoDE jointly optimizes multiple canonical deformation experts (HexPlane, hash-grid, per-Gaussian embeddings) on a shared canonical Gaussian representation with spline-based temporal Top-K gating and asymmetric gradient flow. MoE-GS independently trains heterogeneous experts (including non-canonical polynomial and keyframe-interpolation models plus static experts) then combines them via a volume-aware pixel router that lifts per-Gaussian routing weights; efficiency is recovered by single-pass multi-expert rendering, gate-aware pruning, and distillation. Extensive experiments on N3V, Technicolor, HyperNeRF, PanopticSports and D-NeRF, plus ablations of routers, pruning, distillation, expert candidates, run-to-run stability and multi-view depth consistency, demonstrate that both strategies improve robustness over single-deformation baselines while exposing complementary trade-offs in 3D fidelity, training stability and cost (Tables I–XI, Figs. 5–11).","tokens_in":36450,"tokens_out":768,"duration_ms":18592,"significance":"If the reported gains and trade-offs hold, the work supplies a clear design-space map for multi-deformation dynamic Gaussian representations rather than a single new SOTA method. The explicit contrast between joint canonical composition (MoDE) and decoupled expert-plus-router composition (MoE-GS), the geometry-aware lifting analysis, the efficiency mechanisms, and the released code (https://github.com/cvsp-lab/MoE-GS-studio) make the contribution reusable for future dynamic GS research. The empirical breadth—multiple datasets, large-motion benchmarks, repeated-run stability, and static-region degradation analysis—raises the bar for claims about deformation priors in this literature.","major_comments":[],"minor_comments":[{"comment":"Table II and Fig. 6: the static-region degradation for E-D3DGS-based MoDE is shown for one illustrative case; a short quantitative summary of static vs. dynamic ROI PSNR across all six N3V scenes would make the trade-off fully transparent.","section":"Sec. IV-B3 / Fig. 6"},{"comment":"Eqs. (12)–(15) and Fig. 4: the residual MLP Φ that refines the splatted routing features is described only at a high level; a one-sentence statement of its layer count / channel width would aid exact re-implementation.","section":"Sec. III-C2"},{"comment":"Appendix D3: the Multi-view Depth Consistency formula is clear, yet the precise set of viewpoint pairs and the depth-map resolution used for the curves in Fig. 11 are not stated; adding them would strengthen reproducibility of the geometry claim.","section":"Appendix D3 / Fig. 11"},{"comment":"A few typographical inconsistencies remain (e.g., “V olume-aware”, occasional missing spaces after citations). A final proof-reading pass would polish the manuscript.","section":"Throughout"}],"recommendation":"accept","confidential_remarks":"The manuscript is a natural journal extension of the authors’ own ICLR 2026 MoE-GS conference paper; the addition of MoDE, the systematic design-space analysis, and the expanded experiments justify the TPAMI venue. No novelty or citation issues warrant concern. I see no reason to request further major experiments."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean engineering paper that does something useful: it treats multi-deformation modeling for dynamic Gaussian Splatting as a design problem under two different integration constraints, and actually builds both.\n\nWhat is new is the systematic contrast. MoDE jointly optimizes multiple canonical deformation experts (HexPlane, hash, per-Gaussian embeddings) on a shared Gaussian set with spline-based temporal gating and asymmetric gradient flow. MoE-GS trains heterogeneous experts independently (including non-canonical ones like STG and Ex4DGS) and combines them with a volume-aware pixel router, plus single-pass rendering, gate-aware pruning, and distillation. The paper is explicit that these are alternatives with different trade-offs (Table I, XI), not a single winner. That framing is clearer than most “we add MoE” papers.\n\nThe experiments back it. MoE-GS consistently beats the best single expert on N3V and Technicolor (Tables III–IV), holds under repeated runs (Table IX), and still helps on large-motion sets in the appendix. Router ablations, pruning, and distillation are thorough. Code is released. MoDE’s story is more nuanced and the paper owns it: dynamic-region gains can be offset by static-region loss when the baseline is already strong (Fig. 6, E-D3DGS variants). That honesty is a plus, not a bug.\n\nSoft spots are real but proportional. The expert pool is hand-picked; Table X shows other third experts still help, but it is not an exhaustive search. MoDE’s full-image PSNR is not uniformly better. Free parameters (N, Top-K, spline points, pruning τ, distillation λ) exist, as expected for this class of system. None of that undermines the central design claim. The stress-test note is right: there is no load-bearing inconsistency.\n\nThis is for people working on dynamic 3DGS / free-viewpoint video who need to decide how to compose motion priors. It will not reorganize broader vision, but it is a practical advance with reproducible pieces. I would send it to peer review; the comparative analysis and evidence are strong enough to deserve referee time. Worth reading and, for anyone in the subfield, worth citing.","headline":"Solid design-space paper: two concrete MoE integration strategies for dynamic 3DGS, with real gains and honest trade-offs, not a universal claim.","tokens_in":37037,"tokens_out":563,"would_cite":true,"duration_ms":6434,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Two different MoE integration strategies for multi-deformation Gaussian models outperform any single deformation prior on dynamic scenes.","keywords":["dynamic Gaussian splatting","mixture of experts","deformation modeling","novel view synthesis","3D reconstruction","MoDE","MoE-GS"],"falsifier":"Find a dynamic scene (or large ROI set) in which one fixed deformation model already wins every spatial region and every timestamp; on that data both MoDE and MoE-GS must then match or fall below that single expert’s PSNR rather than improve it.","tokens_in":37103,"feed_emoji":"🌀","tokens_out":814,"duration_ms":17512,"temperature":0.7,"pith_summary":"Real-world dynamic scenes contain heterogeneous motion that no single deformation model captures well: different methods win on different scenes, patches, and timestamps because each deformation prior favors only certain motion regimes. The paper treats multi-deformation modeling as a Mixture-of-Experts problem and shows two concrete ways to combine specialized experts inside a unified 3D Gaussian representation. MoDE jointly optimizes multiple canonical deformation experts that share the same Gaussian primitives and are gated by a continuous temporal spline, adding almost no extra stages. MoE-GS instead trains heterogeneous experts independently (including non-canonical ones) and later blends their rendered images with a volume-aware pixel router. Together the two designs map the trade-offs among reconstruction fidelity, training stability, and compute, giving practitioners clear alternatives rather than a single universal recipe.","feed_headline":"Two MoE designs beat single-deformation Gaussians","feed_subtitle":"Joint and decoupled expert mixtures handle heterogeneous motion better than any one prior.","key_machinery":"Two integration constraints that decide when experts interact: MoDE (shared canonical Gaussians + spline temporal gating + baseline-only gradient flow) versus MoE-GS (independent experts + volume-aware pixel router that splat per-Gaussian routing weights then blends images).","core_discovery":"Performance gaps among existing dynamic Gaussian Splatting methods arise from reliance on one deformation prior, not from lack of capacity. Combining multiple specialized deformation experts under either joint canonical optimization (MoDE) or decoupled optimization plus volume-aware routing (MoE-GS) systematically improves novel-view quality by exploiting complementary motion behaviors across space and time.","pith_inferences":["The same joint-versus-decoupled design choice likely applies to other explicit dynamic representations (meshes, particles, hash grids) whose motion models also carry conflicting inductive biases.","An online residual-driven expert pool that can add or drop deformation modules mid-training would reduce the need for hand-selected candidate sets.","Volume-aware routing may serve as a general post-hoc calibration layer for any ensemble of 3D renderers that lack direct primitive correspondence."],"forward_implications":["When a shared canonical space exists, MoDE gives multi-deformation modeling with only modest extra training time and direct 3D Gaussian output.","When experts are heterogeneous or already trained, MoE-GS yields larger PSNR gains by full specialization, at the price of multiple training runs and a routing stage.","Gate-aware pruning and distillation recover real-time speed while retaining most of the mixture’s quality.","Image-space routing can still be lifted to a coherent post-hoc 3D Gaussian model whose multi-view depth consistency matches or exceeds single experts."],"fun_headline_variants":["MoDE and MoE-GS mix deformation experts for dynamic Gaussians","Two MoE strategies beat single-deformation dynamic Gaussians","Joint vs decoupled experts fix motion gaps in Gaussian Splatting","Multi-deformation MoEs improve dynamic novel-view quality","Shared-canonical MoDE and routed MoE-GS handle heterogeneous motion"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"The chosen experts keep complementary, non-interfering motion priors under the proposed gating so the mixture reliably beats the strongest single expert instead of averaging or destabilizing.","fun_headline_variants_meta":{"raw":{"variants":["MoDE and MoE-GS mix deformation experts for dynamic Gaussians","Two MoE strategies beat single-deformation dynamic Gaussians","Joint vs decoupled experts fix motion gaps in Gaussian Splatting","Multi-deformation MoEs improve dynamic novel-view quality","Shared-canonical MoDE and routed MoE-GS handle heterogeneous motion"]},"model":"grok-4.5","effort":"low","cost_usd":0.001832,"raw_usage":{"total_tokens":920,"prompt_tokens":826,"num_sources_used":0,"completion_tokens":94,"cost_in_usd_ticks":18320000,"prompt_tokens_details":{"text_tokens":826,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":0,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":826,"tokens_out":94,"duration_ms":1579,"temperature":1.0,"reasoning_tokens":0,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T15:36:17.107229+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Find a dynamic scene (or large ROI set) in which one fixed deformation model already wins every spatial region and every timestamp; on that data both MoDE and MoE-GS must then match or fall below that single expert’s PSNR rather than improve it.","supporting_citations":[],"review_version":2}