{"id":"a25c6feb-90fe-4f98-9b50-da2a1ef80292","arxiv_id":"2506.07865","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FreeGave predicts future 3D frames by learning per-Gaussian physics codes and divergence-free velocities from multi-view video, outperforming baselines on extrapolation and unsupervised motion segmentation.","lead":"FreeGave learns 3D scene shape and motion from multiple cameras, then predicts what the scene looks like in future frames. It encodes each 3D Gaussian part with a 'physics code' and uses a divergence-free velocity field to move the parts forward in time, without needing any masks or object labels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'physics learning' claim is not established: future velocity is an MLP extrapolation with no temporal dynamical law, so the divergence-free constraint is only a spatial kinematic prior.","rationale":"I read the paper as an empirical method paper: it introduces a 3DGS-based dynamic scene representation with per-Gaussian physics codes and a divergence-free velocity parameterization, and demonstrates superior future-frame extrapolation and unsupervised motion segmentation across four datasets. The experiments are extensive, with per-scene tables, ablations, a new GoPro dataset, and collision studies; the improvement over strong baselines such as DefGSnvfi is credible. My concern is not with the empirical results but with the conceptual claim that the method 'learns physics.' The divergence-free constraint is a kinematic property of an instantaneous velocity field of the form v = v_lin + ω×p; it is exactly the velocity field of rigid-body motion, and its divergence is identically zero. It does not constrain how Vt evolves in time, and no dynamical law is imposed on f_weight(t). Future frames are generated by querying this MLP at unseen timestamps. Thus the extrapolation quality measures the smoothness and generalizability of a learned temporal function, not the recovery of an underlying physical equation. The collision results in Appendix F (small margins over DefGSnvfi) and the limitation in Appendix H (failure on abrupt motions) corroborate this reading. I therefore agree with the reader's main reservation about f_weight(t) generalization, though I do not think the 'global divergence-free' phrasing is the crux, since the method only claims per-Gaussian divergence-free. I also note the segmentation evaluation uses per-scene choices of cluster count C and position-weighting λ (Appendix I), which introduces priors into the 'unsupervised' demonstration, but that is secondary. Because the empirical contribution is solid and the overclaim is fixable by revising the language and adding a temporal-physics test, the CONDITIONAL verdict is appropriate and I recommend no change.","tokens_in":27788,"tokens_out":14531,"duration_ms":177754,"concrete_test":"Evaluate the ParticleNeRF Spring scene (or a new controlled harmonic-oscillator scene) by training on exactly one partial period and extrapolating over one full additional period. Compare FreeGave's predicted Gaussian centroid trajectories to the analytic SHM solution x(t)=A cos(ωt+φ), fitting A, ω, and φ to the training portion only. If the RMS position error over the extrapolated period grows monotonically (drift) and exceeds 10% of the motion amplitude, the extrapolation behaves as a learned function fit rather than the recovered physical law. A simpler variant: use the existing Spring data (training and test together form one full period) and report per-frame error against the analytic SHM curve for both FreeGave and an SHM baseline fitted to the observed positions; if FreeGave does not statistically beat that baseline, the 'physics learning' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that FreeGave 'learns the underlying physics' of complex 3D scenes. The only physical ingredient is a spatial kinematic constraint: the per-Gaussian velocity v(pt,t)=Vt·B(pt) has zero divergence with respect to pt because Vt is chosen independent of pt (Appendix A). This is a property of rigid-body motion, not a dynamical law. The temporal evolution is produced by f_weight(t) in Eq (6), a neural network queried at future timestamps, with no loss or constraint enforcing any governing equation (Newton's second law, momentum conservation, or a PDE) on the time dimension. Algorithm 1 therefore extrapolates a learned function rather than a physical law. Since each Gaussian has its own Vt, the collection of per-particle velocity fields does not form a single divergence-free global field, so the 'divergence-free' terminology does not make the scene motion physically consistent. The empirical results are strong on smooth, short-horizon motions, but the collision experiments in Appendix F show only a small gain over DefGSnvfi (28.426 vs 28.017 PSNR), and the stated limitation (Appendix H) concedes failure on abrupt motions. Consequently, the abstract's claim that the method learns 'the underlying physics' and 'meaningful 3D physical motion patterns' outruns the evidence; the supported claim is accurate future-frame extrapolation via a smooth per-Gaussian velocity parameterization.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FreeGave, a 3D Gaussian splatting method that learns per-Gaussian velocity fields from multi-view videos without object priors or PINN losses. Each Gaussian kernel is assigned a physics code z=f_code(p0), and its velocity is parameterized as v(p_t,t)=V_t·B(p_t) with V_t=f_neck(z)·f_weight(t), where B(p_t) is a divergence-free basis for rigid-body motion. An auxiliary deformation field aids optimization, and future frames are generated by an interleaved midpoint integration (Algorithm 1). The method is evaluated on three public datasets and a newly collected real-world dataset for future frame extrapolation and unsupervised motion segmentation, with reported improvements over several baselines including a DefGS+NVFi hybrid.","tokens_in":28047,"tokens_out":4980,"duration_ms":58242,"significance":"The empirical contribution is substantial: the evaluation covers four datasets, multiple baselines, per-scene results, ablations, and held-out future frames, and the new FreeGave-GoPro dataset is a potentially useful resource for the community. The future-frame extrapolation results, in particular the large margins on several synthetic and real scenes, are strong evidence that the proposed per-Gaussian velocity parameterization generalizes beyond the training time range. The unsupervised motion segmentation via K-means on bottleneck vectors is also a nice by-product. However, the paper's central claim of learning 'the underlying physics' is not established by the evidence. The divergence-free property is per-Gaussian, not global, and the temporal evolution is an MLP extrapolation rather than a dynamical law. The method is better characterized as a kinematic velocity parameterization with good extrapolation performance, and the 'physics learning' framing should be moderated accordingly.","major_comments":[{"comment":"The divergence-free property is proven only for each basis function B_k(p_t) with V_t fixed. In the actual model, V_t = f_neck(z)·f_weight(t) with z = f_code(p0), so each Gaussian has a different V_t. The collection of per-Gaussian velocity fields does not define a single global vector field on the scene, and the phrase 'divergence-free Gaussian velocity' in the abstract and Section 3.2 describes a per-particle rigid-body kinematic constraint, not a global incompressibility constraint. The paper should define what 'divergence-free' means at the scene level and verify it, or explicitly reframe the contribution as per-particle kinematic parameterization.","section":"§3.2, Eqs. (4)–(6), and Appendix A"},{"comment":"The time dependence of the velocity is carried entirely by f_weight(t), a neural network, with no loss or constraint enforcing a dynamical law (e.g., Newton's second law, momentum conservation, or a PDE) on the time dimension. Future extrapolation is therefore MLP evaluation at unseen timestamps, not integration of learned physics. The strong results on smooth/periodic motions and the modest results on collisions (Appendix F, Table 6: 28.426 vs 28.017 PSNR over DefGSnvfi) are consistent with function extrapolation of a smooth motion field. The abstract's claims of learning 'the underlying physics' and 'meaningful 3D physical motion patterns' should be revised, or the authors should add a temporal dynamics constraint and demonstrate that it improves collision and abrupt-motion cases.","section":"§3.2, Eq. (6), and Algorithm 1"},{"comment":"The limitation section states that the method 'would fail to predict abrupt motions, such as an explosion, primarily because the underlying physics rules are unable to be observed or learned from visual frames.' This is an honest statement, but it directly undercuts the abstract's claim that the method learns 'the underlying physics' of complex dynamic 3D scenes. A method that fails precisely when a dynamical law governs the motion is better described as learning smooth velocity extrapolation. The authors should either provide evidence that the learned physics codes capture invariant physical parameters beyond the observed motion statistics, or revise the central claims to match the actual scope.","section":"Appendix H"}],"minor_comments":[{"comment":"There are numerous typos and formatting inconsistencies in the main text and supplementary material, e.g., 'Canonial3DRepresentation' in Figure 2, 'fdef orm' vs 'f_def orm' vs 'fdeform', and 'DefGSnvf i' with stray spaces. A thorough proofread is needed.","section":"General"},{"comment":"The rotation update R_t ← (I + ∆t ∂v_mid/∂p_mid) R_t' is borrowed from [70] without explaining why this approximation is valid for the per-Gaussian velocity field or whether it preserves the divergence-free property. Please add a short derivation or reference to the specific result.","section":"§3.3, Algorithm 1"},{"comment":"The motion segmentation evaluation uses K-means on bottleneck vectors with per-scene hyperparameters (λ and C) chosen individually for each scene (Appendix I.1). The paper should state how these values were selected and whether the reported high AP/PQ scores are robust to reasonable variations of λ and C.","section":"§4.2 and Appendix I"},{"comment":"The caption says 'Results on four scenes of oscillations or collisions' but the table only reports a single 'Collisions' row. Please clarify whether the results are averaged over two collision scenes and whether the 'oscillations' scenes are reported elsewhere.","section":"Table 6"},{"comment":"The paper cites its own prior work (NVFi, OGC) very frequently; please ensure that the novelty with respect to NVFi is crisply stated, since the velocity parameterization is closely related to the NVFi framework.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The empirical work is solid and the dataset release is valuable, but the 'physics learning' framing is substantially overstated relative to the actual method, which is a per-Gaussian kinematic velocity extrapolation. The divergence-free terminology is misleading at the scene level, and the ablation and collision results do not support the claim that a physical law has been learned. I would encourage the editor to require the authors to either add a genuine temporal dynamics constraint or rewrite the abstract, introduction, and conclusion to describe the contribution as accurate future-frame extrapolation via a smooth per-Gaussian velocity parameterization. The unusually high self-citation rate (NVFi, OGC, and related works from the same group) is also worth noting, though it does not affect the technical assessment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: FreeGave is the best current recipe for future-frame extrapolation in dynamic 3DGS without masks, but it does not learn 'underlying physics' in any dynamical sense. The divergence-free module is a kinematic rigid-motion parameterization, and future velocity is MLP extrapolation. Treat it as an empirical system, not a physics learner.\n\nWhat's new: a per-Gaussian latent physics code, a hard-coded divergence-free velocity basis from rigid-body motion, and a deformation-aided optimization that avoids PINN losses. The combination is new and it works. Evaluation is thorough: four datasets (including a new GoPro scene set), multiple baselines, per-scene numbers, and ablations. The extrapolation gains over DefGSnvfi are large on most scenes. The motion segmentation via K-means on bottleneck vectors is clever and strong, though the per-scene choices of cluster count C and smoothing lambda weaken the 'no priors' claim.\n\nSoft spots: the core claim of 'learning physics' outruns the evidence. The divergence-free property is spatial and kinematic — any rigid-body velocity field has zero divergence, and the Appendix A proof is correct for each basis function. That proves nothing about time dynamics. Eq. (6)'s f_weight(t) is a neural network that is simply evaluated at future timestamps; there is no PDE, no momentum law, no Newton constraint. And because each Gaussian owns a separate V_t, the global velocity is not one divergence-free field. The collision results are only mildly better than DefGSnvfi, and the appendix admits failure on abrupt motions. So the honest contribution is: smooth per-Gaussian velocity parameterization that extrapolates well on smooth, short-horizon motion.\n\nRecommendation: send to peer review. The empirical core is solid, the method is clearly described and ablated, and the conceptual overclaim can be fixed in revision (reword claims, or add a temporal physical constraint if they want to earn the title). I'd cite it for the method, and I'd bring it to reading group. It deserves a serious referee.","headline":"FreeGave is a solid empirical system for future-frame extrapolation with 3DGS whose 'physics learning' framing overreaches: the divergence-free constraint is kinematic, not dynamical.","tokens_in":28609,"tokens_out":3893,"would_cite":true,"duration_ms":46601,"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":"FreeGave learns per-Gaussian physics codes and divergence-free velocity fields from multi-view video, enabling future frame extrapolation and unsupervised motion segmentation without object priors.","keywords":["3D Gaussian splatting","physics learning","divergence-free velocity field","future frame extrapolation","motion segmentation","dynamic scenes","physics code","multi-view video"],"falsifier":"Feed FreeGave a video in which an object's motion changes abruptly after the training frames, such as a ball hitting a wall and reversing; the failure would be extrapolated frames showing the ball passing through the wall, which would show that the per-Gaussian divergence-free velocity fields do not capture interaction physics.","tokens_in":27522,"feed_emoji":"🎥","tokens_out":7136,"duration_ms":65262,"temperature":0.7,"pith_summary":"FreeGave aims to show that the physics of a dynamic 3D scene can be learned from multi-view RGB video alone, without object masks, object-type priors, or physics-loss terms. It treats each 3D Gaussian kernel of a splatting representation as a rigid particle, gives each particle a learned latent physics code, and derives its motion from a divergence-free velocity field built from that code. The paper reports that this approach extrapolates future frames more accurately than prior velocity-field and deformation-based methods across four datasets, and that the learned codes cluster into meaningful object motion patterns in an unsupervised way. If correct, it suggests that per-particle divergence-free velocity learning is a viable replacement for PDE-based physics losses in dynamic scene modeling.","feed_headline":"Per-Gaussian physics codes predict future 3D motion from video","feed_subtitle":"A learnable physics code per Gaussian drives future frames, no masks or object types needed.","key_machinery":"The core object is the per-Gaussian velocity field $v(p_t, t) = V_t \\cdot B(p_t)$, built from a physics code $z = f_{\\text{code}}(p_0)$ shared across time. $V_t$ is produced by $f_{\\text{neck}}(z) \\cdot f_{\\text{weight}}(t)$, a bottleneck that factors motion into $K$ latent pattern types and time-selective weights, while $B(p_t)$ is a $6 \\times 3$ basis whose rows are the linear-velocity directions and the cross-product angular-velocity terms, each divergence-free. Because $V_t$ is independent of $p_t$ and each basis row has zero divergence, the field is divergence-free without any PINN loss. The same velocity field is paired with a deformation-aided optimization that transports canonical Gaussians to nearby timestamps for supervision, and an interleaved mid-point integration advances the particles forward for future prediction.","core_discovery":"The central claim is that embedding a physics code in every Gaussian kernel and transporting each kernel by a divergence-free velocity field teaches the model the scene's underlying dynamics directly from pixels. Velocities are parameterized as $v(p_t, t) = V_t \\cdot B(p_t)$, where $V_t$ is learned from the code via a bottleneck that decomposes motion into $K$ pattern types and time-dependent weights, and $B(p_t)$ is a fixed basis of linear and angular velocity fields whose divergence is zero by construction. This lets future frames be obtained by integrating the velocity field forward, with a deformation field used during training to stabilize optimization. The paper argues that because the divergence-free constraint is baked into the parameterization rather than imposed as a soft PINN loss, boundary motions are learned more sharply and no object priors are needed. Evidence includes higher extrapolation metrics on synthetic and real datasets and near-perfect unsupervised motion segmentation on an indoor benchmark.","pith_inferences":["A natural next step the paper does not take is adding interaction terms between nearby Gaussian codes; the current per-particle independence would likely break at collisions, though the paper's collision experiments suggest some robustness.","The same bottleneck decomposition could transfer across scenes: a physics code learned on one object might be reused to drive the motion of a geometrically different object in a 4D generation pipeline.","One testable extension is to measure how extrapolation error grows with horizon; the paper evaluates short extrapolation windows, and a longer-horizon study would reveal whether the learned velocity field remains stable."],"forward_implications":["If the claim holds, future-frame extrapolation for dynamic scenes no longer requires object masks, object-type classifiers, or expensive PDE sampling; per-Gaussian divergence-free velocities learned from RGB are sufficient.","The physics codes themselves become a representation that can be clustered or edited, so scenes can be re-animated by modifying a code or transferring a code between objects.","Because the model already separates motion patterns unsupervised, robotics and embodied agents could extract moving-object proposals directly from the learned codes without hand-labeled segmentation.","The architecture's bottleneck dimension $K$ controls the number of motion patterns, giving a principled lever for trading expressiveness against generalization in long-horizon prediction."],"supporting_citations":[{"why":"Supplies the velocity-field baseline and the two synthetic datasets used for the main extrapolation comparisons.","marker":"[30]"},{"why":"Provides the 3D Gaussian scene representation that each kernel is treated as a rigid particle in.","marker":"[28]"},{"why":"Provides the deformable-Gaussian baseline whose deformation field the paper adapts for deformation-aided optimization.","marker":"[75]"},{"why":"Supplies the rotation-update approximation used in the interleaved mid-point integration to transport Gaussian orientations.","marker":"[70]"},{"why":"Provides the ParticleNeRF dataset used for an additional future-frame extrapolation benchmark.","marker":"[1]"},{"why":"Supplies the feature-rendering approach used to convert grouped physics codes into segmentation masks for motion segmentation evaluation.","marker":"[78]"},{"why":"Provides the unsupervised segmentation losses used to segment the 3DGS-based baselines for fair comparison.","marker":"[57]"}],"fun_headline_variants":["Per-Gaussian velocity learns 3D physics from video","Divergence-free velocity predicts future frames without masks","Physics codes in Gaussian splatting predict 3D motion","No priors: learn scene physics directly from video"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The premise is that independent per-Gaussian divergence-free velocity fields, with time dependence learned by a weight network, remain physically faithful beyond the training interval even though the assembled global field is not a single divergence-free flow.","fun_headline_variants_meta":{"raw":{"variants":["Per-Gaussian velocity learns 3D physics from video","Divergence-free velocity predicts future frames without masks","Physics codes in Gaussian splatting predict 3D motion","No priors: learn scene physics directly from video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000508,"raw_usage":{"total_tokens":2452,"prompt_tokens":897,"completion_tokens":1555,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":1489}},"tokens_in":513,"tokens_out":1555,"duration_ms":13178,"temperature":1.0,"reasoning_tokens":1489,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:23:53.945905+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed FreeGave a video in which an object's motion changes abruptly after the training frames, such as a ball hitting a wall and reversing; the failure would be extrapolated frames showing the ball passing through the wall, which would show that the per-Gaussian divergence-free velocity fields do not capture interaction physics.","supporting_citations":[{"cited_title":"NVFi: Neural Veloc- ity Fields for 3D Physics Learning from Dynamic Videos","cited_arxiv_id":null,"evidence_quote":"Supplies the velocity-field baseline and the two synthetic datasets used for the main extrapolation comparisons."},{"cited_title":"Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction","cited_arxiv_id":null,"evidence_quote":"Provides the deformable-Gaussian baseline whose deformation field the paper adapts for deformation-aided optimization."},{"cited_title":"Physgaussian: Physics- integrated 3d gaussians for generative dynamics","cited_arxiv_id":null,"evidence_quote":"Supplies the rotation-update approximation used in the interleaved mid-point integration to transport Gaussian orientations."},{"cited_title":"Gaussian grouping: Segment and edit anything in 3d scenes","cited_arxiv_id":null,"evidence_quote":"Supplies the feature-rendering approach used to convert grouped physics codes into segmentation masks for motion segmentation evaluation."},{"cited_title":"OGC: Unsupervised 3D Ob- ject Segmentation from Rigid Dynamics of Point Clouds","cited_arxiv_id":null,"evidence_quote":"Provides the unsupervised segmentation losses used to segment the 3DGS-based baselines for fair comparison."}],"review_version":1}