{"id":"dbb8b348-f351-4151-a533-44265b6b3528","arxiv_id":"2411.17189","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PhysMotion generates physically plausible videos from a single image by simulating 3D object motion with a material point method, then enhancing the rendering with a diffusion model.","lead":"PhysMotion turns a single photo into a short video where the object moves as if pushed, dropped, or stretched, using a physics simulator to drive the motion and a diffusion model to make it look realistic. This matters because it offers a way to create controllable, physically plausible animations from one still image, which could be useful for filmmakers, game designers, and content creators.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never measures whether the diffusion enhancement preserves the MPM motion; VideoPhy PC and user preference only score plausibility, so the 'physics-grounded' claim is not yet supported.","rationale":"The pipeline is well-motivated and the pieces are individually sound: MPM is a principled simulator, the single-image 3D Gaussian reconstruction with depth supervision is a reasonable lifting of the problem, and the diffusion refinement is a sensible way to add texture, background, and temporal consistency. If the enhancement preserves the coarse dynamics, the central claim is credible. But the paper's own evidence does not establish that preservation. Section 3.4's feature injection and cross-frame attention are designed for visual and temporal consistency following TokenFlow, not for physical fidelity, and no sensitivity study is reported for the injection hyperparameters tau_f and tau_A. The only quantitative physics measure is VideoPhy's PC, which is a learned judge of physical commonsense in natural videos; it can reward plausible-looking motion even if that motion is not the simulated one and cannot verify force/torque conditioning. The user study measures preference, not accuracy. Thus the weakest link is the connection between the coarse simulation and the final enhanced video. This is the same concern the reader identified, so I agree with the conditional verdict rather than moving it. A trajectory-fidelity check with a control-motion ablation would settle the matter: low deviation and degraded control scores would support the physics-grounded claim, while large deviation or comparable control scores would require the claim to be weakened to 'physics-inspired' or 'physics-initialized' dynamics.","tokens_in":18775,"tokens_out":3251,"duration_ms":33324,"concrete_test":"Compute a trajectory-fidelity metric between the coarse MPM renderings {I_j} and the final enhanced frames {I*_j}: track the foreground object centroid and key contact points in both videos and report per-frame displacement error plus temporal event timing (e.g., first impact frame, bounce period) for all 13 test scenes. Then repeat the enhancement with a control motion that violates the MPM dynamics, such as time-reversed frames or a rigid translation that ignores the simulated force/torque. If trajectory error is large or if the control achieves comparable VideoPhy PC and user-preference scores, the physics-grounding claim is not supported; if error is small and the control degrades scores, the claim is substantiated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim asserts physics-grounded dynamics, yet the paper provides no measurement that the final enhanced video actually follows the MPM simulation. Section 3.4 injects diffusion features and cross-frame attention into the enhanced sampling, with feature/attention injection timesteps tau_f and tau_A as free hyperparameters that control how much of the coarse simulation survives. The quantitative evaluation in Section 4.2 uses VideoPhy's Physical Commonsense and Semantic Adherence z-scores plus a user preference study; these measure plausibility and prompt alignment, not fidelity to simulated trajectories or to the user-specified force/torque. Consequently, a video could score well on PC while the enhancement stage substantially alters the motion: falling direction, bounce timing, deformation magnitude, or contact behavior could all change without being detected. The paper's own Section 5 admits the diffusion stage may introduce artifacts and color distortions, and no ablation quantifies motion deviation with versus without enhancement. The load-bearing assumption is therefore not the MPM simulator itself, which is standard and off-the-shelf, but the unverified claim that the enhancement acts as a detail refiner rather than a motion rewrite. If the enhancement changes the dynamics, the advertised physics grounding is illusory even though the videos look plausible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PhysMotion, a framework that turns a single input image into a video of physics-driven object dynamics. The method first segments the foreground, reconstructs a 3D Gaussian representation using LGM with depth and color refinement, and then steps that representation forward with a Material Point Method (MPM) simulator using continuum-mechanics elastoplasticity models to produce a coarse video of the object. A diffusion-based enhancement stage, built on DDIM+ inversion, ControlNet depth/edge conditioning, and cross-frame attention, blends the coarse dynamics with the background and restores fine texture and visual quality. The authors evaluate against several image-to-video diffusion baselines using VideoPhy's Physical Commonsense (PC) and Semantic Adherence (SA) z-scores, a user study, and qualitative comparisons, reporting advantages on all quantitative criteria and demonstrating a range of materials and fracture effects. The central claim is that PhysMotion is the first single-image-to-video framework that combines 3D geometry awareness with physics-grounded dynamics, where the physics comes from the MPM simulation and the diffusion stage acts as a detail refiner.","tokens_in":19046,"tokens_out":4767,"duration_ms":43521,"significance":"If the central claim is fully supported, the paper would make a useful step toward controllable, physically plausible video synthesis from a single image, an area where purely data-driven video diffusion models often violate physical laws. The use of standard continuum mechanics and an off-the-shelf MPM solver gives the physics component a solid foundation, and the qualitative results, especially for fracture and elastoplastic deformation, are visually compelling and demonstrate versatility across materials. The method builds on a sensible combination of existing components (PhysGaussian-style physics-integrated 3DGS, LGM reconstruction, TokenFlow-style cross-frame attention, and ControlNet), and the paper is generally clearly written. However, the current manuscript does not release code or the 13-image test set, the quantitative evaluation is thin, and no measurement is provided that the diffusion enhancement preserves the simulated motion. Those gaps directly affect the strength of the 'physics-grounded' claim, and they can be addressed within the manuscript's scope.","major_comments":[{"comment":"The paper's central claim (Abstract and Section 1) is that the output video is physics-grounded because it is guided by MPM simulation. However, no experiment measures whether the diffusion enhancement stage preserves the simulated motion. The feature/attention injection timesteps tau_f and tau_A (Section 3.4.2) are free hyperparameters that control how much of the coarse simulation survives, and the quantitative metrics (VideoPhy PC/SA and the user study) measure plausibility and semantic adherence, not fidelity to the MPM trajectory. Please add a quantitative comparison between coarse and enhanced videos, for example by tracking keypoints or by comparing rendered foreground masks or optical flow of the enhanced video against the MPM output, and include an ablation that varies tau_f and tau_A and reports the resulting motion deviation. Without this, the claim that the enhancement is a detail refiner rather than a motion rewriter is unsupported.","section":"Section 3.4 and Section 4.2"},{"comment":"The user study results are not reported for the proposed method: the 'Ours' row of Table 1 contains only em-dashes for TC, PP, and Overall, yet the text states that 'our proposed method consistently outperforms all baselines across every evaluation criterion in the user preference results.' Please report the actual preference percentages and their statistical significance (for example, bootstrap confidence intervals or a paired test), and clarify whether the baseline percentages in Table 1 denote the preference rate for the proposed method over each baseline or the preference rate for each baseline over the proposed method. The current presentation does not support the claimed superiority.","section":"Table 1 and Section 4.2"},{"comment":"The quantitative evaluation uses only 13 scenes, with per-scene z-score normalization across the six methods, and no error bars or significance tests are provided for the reported average z-scores. With 13 scenes, the reported differences (for example, PC z-score 0.5142 vs. 0.1853) may be within the noise, especially because z-score normalization forces each scene's mean to zero and discards the absolute score scale. Please report the per-scene scores (or include them as a table in the appendix), provide error bars or confidence intervals, and perform a significance test. In addition, the paper should state whether the 13-image test set and the evaluation code will be released, since independent verification is not currently possible.","section":"Section 4.2 and Appendix A.3"},{"comment":"The ablation studies are qualitative only. In particular, the key claim that the enhancement stage is necessary and that it does not degrade physical motion is supported by a single visual example (the bread-tearing case in Fig. 6), while Section 5 concedes that the diffusion stage may introduce artifacts and color distortions. Please provide quantitative ablations for the enhancement stage, such as VideoPhy scores with and without enhancement as well as the motion-deviation metric described above, and for the hard-depth loss, ideally including reconstruction metrics on the input view in addition to the qualitative Fig. 8.","section":"Section 4.4 and Section 5"}],"minor_comments":[{"comment":"The caption cites 'I2VGen-XL [72]' but I2VGen-XL is reference [105] in the bibliography; the reference numbering is inconsistent between the text (Section 4.2 uses [105]) and the figure caption.","section":"Figure 5"},{"comment":"The sentence 'The timesteps for feature and attention injection is controlled by two hyperparameters' has a subject-verb agreement error and should read 'are controlled.' Also, Eq. (12) introduces notation Q_{jk,c} and K_{j1,c} without explicitly connecting it to the previously defined Q_c and K_c, which makes the equational flow hard to follow.","section":"Section 3.4.2"},{"comment":"The phrase 'we randomly choose key-frames every 5 frames' is ambiguous: if key-frames are selected every 5 frames, the selection is deterministic, not random, and the word 'randomly' should be removed or the random mechanism specified.","section":"Appendix A.2.2"},{"comment":"The implementation appendix gives training epochs and learning rates but does not specify the material model parameters (for example, Young's modulus, yield stress, or hardening parameters) used for each showcased material; please either list these parameters or state explicitly that they are inherited from the MPM solver or from PhysGaussian [91].","section":"Section 4.1 and Appendix A.2"},{"comment":"The phrase 'simulation-based methods produce generative dynamics that are physically grounded [4]' cites VideoPhy [4] as support, but VideoPhy is an evaluation benchmark for physical commonsense, not a method that produces physically grounded dynamics; consider rephrasing to avoid conflating evaluation with generation.","section":"Section 2.1"}],"recommendation":"major_revision","confidential_remarks":"The evaluation benchmark VideoPhy [4] and the MPM solver [113] share authors with this manuscript, which introduces a conflict-of-interest risk for the quantitative claims; I recommend asking the authors to release the test scenes, raw per-model scores, and evaluation scripts to allow independent verification. The 'first single image-to-video framework with 3D geometry awareness and physics-grounded dynamics' claim is context-dependent given recent concurrent work (PhysGen, Phy124), but it is not by itself a reason for rejection; the main load-bearing issue is the unverified motion preservation through the enhancement stage. The paper is a reasonable candidate after substantive revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, PhysMotion is a reasonable, well-engineered pipeline: lift a single image to 3D Gaussians with LGM, optimize geometry, run MPM for the dynamics, then refine the coarse renderings with a TokenFlow-style diffusion pass. The combination is new, and the results look good. Second, the paper never checks whether the refinement stage preserves the simulated motion. The evaluation measures plausibility and prompt adherence, not fidelity to the MPM output, so the \"physics-grounded\" claim is, strictly speaking, unsupported.\n\nWhat's genuinely good: the design is honest about the coarse simulation being coarse and uses the diffusion step to add texture and background, which is a pragmatic answer to a real problem. The comparison against five I2V baselines is fair in setup. The ablation of image-space vs latent-space blending is useful, and the material coverage (elastic, plastic, rigid, granular, fracture) is impressive for one pipeline.\n\nThe soft spots are real but not fatal. The biggest is the missing motion-fidelity check. ControlNet depth/edge conditions constrain each frame's geometry, but per-frame constraints don't pin down trajectories, bounces, or deformation rates; the injection timesteps tau_f and tau_A are free parameters that decide how much of the physics survives. A simple tracking or trajectory-error measurement against the coarse simulation would settle it. Second, the \"first\" claim is overstated — Phy124 [50] already does single-image physics-driven generation, and the paper's own distinction is quality, not existence. Third, the quantitative section is thin: 13 scenes, z-scores without error bars or significance tests, and a user-study table that leaves the Ours row blank. Reproducibility is limited by no code, no test set, and no per-scene material/force settings. The self-citation cluster (PhysGaussian, the MPM solver, VideoPhy) is heavy but not improper; the mechanics equations are standard.\n\nWho should read it: people building physics-aware I2V systems or looking for a practical recipe. It deserves a serious referee: the method is novel enough and the problems are fixable. My recommendation is to send it out, with a required revision that quantifies how much the enhancement deviates from the simulation and completes the reporting.","headline":"Plausible single-image physics I2V pipeline, but the paper never checks whether the diffusion stage preserves the simulated motion, so the headline claim outruns the evidence.","tokens_in":19584,"tokens_out":3207,"would_cite":false,"duration_ms":28790,"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":"PhysMotion turns a single image into a physically plausible video by simulating the object's 3D shape with continuum mechanics, then polishing the coarse simulation with a diffusion model.","keywords":["single-image video generation","physics-grounded dynamics","material point method","3D Gaussian splatting","diffusion-based video enhancement","image-to-video generation","controllable dynamics","continuum mechanics"],"falsifier":"Take one generated scene, such as a falling and bouncing object, and track a few surface points in the final enhanced video; compare those trajectories to the trajectories of the corresponding MPM particles. If the enhanced video's motion diverges from the simulation beyond a small tolerance, or if the video fails to respond appropriately to a changed applied force, the paper's physics-grounded claim is not borne out.","tokens_in":18588,"feed_emoji":"🎬","tokens_out":3887,"duration_ms":34561,"temperature":0.7,"pith_summary":"The paper tries to establish that a single photograph is enough to generate a short video whose motion obeys physics: the object is lifted to a 3D Gaussian representation, its dynamics are computed by a continuum-mechanics Material Point Method simulation under user-specified forces and torques, and a diffusion-based enhancement stage then restores fine texture and detail. If this works as claimed, it closes a gap between purely data-driven image-to-video models, which often violate physical laws, and physics simulators, which usually require 3D models or multi-view inputs. A sympathetic reader would take the central contribution to be the combination of a physics-grounded coarse simulation with a generative refinement stage that preserves both physical plausibility and visual fidelity.","feed_headline":"A single image becomes a physics-grounded video","feed_subtitle":"The method reconstructs 3D geometry, simulates it with continuum mechanics, then sharpens the result with diffusion.","key_machinery":"The load-bearing object is the physics-integrated 3D Gaussian representation: each Gaussian carries a deformation gradient F, so the MPM simulation updates the Gaussian centers and covariances through the continuum momentum equation, making the rendering itself differentiable. The second stage leverages DDIM+ inversion, ControlNet depth and edge conditioning, and TokenFlow-style cross-frame attention to propagate enhanced key-frame features through the whole video, which is what keeps the final frames visually consistent with the input image and with each other.","core_discovery":"The central claim is that physics-grounded 3D dynamics can be generated from a single image by first reconstructing the foreground as a set of 3D Gaussians, time-stepping that representation with a differentiable Material Point Method using elastoplastic material models, and then using a text-to-image diffusion model with cross-frame attention to enhance the coarse simulated frames. The paper reports that this pipeline produces videos that score highest among compared image-to-video baselines on physical commonsense and semantic adherence metrics, and that it supports rigid motion, elastic deformation, viscoplastic flow, granular particles, and fracture from a single RGB image plus an applied force or torque.","pith_inferences":["A direct testable extension is to measure how much the final enhanced video deviates from the coarse MPM simulation; if the deviation is small, the pipeline can double as a controllable physical simulator with photorealistic output.","The same coarse-to-fine recipe could be applied to other physics simulators, such as rigid-body or fluid solvers, and to other single-image 3D representations, since the enhancement stage only needs a sequence of coarse frames.","If the physics is preserved, PhysMotion's outputs could serve as pseudo-ground-truth for training or benchmarking video diffusion models on physical plausibility, an application the paper does not discuss."],"forward_implications":["A user can specify an applied force or torque and watch the object respond with deformable, rigid, granular, or fracturing motion, rather than merely interpolating pixels.","Text- or trajectory-conditioned video models that produce physically implausible motion could be replaced or augmented by this simulation-first recipe in applications like visual effects and game asset animation.","Because the pipeline starts from a single image, it extends physics-based video synthesis beyond the 3D-model or multi-view inputs that earlier physics-grounded methods require.","The quantitative results position the method as a baseline for evaluating physical commonsense in image-to-video generation tasks."],"supporting_citations":[{"why":"Supplies the MPM-integrated 3D Gaussian physics simulation framework that PhysMotion inherits for its coarse dynamics.","marker":"[91]"},{"why":"Provides the coarse single-image 3D Gaussian reconstruction that the pipeline optimizes before simulation.","marker":"[79]"},{"why":"Supplies the cross-frame attention and feature propagation mechanism used to keep the enhanced video temporally consistent.","marker":"[23]"},{"why":"Contributes the hard-depth loss used in geometry-aware reconstruction to improve shape accuracy.","marker":"[42]"},{"why":"Provides the DDIM+ inversion approach that the enhancement stage adapts with image-space blending.","marker":"[99]"},{"why":"Supplies the depth and edge control signals that maintain geometric consistency during diffusion sampling.","marker":"[104]"},{"why":"Generates the multi-view images used to supervise the coarse 3D Gaussian reconstruction.","marker":"[73]"},{"why":"Supplies the physical commonsense and semantic adherence metrics used in the quantitative evaluation.","marker":"[4]"},{"why":"Provides the open-sourced MPM solver that carries out the continuum-mechanics simulation.","marker":"[113]"}],"fun_headline_variants":["Single image gets physics-grounded motion","Physics simulation turns one photo into video","One image to physics-based video","From a single picture to physical motion video"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The diffusion-based enhancement stage must preserve the motion produced by the MPM simulation; if the enhancement noticeably changes trajectories, contact timing, or deformation, the resulting video is no longer physics-grounded.","fun_headline_variants_meta":{"raw":{"variants":["Single image gets physics-grounded motion","Physics simulation turns one photo into video","One image to physics-based video","From a single picture to physical motion video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000773,"raw_usage":{"total_tokens":3386,"prompt_tokens":871,"completion_tokens":2515,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":2464}},"tokens_in":487,"tokens_out":2515,"duration_ms":17977,"temperature":1.0,"reasoning_tokens":2464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:24:26.350330+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one generated scene, such as a falling and bouncing object, and track a few surface points in the final enhanced video; compare those trajectories to the trajectories of the corresponding MPM particles. If the enhanced video's motion diverges from the simulation beyond a small tolerance, or if the video fails to respond appropriately to a changed applied force, the paper's physics-grounded claim is not borne out.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DDIM+ inversion approach that the enhancement stage adapts with image-space blending."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the depth and edge control signals that maintain geometric consistency during diffusion sampling."},{"cited_title":"Image Sculpting: Precise Object Editing with 3D Geometry Control","cited_arxiv_id":"2401.01702","evidence_quote":"Provides the open-sourced MPM solver that carries out the continuum-mechanics simulation."}],"review_version":1}