{"id":"ae4518c5-0186-4c22-9faa-9a250d74408f","arxiv_id":"2605.20290","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A training-free pipeline reconstructs a single image into a physically simulated 3D scene and re-renders simulated frames into a controllable, photorealistic video.","lead":"A single photo can be turned into a 3D scene where objects fall, collide, and stack under user-controlled forces, then re-rendered as realistic video. The paper's evidence is not yet independently checkable because the code, test set, and evaluation details are not fully released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"VLM material estimation is explicitly biased toward deformable materials, undercutting the claimed physical accuracy of rigid-body multi-object interactions.","rationale":"The reader's weakest_assumption identified VLM-inferred material parameters as a key concern. I agree, and I sharpen it: the VLM prompt in Appendix D.1 does not merely guess heuristically; it actively prefers deformable interpretations for all non-immovable objects. This is a load-bearing flaw because the paper's central claim is physical consistency and accurate interaction, and rigid-body behavior is prominently featured in the qualitative results (dominoes, bowls, stacked boxes). The paper's own limitations statement admits material parameters are not directly observable, but the prompt's bias is stronger than an occasional miss. The current quantitative evidence cannot rule this out because Physical Commonsense is scored by a GPT-5 VLM on 10 sampled frames, which may reward visually plausible motion rather than correct mechanical behavior. The proposed test directly compares prompted estimation, neutral estimation, and ground-truth parameters against real-world rigid-body dynamics; this would settle whether the bias actually changes outputs. I therefore recommend keeping the reader's CONDITIONAL verdict, since the concern is concrete but potentially fixable and the paper may still be substantially correct in its alignment and controllability contributions.","tokens_in":24150,"tokens_out":3759,"duration_ms":45752,"concrete_test":"Build a benchmark of 20 simple scenes containing known rigid objects (steel ball, ceramic mug, wooden block, glass bottle) with known ground-truth mass, friction, and Young's modulus. Run TelePhysics three ways: (1) original Appendix D.1 prompt; (2) a neutral prompt that chooses rigid materials when warranted; (3) ground-truth parameters injected directly into Genesis. Compare simulated trajectories, rest configurations, and contact forces against high-speed video or analytic rigid-body solutions. If variant (1) produces visible deformations, changed stacking stability, or >10% trajectory deviation relative to variant (3), the claim of physically accurate multi-object interactions is not supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract claims TelePhysics produces 'physically consistent' multi-object dynamics and 'accurate scene-level multi-object interactions.' This hinges on per-object mass, friction, and Young's modulus inferred by Qwen2.5-VL-72B. However, Appendix D.1's system prompt explicitly instructs: 'IMPORTANT: Prefer non-rigid materials --- choose the MOST DEFORMABLE plausible interpretation,' and reserves 'rigid' for 'immovable structures.' This is a systematic bias, not a neutral heuristic. A ceramic bowl or glass bottle will be simulated as an MPM elastic body with default E=3e5 Pa and rho=1000 kg/m^3, not as a rigid body, so collision, stacking, and support behavior will differ materially from true physics. The Limitations section acknowledges that material parameters 'are not always directly observable from a single image and may require heuristic initialization,' but the prompt's explicit preference for deformability is stronger than a fallback; it guarantees common rigid-object scenes are mis-specified. Since the reported Physical Commonsense scores come from a GPT-5 judge rather than ground-truth trajectories, this systematic material bias would not be detected by the current evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"TelePhysics is a training-free pipeline that takes a single RGB image, segments and reconstructs foreground objects as meshes, reconstructs an inpainted/outpainted background, aligns the meshes into a common world frame using a RANSAC/AGMF ground-plane estimate and a coarse-to-fine camera pose optimization, infers per-object material parameters and force fields with a VLM (Qwen2.5-VL-72B), simulates the resulting scene with Genesis (RBD/MPM/PBD solvers), and optionally re-renders the simulation frames into photorealistic video using a video diffusion model running only the final denoising steps. The authors claim that the unified scene representation resolves object penetration and alignment ambiguity and yields physically consistent, controllable multi-object interactions, with real-time interactive previews after a one-time initialization. The paper reports quantitative GPT-5-based scores, a human ranking study, ablations, and runtime figures on a 60-scene test set.","tokens_in":24457,"tokens_out":4443,"duration_ms":50162,"significance":"If the claims hold, the work is a practically useful integration of explicit 3D reconstruction, forward physics simulation, and generative video re-rendering, with the advantage of being training-free and providing explicit control over forces and velocities. The proposed AGMF plane-fitting and coarse-to-fine camera alignment are sensible and are supported by internally consistent ablation results. The release of code and the use of a standard physics engine (Genesis) are strengths. However, the evidence for the central claim of 'accurate scene-level multi-object interactions' is weakened by two load-bearing gaps: (i) the VLM material estimation is explicitly biased toward deformable interpretations, so many everyday rigid objects are simulated with the wrong constitutive model; and (ii) the main quantitative evaluation rests on an unvalidated GPT-5 judge with no error bars, a small test set, and human-study details that are not reported. These issues can likely be fixed within the manuscript's scope, but they require more than local copy-editing.","major_comments":[{"comment":"The VLM system prompt explicitly instructs the model to 'Prefer non-rigid materials --- choose the MOST DEFORMABLE plausible interpretation' and to reserve 'rigid' only for immovable structures. This is a systematic bias, not a neutral heuristic. Consequently, objects such as the ceramic bowls in Fig. 9, glass bottles, or stacked boxes will likely be simulated as MPM/PBD deformable bodies with default parameters (e.g., E=3e5 Pa, rho=1000), rather than as rigid bodies. Collision, support, and stacking dynamics will therefore differ materially from true physics. Because the paper claims 'accurate scene-level multi-object interactions' and 'physically consistent' dynamics in the abstract, this bias is load-bearing. The Limitations section (Sec. 5) acknowledges that material parameters 'may require heuristic initialization,' but it does not disclose that the default prompt systematically pre","section":"Appendix D.1, Sec. 3.3"},{"comment":"The primary quantitative evidence (Table 1) is a GPT-5-based 5-point Likert score. The manuscript does not validate this judge against human ratings or ground-truth physics, does not report per-scene variance or confidence intervals, and uses only 60 scenes. The VideoPhy protocol cited as the basis is a benchmark for evaluating video generation models, not a certified judge. The prompt in Appendix F even says 'Eight sets of 10 evenly spaced frames' while only seven methods are compared. Without validation, the reported superiority in Physical Commonsense and Semantic Adherence could reflect the judge's priors rather than actual physical correctness. I recommend adding human-machine agreement on a subset, reporting error bars / significance tests, and complementing the LLM judge with quantitative trajectory- or contact-based metrics on scenes where ground-truth or analytic expectations ex","section":"Sec. 4.3, Appendix F"},{"comment":"The human evaluation reports Borda scores but omits essential details: number of raters, number of scenes rated by each rater, whether raters were naive or expert, and inter-rater reliability. The reported margins (e.g., Physical Commonsense 6.90 for Ours vs. 2.05 for WonderPlay) are large and, in the absence of rater-count information, impossible to interpret. At minimum, the authors should report n, a measure of agreement (e.g., Fleiss' kappa or Kendall's W), and the distribution of ranks. A small number of raters or a single rater could not support the strength of the claim.","section":"Sec. 4.4, Table 2"},{"comment":"WonderTrace re-renders the raw simulation with a video diffusion model, which can alter object trajectories, contacts, and deformations. The 'Geometric fidelity' evidence in Table 8 is limited to SSIM and centroid displacement on 15 videos; it does not measure whether the final video obeys the specified forces, collisions, or stacking relations. A reader cannot tell whether the photorealistic video truly inherits the simulation's dynamics or merely resembles the raw simulation visually. Since the central claim is controllable and physically consistent output, please provide trajectory/contact fidelity metrics on the final re-rendered videos, or a comparison of final-video object masks/poses against the simulated ones over time.","section":"Sec. 3.4, Table 8"}],"minor_comments":[{"comment":"The manuscript title says 'TelePhysics' while the abstract and project page say 'PhysOmni'. This inconsistency must be resolved in the final version.","section":"Title/Abstract"},{"comment":"The main text says a '60-scene test set' while Appendix F says '59-scene test set'. Please correct the count.","section":"4.1 vs. F.1"},{"comment":"The evaluation prompt states 'Eight sets of 10 evenly spaced frames', but only seven methods are compared. Also, the instruction that 'static or nearly-static sequences are less preferred' in Video Quality may bias scores against physically correct scenes where objects come to rest; consider neutral wording.","section":"Appendix F.2"},{"comment":"The AGMF anchor set in Eq. (6) depends on a threshold δ but no default value or estimation rule is given; Appendix B instead describes selecting the lowest p% of points. Please clarify how δ is chosen and how it relates to the lowest-p% RANSAC strategy.","section":"Eq. (6) and Appendix B"},{"comment":"The captions contain artifact text such as 'RA/glyph1197SAC Mean Plane Fitting' and '/glyph1197ormal Orientation Analysis', likely from font corruption. These should be fixed.","section":"Figs. 4 and 5"}],"recommendation":"major_revision","confidential_remarks":"The framework has merit and the ablations are internally consistent, but the paper currently overclaims physical accuracy while the material estimation is explicitly biased toward deformables, and the headline quantitative evaluation is an unvalidated LLM judge on a small test set. These are fixable. I would not reject, but I would require the authors to address the material-bias issue and strengthen the evaluation evidence before publication. The title inconsistency between 'TelePhysics' and 'PhysOmni' should also be resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you're considering whether single-image-to-VFX pipelines are ready for prime time. The main contribution is an integrated, training-free pipeline: SAM3/SAM3D meshes, LaMa background, AGMF ground-plane fitting, scene-aware pose alignment, coarse-to-fine camera optimization, Genesis simulation, Wan VACE re-rendering. None of the blocks are new, but the combination is, and the alignment parts are genuinely useful. The ablations show AGMF beats vanilla RANSAC on support violations and penetration, and the camera optimization reduces reprojection error from ~39 to ~14 px. That's real engineering work, presented clearly.\n\nThe soft spots are mostly in the evidence and one design choice. The headline evaluation is a GPT-5 judge with no validation, on a 60-scene set, with single-point scores and no error bars. The human study doesn't report rater count. The code is claimed open-source but I couldn't verify, and the paper is internally inconsistent on the name: PhysOmni in the abstract, TelePhysics throughout.\n\nThe bigger issue is the material estimation. Appendix D.1's system prompt tells the VLM to 'Prefer non-rigid materials... choose the MOST DEFORMABLE plausible interpretation,' and reserves rigid for immovable structures. So a ceramic bowl or glass bottle becomes an MPM elastic body with E=3e5 Pa and rho=1000, not a rigid body. That systematically biases collision, stacking, and support behavior away from true physics for the exact rigid-object scenes the paper highlights. The Limitations section acknowledges material parameters are heuristic, but this is a deliberate preference, not a neutral fallback. And because the physical-commonsense score comes from a GPT-5 judge, the bias would pass unnoticed. That doesn't sink the pipeline as a preview tool, but it does sink the 'accurate scene-level multi-object interactions' claim.\n\nWho's this for: people building interactive physics previews from a single image, whether VFX previsualization, game prototyping, or synthetic data for embodied AI. They'll get a working recipe and a useful alignment module. They should not cite it as evidence that VLM-estimated physics is accurate.\n\nSerious referee: yes, but it needs a major revision. Release code and data, validate or replace the GPT-5 judge, report the human study properly, and either remove the deformable bias or soften the physical-accuracy claim to 'visually plausible.' I'd send it to peer review.","headline":"Solid integrated pipeline and honest ablations, but the 'physically accurate' claim is undercut by the VLM prompt that forces deformable materials on rigid objects.","tokens_in":24952,"tokens_out":2361,"would_cite":true,"duration_ms":25548,"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":"One photograph contains enough information to generate a physically consistent, user-controllable multi-object video.","keywords":["physics-based video generation","single-image to 3D","scene-level reconstruction","multi-object dynamics","pose alignment","real-time interaction","diffusion re-rendering","material parameter estimation"],"falsifier":"Choose an object whose material is ambiguous in a single photo (for example, a foam sphere versus a glass sphere of the same color). Run the pipeline on that image with a fixed force and compare the predicted trajectory to the measured trajectory of a physically real copy of the object; if the predicted trajectory matches the wrong material's behavior, the claim of physical consistency is refuted. A cheaper variant: feed the same image to two different material estimators and show the output videos diverge while both claim physical plausibility.","tokens_in":24031,"feed_emoji":"🎬","tokens_out":5870,"duration_ms":58759,"temperature":0.7,"pith_summary":"The paper (PhysOmni, called TelePhysics in the full text) claims that a single photograph can be turned into a video in which several objects interact physically and in a way the user can steer. Its pipeline works in stages: reconstruct each visible object as a 3D mesh, place all meshes in one shared world coordinate frame anchored to a common ground plane, infer physical properties with a vision-language model, simulate dynamics with an explicit multi-solver physics engine, and finally re-render simulation frames with a video-diffusion model so the result looks photorealistic while keeping the simulated motion. The paper argues that this unified spatial frame, rather than object-by-object lifting, is what removes interpenetration and alignment ambiguity, and that decoupling simulation from rendering allows interactive previews at roughly 15 frames per second before final rendering. If correct, this is a practical route from static images to controllable, physics-grounded video without training a new generative model.","feed_headline":"One photo becomes an interactive physics scene at 15 FPS","feed_subtitle":"Rebuilding all objects in one shared 3D world frame lets collisions, stacking, and force fields behave plausibly.","key_machinery":"The load-bearing object is a unified world coordinate system with a canonical ground plane. It is constructed by Anchor-Guided Manifold Fitting (AGMF), which estimates the ground from the lowest vertices of each object rather than from all points, avoiding the bias of naive plane fitting toward vertical structures; a coarse-to-fine camera-pose optimizer then aligns rendered silhouettes to the input so the scene matches the photograph's perspective. Physics runs in this frame through three coupled solvers—rigid body dynamics for non-deformable objects, material point method for fluids and deformable materials, and position-based dynamics for cloth—parameterized by a vision-language model that","core_discovery":"On its own terms, the paper claims that the main failures of prior physics-grounded generation—physically uncontrollable motion, object interpenetration, spatial misalignment, cartoonish appearance, inconsistency with the input image—share a root cause: objects are not reconstructed and anchored in one common world frame. The central discovery is that holistic scene-level reconstruction combined with scene-aware pose alignment and coarse-to-fine camera-pose optimization resolves those failures. Once every mesh sits in the same gravity-aligned coordinate system with the ground at z=0, a physics engine can simulate accurate multi-object contact, support, collision, and stacking, and a partial-","pith_inferences":["The paper demonstrates physical plausibility, not physical accuracy: if material parameters guessed by the vision-language model are wrong, the result can still look physically reasonable even though the true object would behave differently. A direct test would compare predicted trajectories against real drop or bounce experiments for objects of known material.","The same unified-frame idea could be injected into a trained video generator as an architectural bias, for example by conditioning on a ground-aligned scene representation, potentially reducing the need for a separate simulator at test time while retaining long-horizon consistency.","The modular split (perception, alignment, simulation, re-rendering) suggests each stage can be swapped independently; replacing the vision-language material estimate with an estimator that uses multiple views or physical measurements should improve fidelity without changing the rest of the pipeline.","Because the initial reconstruction is monocular, the relative gains over video-prior baselines may partly reflect the vision-language model's prior about objects rather than true measured physics."],"forward_implications":["Users can specify force fields (gravity, wind, vortex, turbulence) or initial velocities and see multi-object consequences that respect contact and support, not just appearance.","Scene-level grounding eliminates the interpenetration and drift that plague independently lifted objects, so stacking and collision chains can survive long simulations.","Because simulation is decoupled from rendering, an interactive preview loop runs at roughly 15 FPS on a single H100 GPU, with the photorealistic re-render as a separate offline step.","The re-rendering stage preserves the simulator's geometry: reported silhouette drift is near zero and mean object-centroid displacement is about 6% of image width.","The framework is training-free, so new scenes can be initialized once (perception plus alignment) and then simulated at interactive rates without per-scene fine-tuning."],"fun_headline_variants":["One photo becomes a real-time physics playground","Single image to interactive physics scene: real-time, collision-aware","PhysOmni: one image yields grounded multi-object physics at 15 FPS","Training-free: from one photo to a physical world you can manipulate","Unified 3D coordinates make single-image physics scenes real-time"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a single image, together with a vision-language model's guesses, contains enough information about geometry and material to specify correct physics; if the guessed mass, friction, or stiffness is wrong (as the paper's own limitations concede), the resulting motion will look plausible without being physically accurate.","fun_headline_variants_meta":{"raw":{"variants":["One photo becomes a real-time physics playground","Single image to interactive physics scene: real-time, collision-aware","PhysOmni: one image yields grounded multi-object physics at 15 FPS","Training-free: from one photo to a physical world you can manipulate","Unified 3D coordinates make single-image physics scenes real-time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1462,"prompt_tokens":731,"completion_tokens":731,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":643}},"tokens_in":475,"tokens_out":731,"duration_ms":8486,"temperature":1.0,"reasoning_tokens":643,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T13:38:12.083418+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose an object whose material is ambiguous in a single photo (for example, a foam sphere versus a glass sphere of the same color). Run the pipeline on that image with a fixed force and compare the predicted trajectory to the measured trajectory of a physically real copy of the object; if the predicted trajectory matches the wrong material's behavior, the claim of physical consistency is refuted. A cheaper variant: feed the same image to two different material estimators and show the output videos diverge while both claim physical plausibility.","supporting_citations":[],"review_version":2}