{"id":"998415af-335d-4d78-b293-199eca1377a5","arxiv_id":"2604.01641","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An interactive pipeline generates expanding 4D worlds with globally coherent environmental dynamics from a single image in roughly 12 seconds per expansion step.","lead":"LivingWorld builds explorable 4D scenes with coherent environmental motion (clouds, water, smoke) from one image by growing a global motion field as the world expands. It targets interactive world creation where prior systems either stay static or need minutes of video-driven optimization.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Global coherence of the progressive motion field is only weakly stress-tested under large expansions and novel viewpoints.","rationale":"The reader correctly isolates the weakest link: that sparse, independently estimated, Kabsch-aligned Eulerian samples suffice to supervise a continuous hash motion field that stays physically plausible under expansion without video refinement. My concern is the same assumption, sharpened to the measurement gap—Table 3 and the ablations only probe consistency on the supervision set itself, not on held-out novel views after multi-step expansion. That gap is load-bearing for the “globally coherent \to interactive 4D” claim, yet does not overturn the empirical evidence of competitive VBench/PhysReal scores, human preference, and ~12 s runtime. Hence the verdict remains CONDITIONAL (pending stronger non-LLM motion metrics and code), with no need to move to REJECT or ACCEPT. The proposed concrete test directly falsifies or corroborates the missing measurement and is feasible with the existing pipeline.","tokens_in":19284,"tokens_out":620,"duration_ms":6516,"concrete_test":"On 10 multi-step expansion sequences, freeze F_\theta after the final update, render the same dynamic Gaussians from 5 held-out novel viewpoints never used for alignment or training, re-estimate 2D Eulerian flow on those renders, lift with the same depth model, and recompute the global cosine similarity and magnitude ratio of Table 3 against the stored F_\theta queries at corresponding 3D points. If either metric drops >15 % relative to the in-distribution numbers, the progressive global-coherence claim is overstated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that independently estimated Eulerian 2D flows, after Kabsch rotation/scale alignment of sparse lifted samples (Eq. 4, Sec. 3.2) and hash-field regression (Eqs. 5–7), yield a continuous F_\theta whose bidirectional Euler advection (Eqs. 8–12) remains directionally and magnitude-consistent across large expansions and novel viewpoints without video refinement. Table 3 and Fig. 5 measure local MCA/FMV and global cosine/magnitude ratio only on the accumulated samples themselves (or nearby pairs), not on held-out novel camera trajectories far from the training views. Qualitative multi-view results (Fig. 4) and VBench/PhysReal scores evaluate rendered appearance, not whether the underlying 3D velocity field stays coherent once the scene has expanded several steps. Because depth (MoGeV2), flow (3D-Cinemagraphy EF), and SAM masks are off-the-shelf and never jointly refined, residual scale/direction errors can accumulate; the closed-form Kabsch + 300-iter SGD refinement only aligns overlapping reprojected points and does not guarantee consistency in newly outpainted regions. Thus the engineering claim of “globally coherent environmental dynamics” under interactive expansion rests on an incompletely measured assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"LivingWorld is an interactive systems framework that generates 4D worlds with scene-scale environmental dynamics (clouds, water, smoke/fog, fire) from a single image. Starting from a static 3D reconstruction pipeline in the WonderWorld style, it progressively builds a continuous global Eulerian motion field: user-guided 2D Eulerian flows are lifted with depth, aligned across views by a geometry-aware Kabsch rotation/scale module plus light refinement (Eq. 4), and regressed into a multi-resolution hash-encoded field F_θ (Eqs. 5–7). During rendering, Gaussians are advected bidirectionally with a linear opacity scheduler (Eqs. 8–12) to produce looping, temporally coherent dynamics without video-based refinement. On a 60-scene benchmark the method is competitive on VBench, best on GPT-based PhysReal, preferred in 2AFC human studies, and completes each expansion step in ~12 s on one RTX 5090 versus minutes for video and prior 4D baselines.","tokens_in":19702,"tokens_out":1462,"duration_ms":19261,"significance":"The paper targets a genuine and timely gap: interactive large-scale 3D world generation has largely produced static geometry, while environmental dynamics that are intrinsically scene-coupled remain underexplored under interactive latency constraints. The contribution is primarily engineering—geometry-aware alignment of sparse scene flow, a compact hash motion field decoupled from per-Gaussian optimization, and bidirectional propagation that avoids expensive video supervision—but it is well motivated and practically useful. Strengths include a clear runtime breakdown (~9 s expansion + ~3 s motion update), ablations of alignment and motion-field design (Table 3, Fig. 5–6), both automatic and human evaluation (Tables 1–2, interactivity Likert study), and an honest statement that no prior method solves the same interactive 4D task. If the global-coherence claim holds under multi-step expansion, the work is a solid systems advance for interactive 4D content creation and a useful foundation for later object-centric dynamics.","major_comments":[{"comment":"The central claim of “globally coherent environmental dynamics” under progressive expansion is only partially stress-tested. Table 3 and Fig. 5 report MCA/FMV and global cosine/magnitude ratio on accumulated (or nearby) scene-flow samples after alignment, not on held-out novel camera trajectories after several expansion steps far from the views that supervised F_θ. Qualitative multi-view results (Fig. 4) and VBench/PhysReal scores assess rendered appearance, not whether the underlying 3D velocity field remains directionally and magnitude-consistent once outpainting has revealed large new regions. Because depth (MoGeV2), Eulerian flow (3D-Cinemagraphy), and SAM masks are off-the-shelf and never jointly refined, residual scale/direction errors can accumulate; Kabsch+refinement (Eq. 4, Sec. 3.2) only aligns overlapping reprojected points. Please add either (i) a multi-step expansion protoco","section":null},{"comment":"Sec. 3.3’s bidirectional Euler advection plus linear opacity schedule (Eqs. 8–12, w(t)=t/T) produces temporally looped dynamics that return Gaussians toward the initial configuration. This is effective for hole reduction (Table 6: 9.8%→0.9%) and seamless looping, but it is not open-ended continuous evolution. The abstract and introduction frame the output as “long and temporally coherent 4D sequences” and “interactive 4D world generation,” which can be read as supporting non-periodic, indefinitely evolving environments. Please clarify in the method and evaluation that the current renderer is designed for stable looping environmental motion, and discuss how (or whether) the same field supports non-looping long-horizon simulation without reintroducing density gaps.","section":null}],"minor_comments":[{"comment":"Table 1 reports LivingWorld runtime as 12 s against multi-minute baselines; the comparison is informative for interactivity but conflates different task formulations (video synthesis vs. explicit 4D scene). A short note in Sec. 4.4 that video models are not interactive 4D systems (already stated in Sec. 4.1) would prevent over-reading the runtime column as a pure quality–speed Pareto claim.","section":null},{"comment":"PhysReal (and PhotoReal in Supp. Table 4) rely on GPT-based scoring. The paper already includes VBench and 2AFC human preference; still, briefly state the prompt template or scoring protocol (or point more explicitly to PhysGen3D) so the metric is reproducible.","section":null},{"comment":"Eq. (5) uses a product/XOR hash notation that is slightly hard to parse in the rendered text (“3M j=1”, “L denotes bitwise XOR”). Align the notation with Instant-NGP-style multi-resolution hash encoding and define all symbols (π_j, T, ℓ) in one place.","section":null},{"comment":"Fig. 3’s caption and body text describe the geometry-aware alignment module, but the figure itself is largely placeholder-like in the manuscript text. Ensure the camera-ready figure clearly shows before/after alignment of scene-flow directions.","section":null},{"comment":"Sec. 4.7 and Fig. 15 correctly note that rigid-body motion is outside the primary scope and can distort geometry. Consider moving a one-sentence limitation of this kind into the main conclusion so readers do not over-generalize the object-centric demo.","section":null},{"comment":"Minor typos/consistency: “LivingW orld” spacing artifacts in the abstract and Table 1; “image sequences” spacing in Sec. 1; arXiv date stamps in the header (v2, Jul 2026) should be cleaned for the journal version.","section":null}],"recommendation":"minor_revision","confidential_remarks":"This is a competent interactive systems paper rather than a deep theoretical contribution: hash encoding, Eulerian cinemagraph flow, and Kabsch alignment are known ingredients, combined for interactive environmental 4D expansion. That combination and the latency result are still valuable for a CV systems venue. I would not reject for lack of a true interactive 4D competitor—the authors state the gap honestly—but I would insist on clearer measurement or scoping of multi-step global motion coherence before final acceptance. Fit is good for a vision/graphics journal that values interactive 3D/4D generation."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean engineering contribution, not a theory paper. The new piece is the full interactive loop: WonderWorld-style expansion plus Eulerian 2D flow lifted to 3D, closed-form Kabsch rotation/scale alignment of sparse scene-flow samples (Eq. 4), a multi-res hash Eulerian field F_θ (Eqs. 5–7), and bidirectional Euler advection with a linear opacity schedule (Eqs. 8–12) so Gaussians do not leave density holes. That combination is not in the cited priors as a finished interactive system, and the runtime claim is real—~9 s expansion + ~3 s motion update on one RTX 5090 versus minutes for the video and 4D baselines (Table 1, runtime breakdown).\n\nWhat they do well: honest baseline framing (no true interactive 4D competitor), ablations that actually move the needle (Table 3 / Fig. 5 on alignment; Fig. 6 on hash field vs raw flow and forward-only), 2AFC human preference that favors them on motion/flicker, and a short object-centric integration demo. Math is standard and correctly applied; citations cover WonderWorld, 4DGS-Cinemagraphy, Eulerian cinemagraphs, Instant-NGP, and the video generators without obvious padding. Circularity is low—off-the-shelf SAM/MoGe/EF supply supervision, then perceptual and human metrics evaluate the result.\n\nSoft spots, in proportion. The stress-test note is partly right: MCA/FMV and global cosine/mag ratio are measured on accumulated samples (or nearby pairs), not on held-out novel trajectories after several expansion steps, so “globally coherent under large expansion” is supported more by multi-view renders and VBench/PhysReal than by direct 3D velocity-field checks far from training views. Residual depth/flow error can accumulate; Kabsch only aligns overlapping reprojected points. PhysReal is GPT-based, so treat it as a soft signal. No public code yet. None of these sink the engineering claim; they bound how far you should trust the physics story.\n\nWho it is for: people building interactive world tools, simulators, or embodied-AI environments who need fluid-like scene motion at interactive latency. Worth a serious referee. I would bring it to reading group and would cite the pipeline if I work in this space. Accept for peer review.","headline":"Solid interactive systems paper: progressive hash motion field + Kabsch alignment gets environmental 4D dynamics to ~12 s/step with competitive quality; global-coherence claim is only partly stress-tested.","tokens_in":20282,"tokens_out":598,"would_cite":true,"duration_ms":6633,"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":"LivingWorld builds interactive 4D worlds with coherent clouds, water, and smoke from one image by growing a global motion field in seconds.","keywords":["interactive 4D world generation","environmental dynamics","global motion field","geometry-aware alignment","hash-based motion field","bidirectional motion propagation","Gaussian Splatting","single-image animation"],"falsifier":"Generate a large multi-view expansion of a scene with strong environmental motion (e.g., a river or drifting clouds), render novel trajectories never used during construction, and check whether motion directions and magnitudes remain consistent with the original views and with physical expectations; systematic drift, holes, or reverse-direction artifacts would falsify the claim that alignment-plus-hash field is sufficient without video refinement.","tokens_in":20174,"feed_emoji":"🌍","tokens_out":1015,"duration_ms":8540,"temperature":0.7,"pith_summary":"Most interactive 3D scene generators rebuild geometry and appearance from a single photo but leave the world static: rivers do not flow and clouds do not drift. LivingWorld claims that environmental dynamics can be added interactively by progressively constructing one continuous global motion field as the user expands the scene with new camera views. Newly estimated 2D flow is lifted to 3D, aligned across views by a geometry-aware module that corrects direction and scale, then stored in a compact multi-resolution hash field that can be queried anywhere. Bidirectional advection of Gaussians plus a simple opacity schedule then produces long looping 4D sequences without expensive video refinement. On one high-end GPU each expansion step finishes in about twelve seconds, and the resulting motion scores highest on physical-plausibility metrics while remaining competitive on visual quality. A sympathetic reader cares because this turns static world generators into living environments that respond to user motion hints in near real time.","feed_headline":"One image becomes a living 4D world in about 12 seconds","feed_subtitle":"A growing global motion field keeps clouds, water and smoke coherent as the user expands the scene.","key_machinery":"The hash-based global motion field F_θ : R³ → R³, trained on scene-flow samples that have first been aligned across views by a Kabsch-plus-refinement geometry-aware module; it supplies continuous velocity queries so Gaussians can be advected forward and backward while an opacity scheduler keeps density and temporal continuity.","core_discovery":"Environmental dynamics that span an expanding 3D scene can be represented as a single continuous global motion field that is built incrementally from user-guided Eulerian flow, kept consistent by geometry-aware 3D alignment of sparse scene-flow samples, and rendered by bidirectional propagation, yielding temporally coherent 4D worlds at interactive speeds without video-based optimization.","pith_inferences":["If the alignment step is the main bottleneck for global consistency, replacing or augmenting Kabsch with learned cross-view flow correspondences could further reduce residual directional errors in non-overlapping regions.","The same progressive hash-field construction may transfer to other continuous scene attributes (temperature, wind fields, soft-body deformation) that must stay coherent under interactive expansion.","Because the method never relies on multi-view video supervision, it could serve as a lightweight motion prior that later video models fine-tune only where higher fidelity is required.","Boundary-hole rates already drop sharply with bidirectional propagation; adding a lightweight density-repair step would likely close the remaining gaps without sacrificing interactivity."],"forward_implications":["Interactive world builders can add user-controllable environmental motion (clouds, water, smoke, fire) as a native 3D layer rather than as post-hoc video synthesis.","Each new camera expansion can update both geometry and a coherent global motion field in roughly twelve seconds on a single modern GPU, supporting real-time feedback loops.","Long looping 4D sequences become available without the minutes-to-hours cost of video-driven 4D Gaussian optimization.","Localized object motion can be overlaid on the same environmental field, giving a path toward mixed rigid and fluid dynamics in one scene.","Downstream simulation, perception, or embodied-agent pipelines gain access to dynamic environments that remain geometrically consistent under free camera navigation."],"fun_headline_variants":["One image to interactive 4D world with coherent clouds water smoke","Expanding global motion field yields living 4D scenes from a single view","Interactive 4D worlds grow with consistent environmental dynamics","Single image seeds 4D world whose motion stays coherent as it expands","Geometry-aware motion field makes 4D dynamics interactive and global"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That independent 2D Eulerian flow estimates, after a simple rotation-and-scale alignment of the sparse 3D samples they produce, already give enough correct supervision for a continuous motion field whose bidirectional advection stays plausible under large expansions and new viewpoints without any later video refinement.","fun_headline_variants_meta":{"raw":{"variants":["One image to interactive 4D world with coherent clouds water smoke","Expanding global motion field yields living 4D scenes from a single view","Interactive 4D worlds grow with consistent environmental dynamics","Single image seeds 4D world whose motion stays coherent as it expands","Geometry-aware motion field makes 4D dynamics interactive and global"]},"model":"grok-4.5","effort":"low","cost_usd":0.005902,"raw_usage":{"total_tokens":1557,"prompt_tokens":764,"num_sources_used":0,"completion_tokens":93,"cost_in_usd_ticks":59020000,"prompt_tokens_details":{"text_tokens":764,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":700,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":764,"tokens_out":93,"duration_ms":5808,"temperature":1.0,"reasoning_tokens":700,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T14:16:56.773722+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Generate a large multi-view expansion of a scene with strong environmental motion (e.g., a river or drifting clouds), render novel trajectories never used during construction, and check whether motion directions and magnitudes remain consistent with the original views and with physical expectations; systematic drift, holes, or reverse-direction artifacts would falsify the claim that alignment-plus-hash field is sufficient without video refinement.","supporting_citations":[],"review_version":1}