{"id":"7c5161e4-8270-49d5-bcc0-070ab12440d8","arxiv_id":"2502.09631","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"VNCA trains a 3D neural cellular automaton on a single smoke frame and generalizes to stylize the entire sequence, and even unseen smoke datasets, in near real-time.","lead":"This paper introduces Volumetric Neural Cellular Automata (VNCA), a method that stylizes 3D smoke simulations using a reference style image while keeping the result consistent over time and across camera views. It matters for visual effects and animation because it claims to reduce training time by over an order of magnitude compared with existing volumetric style transfer approaches.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central temporal-coherence claim is not yet established: motion supervision uses 2D rendered optical flow, while the local 3×3×3 update rule can move texture at most one voxel per update, and no 3D motion metric is reported.","rationale":"The paper's central and differentiating claim is that VNCA replaces explicit fluid advection with learned automaton motion that aligns with the smoke. Temporal coherence, multi-view motion consistency, and generalization to unseen smoke all depend on this alignment. The reader correctly identifies the 2D projection issue in Section 3.4 as the weakest assumption. I agree and add a concrete mechanical reason: the local update rule cannot move features more than one voxel per update, so the model must use enough update steps to match the smoke speed, and the paper provides no such measurement or argument. The proposed multi-view scene-flow test would settle whether the concern lands: if the recovered 3D flow matches the ground-truth velocity, the mechanism is credible; if not, the central claim fails. Secondary issues are also real—the 0.8 s/frame runtime contradicts the abstract's 'real-time' claim, the training-time speedup is asserted without measurement, and the user study lacks significance testing—but those are presentation/quantification problems that do not attack the mechanism itself. Since the reader's CONDITIONAL verdict already requires additional validation, I recommend no change to the verdict.","tokens_in":128,"tokens_out":10895,"duration_ms":179439,"concrete_test":"Run VNCA on a smoke sequence with known velocity and maximum speed of several voxels per frame (e.g., 4–8). Render pairs of consecutive inference frames from 8–12 calibrated camera views, compute 2D optical flow, and back-project the correspondences into 3D scene flow. Compare the recovered per-voxel trajectories against the ground-truth velocity integrated over one frame, using mean endpoint error and median angular error. Also check whether the achieved displacement saturates when the flow magnitude is large, independent of N. If mean endpoint error exceeds roughly one voxel per frame, or if the recovered 3D motion does not match the velocity field in direction and magnitude, the 2D flow supervision in Eq. 14 is insufficient to enforce true 3D advection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing mechanism is the flow-guided motion supervision in Section 3.4 (Eqs. 12–14). It aligns the emergent NCA motion with the smoke by matching 2D optical flow of rendered views to a 2D projection of the input velocity field. This is not sufficient to establish the central claim of true 3D advection. First, rendered optical flow is an aggregate over all voxels along each ray; a local update can create and destroy texture elements in depth while still producing a plausible 2D projection, so the loss is necessary but not sufficient for 3D correspondence. Second, the update rule in Eqs. 2–3 uses only 3×3×3 Sobel/Laplacian stencils and an MLP that writes to the center cell; no operation transports a cell state across voxels, so a pattern can shift at most one voxel per update. If the smoke velocity exceeds one voxel per frame, the model can only keep up if N update steps are at least the maximum velocity in voxels/frame, but the paper never reports N or the velocity magnitudes. The provided evidence does not close this gap: the multi-view ablation in Figure 6 evaluates static style only, and the generalization results in Figure 10 are qualitative. Thus the claim that VNCA dynamically aligns the texture volume with the smoke motion is unvalidated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Volumetric Neural Cellular Automata (VNCA), a 3D extension of neural cellular automata for stylizing volumetric smoke simulations from a single 2D reference image. VNCA maintains a per-voxel state with RGB and density-residual channels, updates it with a shared MLP acting on 3D Sobel/Laplacian perception vectors, and conditions the update on positional, density, and velocity encodings. A differentiable NeRF-style volume renderer turns the state into images, and training uses a VGG-based appearance loss together with an optical-flow motion loss that aligns rendered 2D motion with the projected input velocity field. The authors claim that, after training on a single density frame, VNCA stylizes the whole sequence in real-time, is multi-view consistent and temporally coherent, generalizes to unseen frames and datasets, and reduces training time by over an order of magnitude compared with prior volumetric style transfer methods. The paper also demonstrates an extension to solid mesh texturing.","tokens_in":12892,"tokens_out":6072,"duration_ms":53941,"significance":"The core idea is appealing: replacing explicit fluid advection and inter-frame smoothing with the emergent dynamics of a small recurrent neural cellular automaton is a genuinely different route to volumetric smoke stylization, and the paper contains several useful ablations (multi-view supervision, density encoding, velocity encoding, and motion-loss weighting) that support the existence of the proposed mechanism. If the claims were established, the practical value would be high, since training on one frame and generalizing to sequences and datasets would give artists a lightweight stylization tool. However, the paper does not currently support its strongest claims: 'real-time' is contradicted by Table 2's 0.8 s/frame, the training speedup is asserted without any training-time measurement, and the central temporal-coherence claim rests on 2D projected-flow supervision with no reported 3D motion correspondence. These gaps are substantial but appear fixable with additional measurement, reporting, and re-framing, so I recommend major revision rather than rejection.","major_comments":[{"comment":"The abstract and Section 1 claim that VNCA synthesizes stylized smoke 'in real-time', but Table 2 reports 0.8 s/frame for VNCA on the Smoke Jet dataset, which is roughly 1.25 frames per second. This is not real-time in any standard sense. The authors should either report a runtime that meets a stated frame-rate target, provide a head-to-head comparison at a common resolution, or qualify the claim as near-real-time or as interactive per-frame processing with a clear operating definition.","section":"Abstract and Table 2"},{"comment":"The abstract and Section 1 claim that VNCA reduces training time by 'over an order of magnitude', but no training-time measurement is reported anywhere in the paper. Section 4.2 reports only per-frame inference time in Table 2; the training times of prior works (20 hours, 70 hours, 18 hours) are quoted from the literature in Section 2.3 rather than measured on the same hardware. The training-speedup claim needs an end-to-end wall-clock comparison of VNCA and the baselines on the same GPU and dataset.","section":"Abstract, Sections 1 and 4.2"},{"comment":"The temporal-coherence claim is not yet established. The motion loss in Eqs. 12-14 compares 2D optical flow of rendered views with a 2D projection of the input velocity field, while the update rule in Eqs. 2-3 uses only 3x3x3 Sobel/Laplacian stencils and an MLP writing to the center cell, so the texture can move at most one voxel per update step. The paper never reports N (the number of VNCA update steps applied between adjacent density frames at inference) or the velocity magnitudes in voxels/frame, so there is no check that the automaton can keep up with the smoke advection. The provided visual evidence does not close this gap: Figure 6 evaluates static multi-view consistency only, and Figure 10 is qualitative. A quantitative 3D motion metric (for example, comparing rendered optical flow against projected velocity from multiple camera angles, or measuring divergence of the synthesized texture from the advected density field) and a stated relation between N, voxel velocities, and the stencil size are needed to support the claim that VNCA dynamically aligns in 3D with the smoke motion.","section":"Section 3.4 (Eqs. 12-14) and Section 4"},{"comment":"The user study that supports the inter-frame consistency and motion-preservation claims is reported as raw percentages with no confidence intervals, significance tests, or description of randomization and viewing conditions. With 44 to 50 responses per question, differences of 10 to 20 percentage points may or may not be statistically meaningful. As presented, the table cannot carry the quantitative weight of the coherence and motion-preservation claims; the authors should add statistical analysis or reframe the study as a small pilot preference report.","section":"Table 2"},{"comment":"The generalization claim (trained on one frame of Bunny, then applied to unseen Smoke Jet data) is demonstrated only qualitatively in Figure 10. It is also unclear in what sense Smoke Jet is 'unseen' given that Table 2 reports experiments on the same dataset name; if a different Smoke Jet sequence was used, that must be stated. Generalization should be quantified on held-out frames and datasets, for example by reporting the appearance and motion losses on the unseen data, so that the reader can distinguish true generalization from within-dataset interpolation.","section":"Section 4.4"}],"minor_comments":[{"comment":"The notation for update steps is inconsistent: Section 3.2 says N steps are mapped to adjacent density frames, while Section 3.4 and Eqs. 12-13 use a random n training steps and then refer to N without defining its value. These variables should be defined together and used consistently.","section":"Sections 3.2 and 3.4"},{"comment":"Key hyperparameters are missing: the transmittance absorption constant γ, the loss weights λ_app and λ_motion, the number of VNCA update steps N and n, the exact MLP architecture, the number of training epochs, and the volume resolution at which the Table 2 runtime was measured. These are needed to reproduce the experiments and to interpret the ablation in Figure 9.","section":"Section 4 and Appendix"},{"comment":"Equation 3 writes the prior as zijk ∥ Pijk ∥ Dijk, but the text states that the perception vector also encodes velocity; the velocity encoding V is not visible in the equation. The authors should clarify where and how V is concatenated.","section":"Section 3.3"},{"comment":"There are several typographical and formatting errors: 'V olumetric' appears at the start of the title, the author email address 'alaa.abboud' does not match the author list, 'omni-view' and 'omniview' are used inconsistently, and the (i)/(ii) labels in the Figure 6 caption are not explained in the caption text.","section":"Throughout"},{"comment":"The Smoke Jet, Bunny, and Billowy datasets are not cited or described with generation parameters, which makes it difficult for readers to reproduce the experiments or to assess how challenging the test cases are.","section":"Datasets"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is creative and the method is well described, but the evidence is consistently one step behind the claims. The 'real-time' statement is contradicted by the paper's own runtime table, the training speedup is unsupported by measurements, and the 3D temporal-coherence claim needs a direct test. These are reporting and measurement gaps rather than obvious fatal flaws, so the authors should be given a revision opportunity. The journal should also decide whether 'real-time' is a hard requirement; if it is, the 0.8 s/frame result is a blocking issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThe core idea here is genuinely new: train a 3D neural cellular automaton on a single density frame and let its emergent motion replace explicit advection during volumetric smoke stylization. That is worth engaging with. The paper does several things well. The method description is clear, the ablations on density encoding, velocity encoding, and multi-view supervision are real and the numbers in Table 3 support those design choices. The qualitative results look plausible, and the mesh-texturing extension is a nice bonus. The authors also honestly note the limitation that the homogeneous update rule limits heterogeneity.\n\nWhere the paper falls short is the gap between claims and measurements. “Real-time” is contradicted by their own Table 2: 0.8 seconds per frame is not real-time by any reasonable definition, even if it beats the baselines’ 10–463 seconds. The “order of magnitude” training speedup is asserted, but no training-time measurements appear anywhere. The user study is a preference poll with no significance tests or confidence intervals. No code is released, which limits reproducibility. None of these are fatal to the method, but they are load-bearing claims that need substantiation.\n\nThe stress-test note is on target. The motion supervision in Eqs. 12–14 matches 2D optical flow of rendered views to a 2D projection of the input velocity field. With only a few camera views, this under-constrains the 3D motion of the texture volume. Also, each VNCA update is a local 3×3×3 stencil plus an MLP write to the center cell, so a pattern can move at most one voxel per update. To track smoke moving at, say, 10 voxels per frame, you need at least N=10 updates per frame. The paper never reports N or the velocity magnitudes of the data, and no direct 3D motion metric is reported. The multi-view ablation in Fig. 6 only tests static style consistency, not motion coherence, and the generalization results are qualitative only.\n\nNone of this makes the paper unserious. The mechanism is plausible, the ablation design is above average, and the authors are not hiding the missing pieces—they just have not filled them in. A serious referee could push for training-time measurements, a precise definition of real-time, a properly powered user study, and at least one quantitative 3D motion evaluation (e.g., synthetic ground-truth advection or voxel correspondence tracking). Given the novelty and the potential utility for VFX, I would send this to peer review, expecting major revision. I would not yet cite it as evidence for spatiotemporal coherence, but it is a legitimate new application of NCA to volumetric styling.\n\nIn short: merits a serious referee, not ready as is.","headline":"Promising 3D NCA for smoke stylization, but the real-time, order-of-magnitude speedup, and true 3D motion alignment claims are not backed by the reported evidence.","tokens_in":13509,"tokens_out":4198,"would_cite":false,"duration_ms":34989,"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":"A volumetric neural cellular automaton trained on a single density frame stylizes entire 3D smoke sequences in real time with multi-view and temporal coherence, cutting training time by more than an order of magnitude.","keywords":["neural cellular automata","volumetric style transfer","smoke stylization","dynamic texture synthesis","temporal coherence","multi-view consistency","flow-guided motion supervision","solid texture synthesis"],"falsifier":"Train VNCA on a smoke sequence with strong depth-wise motion, then stylize held-out frames and render them from side and top views. If distinctive texture features visibly slide relative to the underlying smoke density, or if the paper's motion-loss terms computed from novel views grow well beyond their training-view values, then the 2D flow supervision does not pin down the 3D texture motion and the spatiotemporal-consistency claim fails.","tokens_in":12422,"feed_emoji":"💨","tokens_out":20979,"duration_ms":151350,"temperature":0.7,"pith_summary":"This paper claims that a Volumetric Neural Cellular Automaton (VNCA) — a small recurrent network that updates a 3D voxel grid with one shared local rule — can stylize an entire 3D smoke simulation from a single reference image in real time. It is trained on just one density frame of the simulation, yet the same trained model stylizes every other frame of that sequence and even smoke datasets it never saw. The argument for why this works is that the automaton's self-emerging motion substitutes for explicit fluid-advection modeling and for the inter-frame smoothing used in previous volumetric style-transfer methods, and the paper reports that this substitution cuts training time by over an order of magnitude. If the claim is right, artists could get multi-view-consistent, temporally coherent stylized smoke at interactive rates using a style image as the only input, with training cost low enough for a single GPU.","feed_headline":"Train on one frame, stylize the whole 3D smoke sequence in real time","feed_subtitle":"A single small network replaces costly advection modeling and cuts training time by more than an order of magnitude.","key_machinery":"The central object is the Volumetric Neural Cellular Automaton (VNCA), a 3D voxel grid in which every cell carries a state vector (first three channels: RGB color; fourth: a density residual $\\Delta d$; the rest hidden), updated by a shared two-layer MLP $f_\\theta$ whose input is a perception vector built from the cell state, its 3D Sobel gradients, and a 27-point discrete Laplacian, concatenated with positional, density, and velocity encodings; each cell updates asynchronously through a randomized binary mask. The mechanism that carries the argument replaces explicit advection: the emergent motion of the trained automaton stands in for fluid transport, so no inter-frame smoothing loss or explicit transport solver is needed. That motion is steered by the flow-guided motion loss (Eqs. 12-14), which compares the optical flow between two rendered views of the evolving texture with the 2D projection of the input velocity field from the same camera pose, aligning direction first and then magnitude normalized by the number of update steps.","core_discovery":"The paper's central claim is that volumetric style transfer for smoke can be reframed as dynamic volumetric texture synthesis: rather than optimizing or regressing stylized density fields frame by frame, one learns a single local update rule for a 3D neural cellular automaton whose repeated application grows a texture volume that matches the reference style image and moves with the smoke. The cell state carries RGB color plus a density residual, and the shared update rule reads neighboring cell states through fixed 3D Sobel and Laplacian filters, conditioned on positional, density, and velocity encodings, with stochastic asynchronous updates. Training uses a differentiable volume renderer: rendered views of the evolving texture are matched to the style image with a VGG-based appearance loss, and a flow-guided motion loss asks the optical flow between two rendered views to align in direction and magnitude with the 2D projection of the simulation's velocity field. The key asserted consequence is that the automaton's self-emerging motion replaces explicit fluid advection and inter-frame smoothing, so a model trained on a single density frame generalizes to the entire sequence and to unseen smoke datasets while cutting training time by over an order of magnitude.","pith_inferences":["Beyond the paper: the motion loss only sees a 2D projection of the motion, so smoke moving mostly toward or away from the camera is the natural stress test; rendering stylized frames from side or top views and tracking whether texture features stay locked to the smoke would directly probe whether the 2D supervision pins down the 3D alignment.","Beyond the paper: the paper's stated limitation that the homogeneous update rule cannot produce clear foreground/background structure implies that styles with strong figure-ground organization will not stylize well, and a natural extension would be spatially conditioning the update rule on a segmentation or depth cue.","Beyond the paper: because the motion loss aligns direction before magnitude (Eq. 14), the automaton's texture speed matches the fluid speed only on average; that is fine for artistic output, but the synthesized motion should not be treated as physical advection if the output were reused in a simulation loop.","Beyond the paper: the demonstrated generalization to unseen datasets suggests a single VNCA could act as a reusable style capsule across simulations, and conditioning the appearance loss on the style image itself might let one model accept arbitrary reference images at inference without retraining."],"forward_implications":["A converged VNCA stylizes the full smoke sequence without retraining, so artists get a single trained model that handles every frame of a simulation in real time (about 0.8 s per frame in the paper's measurements, versus 10 s and 463 s for the compared baselines).","Because training uses a single density frame and a rotating camera, the resulting stylization is consistent across views and across frames by construction, eliminating the separate temporal-smoothing pass that slows optimization-based methods.","The order-of-magnitude training speedup makes volumetric style transfer practical on a single GPU, which the paper argues is what previously blocked wider use of such tools.","The same texture-volume formulation extends to mesh stylization, giving solid-texture-style results on meshes without UV maps.","VNCA generalizes to smoke datasets unseen in training, so a model trained on one simulation can be applied to new simulations with the same style reference."],"supporting_citations":[{"why":"Self-organising textures, the original neural-cellular-automaton texture synthesis work showing self-emerging, temporally varying motion from a static exemplar; this is the premise that NCA motion can carry stylization.","marker":"[19]"},{"why":"Cellular automata as convolutional neural networks, establishing that CA update rules can be parameterized by neural nets; the basis of the NCA and hence VNCA update rule.","marker":"[7]"},{"why":"DynCA, the controllable 2D dynamic-texture NCA that VNCA extends to 3D; supplies the update-rule machinery, positional encoding, and the idea of supervising NCA motion with optical flow.","marker":"[21]"},{"why":"Neural radiance fields, whose differentiable volume-rendering scheme lets VNCA render cell states into colored images for the appearance and motion losses.","marker":"[17]"},{"why":"The deep VGG network used as the pretrained feature extractor defining the style-matching and moment-matching appearance losses.","marker":"[25]"},{"why":"The relaxed optimal transport style loss whose formulation the appearance loss is built from.","marker":"[14]"},{"why":"The two-stream optical-flow network that measures the rendered texture's motion for flow-guided motion supervision.","marker":"[26]"},{"why":"Transport-based neural style transfer (TNST), the Eulerian optimization baseline whose explicit advection modeling and temporal smoothing VNCA replaces; the main comparison for stylization quality and runtime.","marker":"[12]"},{"why":"Lagrangian neural style transfer (LNST), the Lagrangian baseline VNCA is compared against for inter-frame consistency and runtime.","marker":"[13]"},{"why":"The stylizing kernel prediction network, a feed-forward volumetric stylization baseline whose tens-of-hours training cost motivates the order-of-magnitude speedup claim.","marker":"[8]"}],"fun_headline_variants":["Single-frame training stylizes entire 3D smoke sequence in real time","Neural cellular automata replace advection for real-time smoke style transfer","Volumetric NCA synthesizes stylized smoke textures in real time","Automaton's self-emerging motion replaces fluid advection for stylization","One update rule grows a stylized smoke volume and moves with it"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 2D projection of the smoke's velocity, seen from whichever camera angle is used in training, carries enough information for the automaton to learn full 3D motion alignment; if the out-of-plane components of that motion are under-constrained, the stylized texture could drift in depth even on frames where the front view looks correct.","fun_headline_variants_meta":{"raw":{"variants":["Single-frame training stylizes entire 3D smoke sequence in real time","Neural cellular automata replace advection for real-time smoke style transfer","Volumetric NCA synthesizes stylized smoke textures in real time","Automaton's self-emerging motion replaces fluid advection for stylization","One update rule grows a stylized smoke volume and moves with it"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001251,"raw_usage":{"total_tokens":5130,"prompt_tokens":951,"completion_tokens":4179,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":4083}},"tokens_in":567,"tokens_out":4179,"duration_ms":23838,"temperature":1.0,"reasoning_tokens":4083,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T05:16:49.936988+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train VNCA on a smoke sequence with strong depth-wise motion, then stylize held-out frames and render them from side and top views. If distinctive texture features visibly slide relative to the underlying smoke density, or if the paper's motion-loss terms computed from novel views grow well beyond their training-view values, then the 2D flow supervision does not pin down the 3D texture motion and the spatiotemporal-consistency claim fails.","supporting_citations":[{"cited_title":"Self-organising textures","cited_arxiv_id":null,"evidence_quote":"Self-organising textures, the original neural-cellular-automaton texture synthesis work showing self-emerging, temporally varying motion from a static exemplar; this is the premise that NCA motion can carry stylization."},{"cited_title":"Cellular automata as convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Cellular automata as convolutional neural networks, establishing that CA update rules can be parameterized by neural nets; the basis of the NCA and hence VNCA update rule."},{"cited_title":"Dynca: Real-time dynamic texture synthe- sis using neural cellular automata","cited_arxiv_id":null,"evidence_quote":"DynCA, the controllable 2D dynamic-texture NCA that VNCA extends to 3D; supplies the update-rule machinery, positional encoding, and the idea of supervising NCA motion with optical flow."},{"cited_title":"Srinivasan, Matthew Tancik, Jonathan T","cited_arxiv_id":null,"evidence_quote":"Neural radiance fields, whose differentiable volume-rendering scheme lets VNCA render cell states into colored images for the appearance and motion losses."},{"cited_title":"Style transfer by relaxed optimal transport and self-similarity","cited_arxiv_id":null,"evidence_quote":"The relaxed optimal transport style loss whose formulation the appearance loss is built from."},{"cited_title":"Two-stream convolutional networks for dynamic texture synthesis","cited_arxiv_id":null,"evidence_quote":"The two-stream optical-flow network that measures the rendered texture's motion for flow-guided motion supervision."},{"cited_title":"Azevedo, Markus Gross, and Barbara Solenthaler","cited_arxiv_id":null,"evidence_quote":"Transport-based neural style transfer (TNST), the Eulerian optimization baseline whose explicit advection modeling and temporal smoothing VNCA replaces; the main comparison for stylization quality and runtime."},{"cited_title":"Azevedo, Markus Gross, and Barbara Solenthaler","cited_arxiv_id":null,"evidence_quote":"Lagrangian neural style transfer (LNST), the Lagrangian baseline VNCA is compared against for inter-frame consistency and runtime."},{"cited_title":"V olumetric appearance styl- ization with stylizing kernel prediction network.ACM Trans- actions on Graphics (Proceedings of SIGGRAPH) , 40(4): 162–1, 2021","cited_arxiv_id":null,"evidence_quote":"The stylizing kernel prediction network, a feed-forward volumetric stylization baseline whose tens-of-hours training cost motivates the order-of-magnitude speedup claim."}],"review_version":1}