{"id":"0af1a990-8b1b-4844-9d25-a193632573a3","arxiv_id":"2511.22459","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A three-camera, flow-initialized 3D-Gaussian pipeline reconstructs temporally coherent 4D fire by separating static background from flame dynamics with monocular-depth regularization and hardware sub-frame synchronization.","lead":"This paper reconstructs 3D fire from just three commodity cameras by fusing multi-view optical flow into a 3D velocity field that initializes moving 3D Gaussian splats, with custom LED timing hardware for microsecond sync. It reports sharper flame-region renders than two 4D-Gaussian baselines and supplies a 17-scene real-fire dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Held-out evaluation and the circular depth metric do not rule out per-view 2D overfitting, so the central claim of genuine 3D reconstruction from three views is not yet supported.","rationale":"The reader's weakest_assumption focuses on the under-determined 3-view flow fusion. I believe the more load-bearing issue is upstream of that: the evaluation never demonstrates that the optimized model is a 3D reconstruction rather than a per-view appearance fit. The quantitative comparisons use held-out frames in time but the same three camera poses, so they measure temporal interpolation (which FreeTimeGS's overlapping lifetimes can do well) and do not test spatial generalization. The depth metric is the only quantitative 3D check, and it is circular for the static scene: DepthAnythingV2 is used both as a training regularizer (Eq. 3) and as the reference for RMSE_depth, so low error on background pixels is expected. For fire pixels, monocular depth is not a ground truth for an emissive, semi-transparent volume. The supplementary's admission that 4DGS(Wu) reduces to per-camera 2D video highlights the failure mode; the proposed method could similarly memorize per-view appearance with many Gaussians, and the paper's ablations do not rule it out—RandFlow still has 27.01 PSNR_flame, indicating that photometric fitting alone is quite powerful. Thus the central claim's '3D from three cameras' component is not yet supported by the evidence as written. A held-out-camera experiment directly tests 3D generalization and would settle the concern. This reinforces the reader's CONDITIONAL verdict: the condition should include such a test, not just (or primarily) the flow-fusion issue. Hence verdict UNCHANGED.","tokens_in":22259,"tokens_out":10075,"duration_ms":97334,"concrete_test":"Hold out one of the three cameras completely (e.g., train on left+right, never on center) and render the center camera's 100-frame sequence from the reconstructed Gaussians; compute PSNR_flame/SSIM_flame in the same fire mask used in Tab. 1. If the held-out-view metrics are within ~2-3 dB of the three-view results, true 3D structure is present. If they drop by >5 dB or show ghosting, the three-view results are dominated by per-view overfitting, and the central claim fails. This test uses the existing 17-scene dataset and requires no new capture.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the method reconstructs genuine 3D spatiotemporal flame dynamics from only three views. The reported evidence does not establish this. PSNR_flame/SSIM_flame are computed on temporally held-out frames rendered from the same three training viewpoints; a per-view 2D representation that interpolates appearance in time can score well on such a test. The supplementary even notes that one baseline degenerates to a per-camera 2D video, yet no experiment distinguishes the proposed method from a more graceful version of the same failure. The only quantitative 3D evidence is RMSE_depth against DepthAnythingV2 monocular depth, but that same network is used to regularize the static scene (Sec. 3.1, Eq. 3), so low background depth error is partly by construction; for the transparent/emissive fire region, monocular depth is not a reliable geometric reference. The flow-fusion weakness identified by the reader is real, but secondary: RandFlow still reaches 27.01 PSNR_flame versus 27.46, showing appearance can be fit without correct velocities. What needs testing is whether the representation actually generalizes across viewpoints, not just in time.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Gaussians on Fire, a pipeline for reconstructing 4D fire from three calibrated, hardware-synchronized consumer cameras. The static background is reconstructed with vanilla 3DGS initialized from stereo depth and regularized by aligned DepthAnythingV2 monocular depth; the dynamic fire is represented with FreeTimeGS-style Gaussians carrying linear velocity, lifetime, and temporal center, initialized from a Tikhonov-regularized fusion of per-view dense optical flow into a voxel grid. The authors introduce a real dataset of 17 fire scenes at 400 Hz, a custom ESP32/LED synchronization rig, and a rolling-shutter model. They report PSNR_flame 27.46 ± 1.76 versus 18.24/25.51 for the two 4DGS baselines, RMSE_depth 0.0412 versus ~0.27, and ablations for sync, rolling shutter, flow initialization, and static pre-training.","tokens_in":22435,"tokens_out":4202,"duration_ms":38010,"significance":"If the central claim were firmly established, this would be a useful contribution: it would show that high-frequency dynamic fire can be reconstructed from a minimal three-camera rig and that explicit per-Gaussian velocity and lifetime parameters are a practical representation for emissive, transparent scenes. Strengths include a real captured dataset, a concrete hardware synchronization design with measured temporal precision, a rolling-shutter correction model, and ablations that isolate the main components. The code and data are promised for release, which would help reproducibility. However, the evidence as presented does not yet establish genuine 3D generalization across viewpoints, and the only quantitative geometric metric is partially circular.","major_comments":[{"comment":"All quantitative metrics are computed on temporally held-out frames rendered from the same three training viewpoints. The supplementary acknowledges that 4DGS (Wu et al.) degenerates to a separate per-camera 2D video representation, and PSNR_flame/SSIM_flame would still reward such a 2D model on a temporal test. The central claim is '3D reconstruction from three views', so the evaluation must include a held-out viewpoint: for example, train on two views and test on the third, or render a calibrated novel camera path and compare against a held-out view quantitatively. The qualitative novel-view video is a useful start but does not separate true view interpolation from appearance memory.","section":"Sec. 5, Tables 1–2, Supp. Sec. 8"},{"comment":"RMSE_depth is computed against DepthAnythingV2 predictions after linear alignment, while the static reconstruction is trained with λ_depth increased by a factor of 100 against those same aligned predictions. The large gap between 0.0412 and ~0.27 therefore partly measures agreement with the training regularizer, not geometric accuracy. For the transparent/emissive fire region, monocular depth is an unreliable geometric reference. Please report depth against a geometrically grounded reference not used in training, or at least evaluate with the depth regularization switched off and clearly state the change in the comparison.","section":"Sec. 3.1 Eq. (3), Sec. 5 Table 1"},{"comment":"The fused 3D flow is underdetermined: with m=3 cameras, each voxel provides only three scalar projection constraints on a 3D vector, and closure relies on Tikhonov regularization with hand-set α0. The RandFlow ablation retains PSNR_flame 27.01 versus 27.46, showing that the photometric loss largely compensates for incorrect velocity initialization. Since the paper claims reconstruction of flame dynamics, please provide a direct evaluation of the velocity field—e.g., predicted 3D trajectories projected into a held-out view and compared with tracked flame features, or comparison with a physically motivated vertical-motion prior. The paper's own limitation statement in Sec. 6 concedes that optical-flow errors propagate into the 3D motion field, so an independent validation of the flow component is needed.","section":"Sec. 3.2 Eqs. (5)–(8), Table 2 (RandFlow)"}],"minor_comments":[{"comment":"Typographical issues: 'RSMEdepth' should be 'RMSE_depth', 'archives' should be 'achieves', and 'high offerfitting' should be 'high overfitting'.","section":"Sec. 5"},{"comment":"The notation π^{-1}_i(..., x) is nonstandard, and the expression f_i(π(x)) + π_i(x) mixes 2D and 3D operands. Please rewrite with explicit coordinates or provide a reference for this back-projection operation.","section":"Eq. (5)"},{"comment":"The measured readout time is stated as 2.85 μs per line in Sec. 3.3 and t_row ≈ 3 μs in Sec. 3.4; clarify whether these are the same quantity or different measurements.","section":"Secs. 3.3–3.4"},{"comment":"References [91] and [92] appear to be the same paper (Wu et al., CVPR 2024) listed twice. Please deduplicate.","section":"References"},{"comment":"The description 'increased the depth loss by a factor of 100, yielding an exponentially decreasing contribution with an initial value of 100 and a final value of 1' is ambiguous about the scheduling. Specify the exact annealing schedule.","section":"Supp. Sec. 10"}],"recommendation":"major_revision","confidential_remarks":"The main issue is not the soundness of the optimization but the absence of a quantitative cross-view generalization test. The depth metric is partially circular, and the flow-initialization ablation shows that appearance metrics alone cannot validate the claimed 3D dynamics. These are fixable within the manuscript's scope with added experiments and metric changes, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a serious, well-engineered paper with a valuable dataset and a sensible pipeline, but the headline claim—real 3D fire reconstruction from three views—is not yet proven by the reported experiments. The evaluation renders held-out temporal frames from the same three training viewpoints; that can be matched by a per-view 2D video model. The stress-test note is right: nothing here rules out a graceful version of the per-camera 2D degeneration the authors themselves identify in one baseline.\n\nWhat is actually new: the 17-scene, 400 fps real-fire dataset captured with three GoPro Hero 13 cameras and an ESP32/LED synchronization rig. That is a real contribution and, as far as I know, the first public dataset of its kind. The pipeline combines known pieces sensibly—minimum-intensity projection for static/dynamic decoupling, stereo-plus-monocular depth initialization, flow-fused FreeTimeGS—and the ablations show the expected dependencies: removing sync hurts, joint optimization destroys depth consistency, random flow reduces quality. The rolling-shutter analysis is thoughtful, and the authors honestly report its quantitative impact is small.\n\nThe soft spots are real but fixable. First, the depth metric is circular: DepthAnythingV2 regularizes the static scene (lambda_depth=100) and also serves as the reference after linear alignment. That explains the large RMSE_depth gap. Second, there is no cross-view generalization experiment. A held-out camera or a viewpoint far from the training cameras would distinguish a genuine 3D model from per-view 2D interpolation. The supplementary video apparently shows novel-view trajectories, but that qualitative claim isn't quantified and we didn't inspect it. Third, the RandFlow ablation (27.01 vs 27.46 PSNR_flame) shows appearance can be fit without correct velocities, so the flow field's geometric correctness isn't established. Fourth, baseline comparisons are not controlled: DyNeRF flame-salmon hyperparameters, a Gaussian cap for 4DGS (Yang), and no FreeTimeGS baseline. The fire-region mask is not described, which matters for masked metrics.\n\nWho it's for: researchers working on fire/smoke capture, sparse-view dynamic reconstruction, and anyone who can use a high-speed real-fire multi-view dataset. The dataset and the capture rig alone justify a serious referee. I would send it to review, with the request that the authors add a held-out-view test and replace or supplement the circular depth metric.","headline":"Valuable dataset and capture rig; the central 3D claim needs a held-out-view test before it is proven.","tokens_in":23102,"tokens_out":2790,"would_cite":true,"duration_ms":26804,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A pipeline reconstructs high-frequency 3D flame dynamics from just three calibrated, hardware-synchronized cameras, using Gaussian splats that each carry a lifetime and a linear velocity.","keywords":["dynamic scene reconstruction","3D Gaussian splatting","fire reconstruction","optical flow fusion","sparse multi-view video","temporal Gaussian parameterization","rolling shutter synchronization","novel view synthesis"],"falsifier":"Run the pipeline on a synthetic fire sequence with known ground-truth 3D velocity, using only three views; measure per-Gaussian velocity error against ground truth while keeping PSNR high. If the recovered trajectories diverge substantially from ground truth even though the images match, the central claim of reconstructing flame dynamics is falsified — the flow-initialized Gaussians would be fitting appearance, not motion.","tokens_in":21996,"feed_emoji":"🔥","tokens_out":7612,"duration_ms":68896,"temperature":0.7,"pith_summary":"The paper sets out to prove that high-frequency flame dynamics can be reconstructed as photorealistic 4D models from only three calibrated, hardware-synchronized consumer cameras, an under-constrained setting where previous dynamic Gaussian methods collapse into per-camera 2D reconstructions. Its strategy is to split the problem: reconstruct the static background with depth-regularized 3D Gaussian splats, then model the fire as a cloud of Gaussians whose positions are seeded by a 3D flow field fused from per-view 2D optical flow, and whose temporal behavior is explicitly parameterized by a lifetime and a linear velocity per splat. On 17 real fire sequences, the method reports fire-region PSNR of 27.46 and SSIM of 0.847, with a depth error of 0.0412, against 18.24 and 25.51 PSNR for two existing dynamic-Gaussian baselines. If correct, it would make high-fidelity fire capture practical with cheap hardware and would expose the flame motion field directly for analysis rather than hiding it inside a learned scene representation.","feed_headline":"Three cameras now rebuild fire as 4D velocity-carrying Gaussians","feed_subtitle":"Sub-frame LED sync and optical-flow fusion let three cameras beat dense-array baselines on fire","key_machinery":"The load-bearing machinery has three coupled parts. First, the static/dynamic split: the background is reconstructed once with vanilla 3D Gaussian splats whose geometry is initialized by dense stereo depth and regularized by aligned monocular depth, so the under-constrained three-view problem never touches the flames. Second, the flow-fusion initialization: per-view dense optical flow is back-projected into a voxel grid, and at each voxel the unknown 3D velocity F must satisfy (F - u_i) . u_i = 0 for each camera's back-projected 2D flow u_i; because three constraints do not determine a 3D vector, a Tikhonov term with a voxel-dependent weight selects a velocity, and the resulting volume seeds","core_discovery":"The discovery is that a dynamic 3D Gaussian representation can reconstruct fire from extremely sparse views, provided the Gaussians' temporal parameters are initialized from observed motion rather than learned from scratch. The key move is to treat each flame Gaussian as a particle: it has a birth time, a lifetime, and a constant velocity, so its world-space trajectory is x(t) = x0 + (t - t_mu)v. This parameterization is seeded by fusing dense 2D optical flow from all three cameras onto a voxel grid; at each voxel the back-projected 2D flows constrain the unknown 3D velocity to satisfy a projection constraint for each camera, and a regularization term selects the smallest admissible velocity","pith_inferences":["The same split-and-seed recipe should transfer to other semi-transparent volumetric phenomena such as smoke, steam, dust, or spray; the paper only evaluates combustion by-products, not whether the velocity-carrying Gaussian form is the general mechanism.","The explicit velocity field could be validated against combustion physics (buoyant rise profiles, turbulent spectra) as a way to separate reconstructed motion from photometric artifacts; the paper does not do this comparison.","A testable scaling prediction follows: if the flow fusion is the active ingredient, reconstruction quality in fire regions should degrade smoothly as cameras are removed from three to two, and the regularization bias toward small velocities should become visible in the recovered trajectories; the paper reports only the three-view case."],"forward_implications":["Fire capture moves from dense camera arrays or physics simulators to three synchronized consumer cameras, lowering hardware cost and setup complexity.","The reconstructed model can be rendered from novel viewpoints outside the training rig and at arbitrary times, because each Gaussian's position and opacity are defined for all times by its velocity and lifetime.","Flame motion becomes an explicit, queryable quantity: every Gaussian's velocity is a number, not a learned latent, so trajectories can be exported for analysis, animation, or reactive systems.","Sparse-view dynamic Gaussian reconstruction is substantially more stable when velocities are initialized from observed optical flow rather than randomly, per the paper's RandFlow ablation.","Separating static background from dynamic fire during training yields not only sharper fire rendering but also geometrically plausible depth, whereas joint optimization collapses depth accuracy."],"fun_headline_variants":["Fire in 4D from 3 cameras via velocity-carrying Gaussians","3 views suffice: flow-seeded Gaussians reconstruct dynamic fire","Particle-like Gaussians rebuild fire from 3 cameras","Sparse views no barrier: flow-initiated Gaussians rebuild fire"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole method stands or falls on the 3D flow field stitched together from three 2D motion estimates: with only three cameras, each voxel is under-constrained, and the paper's own limitation section (Sec. 6) concedes that optical-flow errors in bright or low-texture flame regions propagate directly into the 3D motion field and into every Gaussian seeded from it.","fun_headline_variants_meta":{"raw":{"variants":["Fire in 4D from 3 cameras via velocity-carrying Gaussians","3 views suffice: flow-seeded Gaussians reconstruct dynamic fire","Particle-like Gaussians rebuild fire from 3 cameras","Sparse views no barrier: flow-initiated Gaussians rebuild fire"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000919,"raw_usage":{"total_tokens":3761,"prompt_tokens":709,"completion_tokens":3052,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":2986}},"tokens_in":453,"tokens_out":3052,"duration_ms":19273,"temperature":1.0,"reasoning_tokens":2986,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:47:26.945589+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a synthetic fire sequence with known ground-truth 3D velocity, using only three views; measure per-Gaussian velocity error against ground truth while keeping PSNR high. If the recovered trajectories diverge substantially from ground truth even though the images match, the central claim of reconstructing flame dynamics is falsified — the flow-initialized Gaussians would be fitting appearance, not motion.","supporting_citations":[],"review_version":1}