{"id":"42a51b00-9d73-427d-8626-a6fbf0b8cd2e","arxiv_id":"2501.05226","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A diffusion-prior-guided differentiable volume renderer (PDPS) reconstructs 3D clouds from a single image, using a new monoplanar latent representation and a synthetic cloud dataset.","lead":"This paper combines a diffusion model trained on synthetic clouds with a differentiable renderer that accounts for light scattering, to reconstruct a full 3D cloud volume from a single 2D image. The authors report higher-quality single-view reconstructions than optimization-only baselines, but the test clouds appear to come from the same data the diffusion model was trained on, which would make the comparison optimistic.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central evaluation lacks a documented train/test split: Section 4.3 trains the diffuser on all 1,000 Cloudy volumes while Table 2 evaluates on 16 'unseen clouds' from the same dataset, so the claimed generalization may be memorization.","rationale":"Good-faith reading: the method is a reasonable engineering contribution—an unconditional latent diffusion prior on monoplanar volume codes coupled to a differentiable path tracer, with code and data released. The theoretical scheme is not obviously unsound, and several auxiliary experiments (super-resolution, inpainting, transmittance-based recovery, light-condition recovery) are suggestive. My sole load-bearing concern is the evaluation's train/test integrity. The paper never states that the 16 'unseen clouds' are disjoint from the 1,000 used to train the diffuser; in fact Section 4.3 explicitly says all 1,000 are used for the latent training space. This makes the term 'unseen' ambiguous at best: it could mean 'not used to construct the test images' while still being in the diffusion prior's training set. If so, the prior has memorized the ground-truth volumes, and the strong N-LPIPS numbers in Table 2 are expected even for a system with no generalization. The DPS1 vs SPS1 T-LPIPS comparison (0.0205 vs 0.0123) already shows the paper's single-view advantage is not uniform; the decisive metric is N-LPIPS, and that is exactly the one most vulnerable to memorization. I agree with the reader's weakest_assumption. The concrete test above would settle the issue by checking overlap in the released code/data and, if needed, retraining on a clean split. If the split exists, the verdict should move toward conditional acceptance pending error bars and out-of-distribution tests; if it does not, the central claim is unsupported and rejection is appropriate.","tokens_in":20413,"tokens_out":4736,"duration_ms":42423,"concrete_test":"Download the public Cloudy dataset and training code from the provided GitHub repository, and determine whether any of the 16 clouds used for the 32 Table 2 test cases appear (possibly up to the 14x8 flips/transposes/rotations/scales) among the 1,000 volumes whose latents were optimized in Section 4.3. If overlap exists, retrain the diffusion model on the remaining 984 clouds and rerun the DPS/DRT/SPS comparison on the same 16 test clouds; if DPS3x3's N-LPIPS advantage over SPS3/DRT3 shrinks or reverses, the central generalization claim is unsupported. If no overlap can be verified, the reader's objection is resolved and the verdict should be reconsidered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is that PDPS gives single-view reconstruction of volumetric clouds at a previously unattainable quality. The load-bearing evidence is Table 2, where DPS beats DRT and SPS on novel-view LPIPS on 32 test cases built from 16 'unseen clouds.' But Section 4.3 states that the diffusion model's latent space is generated by considering all 1,000 clouds from the Cloudy dataset, plus 14x8 transformed instances, and no holdout split is described anywhere in the paper or supplement. If the 16 evaluation clouds are a subset of those 1,000 (or transform-equivalent to them), the unconditional prior has already seen the exact density fields being reconstructed. Then the diffusion guidance is not a generalizing prior; it is a retrieval mechanism, and the N-LPIPS improvements over DRT and SPS in Table 2 reflect training-data recall rather than reconstruction ability. This is not a stylistic shortcoming: every quantitative claim in Section 5.5 and the abstract's 'previously unattainable quality' depends on the test clouds being disjoint from training. The paper provides no protocol, no split metadata, and no out-of-distribution experiment to establish that.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for reconstructing 3D volumetric density fields (specifically cumulus clouds) from one or a few images by combining an unconditional latent diffusion model with a physically-based differentiable volume renderer. The authors introduce a new dataset (Cloudy) of 1,000 synthetic cloud volumes, a novel 'monoplanar' implicit representation that compresses each volume into a 128x128x32 latent grid, and a parametric diffusion posterior sampling (PDPS) scheme (Algorithms 1 and 2) that alternates DPS-style guidance with optimization of unknown physical parameters such as background radiance. The method is evaluated on single-view and three-view reconstruction against DRT and SPS (Table 2), on super-resolution, inpainting, transmittance-based reconstruction, and joint lighting recovery. The central quantitative claim is that PDPS yields novel-view LPIPS substantially better than the prior-free baselines, supporting the abstract's assertion of 'previously unattainable quality.'","tokens_in":20667,"tokens_out":7051,"duration_ms":65356,"significance":"If the experimental results are valid, this is a strong contribution: it demonstrates a practical way to inject a learned shape prior into a physically-accurate differentiable renderer, a known hard problem due to the ill-posedness of inverse volume rendering. The dataset release, the compact monoplanar representation, and the public code are valuable assets for the community. The paper is also transparent about important limitations, such as the ambiguity between the latent volume θ and the physical parameters ϕ, and the high computational cost. However, the significance hinges on the evaluation being conducted on clouds that are truly unseen by the diffusion prior; currently that is not established, and the headline claim of 'previously unattainable quality' is therefore not supported as written.","major_comments":[{"comment":"The diffusion model is trained on latent codes derived from all 1,000 Cloudy volumes (Section 4.3), while Table 2's caption states that the 32 test cases are constructed from 16 'unseen clouds' from the same dataset. The paper never states that these 16 clouds are excluded from the 1,000 used to generate the training latent codes, and no split protocol (e.g., cloud identifiers or an explicit 'held-out' statement) is provided anywhere in the main text or supplement. If the test clouds are a subset of the training set, the prior can encode the exact ground-truth volumes, and the reported N-LPIPS improvements over DRT and SPS would reflect memorization rather than reconstruction ability. This is load-bearing for every quantitative claim in Section 5.5 and for the abstract. Please specify the exact split, provide the cloud IDs or a reproducible protocol, and re-evaluate on clouds that are provably held out from the diffusion training set; alternatively, report the overlap statistics to demonstrate disjointness.","section":"Section 4.3 vs. Table 2"},{"comment":"The abstract claims single-view reconstruction at 'previously unattainable quality,' but the only quantitative comparison in Table 2 is against DRT and SPS, which are prior-free differentiable optimization baselines. There is no comparison with recent learning-based single-view reconstruction methods (e.g., pixelNeRF, Zero-1-to-3, or latent-diffusion NeRF pipelines), and no ablation that isolates the contribution of the diffusion prior from the differentiable renderer. As a result, the strength of the headline claim is not fully supported by the evidence presented. A comparison with at least one learning-based baseline, or a careful rephrasing of the claim to 'previously unattainable with optimization-only differentiable rendering,' would make the contribution statement proportionate.","section":"Section 5.5 and Abstract"}],"minor_comments":[{"comment":"The sentence 'Karras et al. [49] and Po et al. [49] provide thorough overviews' appears to cite the same reference twice; [49] is Po et al., and the Karras et al. entry is missing from the reference list.","section":"Section 2"},{"comment":"The metrics T-LPIPS and N-LPIPS are not defined in the main text or the table caption; please define them (e.g., test-view LPIPS and novel-view LPIPS) and state the exact computation procedure.","section":"Table 2"},{"comment":"The augmentation operations are described only as '14 such operations' and '8 equivariant transformations (flips and transposes)'; please enumerate the exact operations for reproducibility.","section":"Section 4.3"},{"comment":"The transmittance-based reconstruction is presented with qualitative figures only; adding a quantitative metric (e.g., PSNR or LPIPS on the reconstructed volume or rendered views) would strengthen the claim.","section":"Section 5.4"},{"comment":"The lighting recovery experiment states that the optimization 'closely matches' the ground-truth lighting, but no quantitative error is reported for the recovered background or environment map; please add a numerical measure.","section":"Section 5.6"},{"comment":"Several reference entries contain incorrect author attributions (e.g., 'Yong Jae Wang' appears in papers that should be attributed to different authors); please proofread the reference list against the original sources.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The train/test overlap issue is the central concern: if the 16 'unseen clouds' are actually part of the 1,000-cloud training set, the paper's main quantitative results are invalid. I recommend major revision rather than outright rejection because the methodological contribution (PDPS + monoplanar representation + public dataset/code) is potentially sound and verifiable, and the split can be documented or the experiments re-run on truly held-out clouds. The authors should also be asked to tone down the 'previously unattainable' claim unless they add a learning-based baseline or clearly qualify the comparison scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the real novelty is coupling latent diffusion posterior sampling with a physically-based differentiable volume renderer for participating media, plus the monoplanar latent representation. That combination is sensible and potentially useful. But the central quantitative evidence rests on an evaluation where the diffusion prior probably trained on the very clouds used for testing, so the reported gains are not yet evidence of generalization.\n\nWhat's new: the integration of DPS with a multi-scattering volume renderer, the Cloudy dataset of 1,000 simulated clouds, and the monoplanar compression scheme. The writing is clear, the method details are fairly complete, and the authors ship code and data, which is good practice. Table 1 suggests their monoplanar representation does beat triplanar and grid at equal parameter count in their setting.\n\nThe soft spot is not small. Section 4.3 says the diffuser's latent codes are generated from all 1,000 Cloudy volumes, with 14x8 transformed copies. Table 2 then evaluates on 32 test cases built from 16 'unseen clouds' from the same dataset, and no train/test split is described anywhere. If those 16 clouds are among the 1,000 (or transform-equivalent), the unconditional prior may have memorized their density fields. Then DPS is retrieving training data, not reconstructing unseen structure, and the N-LPIPS improvements over DRT and SPS in Table 2 are not meaningful. This is load-bearing: the abstract's 'previously unattainable quality' depends on it.\n\nThere are secondary issues: no error bars, no real-image test, no comparison against conditional generative baselines. The paper does acknowledge the ambiguity between density and lighting parameters in the conclusion. But the split problem is the one that matters most.\n\nWho this is for: graphics and inverse-rendering researchers working on volume reconstruction. The method and dataset could be useful, but the claims need a corrected evaluation before I'd rely on them. My recommendation: treat this as a paper worth reviewing, but the referee should demand a proper held-out split or an out-of-distribution test, such as clouds from a different simulator or real photographs. If the authors can show the same gains on genuinely unseen volumes, this becomes a solid contribution. As it stands, the central result is unproven.","headline":"The idea is good and the monoplanar representation is neat, but the main evaluation likely lets the diffusion prior see the test clouds, so the headline quality claim is unsupported.","tokens_in":21180,"tokens_out":2678,"would_cite":false,"duration_ms":25574,"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 diffusion prior guided by light-transport gradients reconstructs volumetric clouds from a single view at a quality earlier methods did not reach.","keywords":["single-view reconstruction","volumetric clouds","diffusion posterior sampling","latent diffusion model","differentiable volume rendering","monoplanar representation","multiple light scattering","inverse volume rendering"],"falsifier":"Check the released dataset and training scripts for overlap between the 16 test clouds and the training set; if any overlap exists, retrain the diffuser on only the non-test clouds and re-run the Table 2 comparison to see whether the N-LPIPS advantage over DRT and SPS survives.","tokens_in":20185,"feed_emoji":"☁️","tokens_out":9098,"duration_ms":84903,"temperature":0.7,"pith_summary":"This paper tries to establish that volumetric cloud reconstruction from a single photograph becomes tractable when a learned prior over 3D density is combined with gradients from a physically-based volume renderer. The authors introduce a monoplanar latent representation that compresses simulated cloud volumes, train an unconditional diffusion model on that latent space, and then run diffusion posterior sampling in which the renderer's gradients push the latent code toward agreement with the observed image while the prior keeps the volume cloud-like. They report that this scheme improves perceptual fidelity over prior-free differentiable rendering baselines using one or three views, runs in minutes rather than roughly half an hour, and can recover unknown background and environment lighting along with the density field.","feed_headline":"Light-aware diffusion sampling rebuilds 3D clouds from one image","feed_subtitle":"A learned cloud prior and light-transport gradients recover density and lighting from one image.","key_machinery":"The central object is a parametric diffusion posterior sampling loop (PDPS, Algorithm 2) running in the latent space of an unconditional diffusion model, with a monoplanar representation as the compression layer. Monoplanar means a single 2D feature grid on the xy-plane whose feature vectors are sampled along z and decoded by a shared MLP; it compresses each cloud volume from roughly 100 MB to a 2 MB latent code. The load-bearing mechanism is that gradients of the image loss are computed through a physically-based differentiable volume renderer, so the update direction $\\nabla_{\\theta_t}\\|y - R(D(\\hat\\theta_0), \\phi)\\|_2^2$ carries real light-transport information, while the diffusion prior keeps the density distribution cloud-like. The outer loop (Algorithm 1) alternates SGD on physical parameters $\\phi$ with posterior samples $\\theta$, starts denoising from a partially noisy latent, and scales the guidance weight $\\zeta$ from 0.1 to 1.","core_discovery":"On its own terms, the paper's central claim is that an unconditional diffusion prior over compressed cloud densities, coupled to a differentiable path-tracing volume renderer, turns single-view reconstruction of highly scattering volumes from a severely underdetermined optimization into a workable posterior sampling problem. The authors compress each roughly 100 MB density field into a roughly 2 MB monoplanar latent code, train a latent diffusion model on 1,000 simulated cumulus clouds augmented to 112,000 instances, and then alternate SGD on physical scene parameters with diffusion posterior samples in the latent space. In their Table 2 comparison, the single-view and three-view variants outperform the gradient-only baselines DRT and SPS on the perceptual metric N-LPIPS for novel views, produce cloud-like interiors, and the same loop recovers unknown background or environment lighting in Section 5.6.","pith_inferences":["The paper does not describe a train/test split between the 1,000 simulated clouds and the 16 'unseen' test clouds, so the headline comparison should be re-run after confirming that the tested clouds are outside the diffusion model's training set.","Because the prior is trained only on simulated cumulus clouds, its ability to generalize to smoke, fire, or non-cumulus morphology is an open question; a natural stress test is reconstruction from an independently simulated or analytic volume.","The $\\zeta$ schedule, with weak prior guidance early and stronger guidance later, is a tunable knob; adapting it to observation noise or view count could extend the method to noisier or more ambiguous inputs without changing the algorithm."],"forward_implications":["A single cloud photograph can yield a full 3D density field suitable for novel-view rendering and relighting, not just a 2D appearance match.","The same posterior-sampling machinery transfers to other observation models, such as transmittance images, coarse-grid super-resolution, and inpainting, because the forward model is swapped without retraining the diffuser.","Unknown scene parameters such as background radiance and environment lighting are recoverable alongside the density field, reducing the need to calibrate lighting before reconstruction.","Three-view input with three diffusion restarts (DPS3x3) gives the best reported quality, so the method scales gracefully as views are added.","Reconstructions take minutes instead of the roughly half-hour runs of the DRT and SPS baselines, making iterative volume reconstruction more practical."],"supporting_citations":[{"why":"Supplies the diffusion posterior sampling gradient-guidance step at the heart of the proposed method.","marker":"[9]"},{"why":"Provides the differentiable volume path-tracing method with singular path sampling and serves as the SPS baseline.","marker":"[27]"},{"why":"Provides the DRT baseline and differential tracker gradients used for volume rendering comparisons.","marker":"[46]"},{"why":"Contributes radiative backpropagation, the adjoint differentiation principle used to send gradients into the volume.","marker":"[45]"},{"why":"Path replay backpropagation gives the constant-memory gradient reuse the optimization relies on.","marker":"[67]"},{"why":"The fluid simulator used to generate the 1,000-cloud benchmark dataset.","marker":"[21]"},{"why":"LPIPS, the perceptual metric used for the quality comparison in Table 2.","marker":"[82]"},{"why":"Hard data consistency refinement used in the optional middle-phase optimization of Algorithm 1.","marker":"[57]"}],"fun_headline_variants":["Single-view clouds from diffusion and light transport","One-shot 3D clouds via light-aware diffusion sampling","Diffusion plus path tracing rebuilds 3D clouds from one photo","Cloud volumes from a single image: diffusion meets light physics","Light-aware latent diffusion reconstructs 3D clouds from one view"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported quality gains assume the 16 test clouds were never among the 1,000 clouds or their 111,000 augmented copies used to train the diffusion model, and the paper does not describe such a holdout split.","fun_headline_variants_meta":{"raw":{"variants":["Single-view clouds from diffusion and light transport","One-shot 3D clouds via light-aware diffusion sampling","Diffusion plus path tracing rebuilds 3D clouds from one photo","Cloud volumes from a single image: diffusion meets light physics","Light-aware latent diffusion reconstructs 3D clouds from one view"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000137,"raw_usage":{"total_tokens":1100,"prompt_tokens":848,"completion_tokens":252,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":169}},"tokens_in":464,"tokens_out":252,"duration_ms":2852,"temperature":1.0,"reasoning_tokens":169,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:13:39.761354+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check the released dataset and training scripts for overlap between the 16 test clouds and the training set; if any overlap exists, retrain the diffuser on only the non-test clouds and re-run the Table 2 comparison to see whether the N-LPIPS advantage over DRT and SPS survives.","supporting_citations":[{"cited_title":"Image-based re- construction of heterogeneous media in the presence of mul- tiple light-scattering","cited_arxiv_id":null,"evidence_quote":"Provides the differentiable volume path-tracing method with singular path sampling and serves as the SPS baseline."},{"cited_title":"Unbiased inverse volume rendering with differential trackers","cited_arxiv_id":null,"evidence_quote":"Provides the DRT baseline and differential tracker gradients used for volume rendering comparisons."},{"cited_title":"Radiative backpropagation: an adjoint method for lightning-fast differentiable rendering","cited_arxiv_id":null,"evidence_quote":"Contributes radiative backpropagation, the adjoint differentiation principle used to send gradients into the volume."},{"cited_title":"Path re- play backpropagation: differentiating light paths using con- stant memory and linear time","cited_arxiv_id":null,"evidence_quote":"Path replay backpropagation gives the constant-memory gradient reuse the optimization relies on."},{"cited_title":"Embergen: Real-time fluid simulation software,","cited_arxiv_id":null,"evidence_quote":"The fluid simulator used to generate the 1,000-cloud benchmark dataset."}],"review_version":1}