{"id":"ba3a4aeb-ab58-4d3c-a2a6-f13f1678832f","arxiv_id":"2507.00476","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A frequency-rectified loss on spherical harmonic coefficients modestly improves neural BRDF reconstruction and editing on MERL, but reproducibility and evaluation issues weaken the claim.","lead":"FreNBRDF adds a spherical-harmonics-based frequency loss to a neural BRDF autoencoder and reports higher PSNR and SSIM than two baselines on the MERL material dataset. It is a short workshop paper with mixed RMSE results, an evaluation metric that duplicates its training loss, and no release of code or hyperparameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The frequency-rectified loss is built on an invalid spherical-harmonic transform: Eq. 3 fixes the wrong azimuth for isotropic BRDFs and leaves a quarter-sphere domain, so Lfre measures an arbitrary extension rather than material frequency content.","rationale":"The reader correctly identified the frequency-extraction step as the weakest link, but the concern is more fundamental than unreported hyperparameters. The paper's Eq. 3 and the surrounding text contain an internal inconsistency: Sec. 3.1 says isotropic BRDFs are independent of phi_D, while Sec. 4.1 says the MERL data are indexed by (theta_H, theta_D, phi_D). Standard Rusinkiewicz parameterization, which the paper cites, says isotropic BRDFs are independent of phi_H, not phi_D. If the literal notation is wrong, the frequency loss is computed on a slice that does not correspond to the material's actual directional variation. Additionally, even after fixing one azimuth, the remaining two angles for MERL occupy only a quarter of S2; no extension is specified, so the integrals in Eq. 9 are mathematically undefined. This means the claimed 'frequency analysis' is not a well-defined operation on BRDFs, and the reported Lfre improvement is not trustworthy evidence because Lfre is the loss being minimized. The independent metrics show only modest gains, and RMSE in Table 1 is worse than the naive NBRDF baseline. These issues are addressable in a revision: the authors could correct the parameterization, define an explicit domain extension or justify a restricted transform, and provide an ablation with eta=0. Until then, the central claim is not established. I therefore retain the reader's CONDITIONAL verdict; the condition should be fixing the SH transform, not merely reporting k, sigma, and L_max.","tokens_in":7007,"tokens_out":8604,"duration_ms":106133,"concrete_test":"Independently re-derive Eqs. 3-9 using the Rusinkiewicz parameterization and run two checks. (1) Verify whether isotropic BRDFs are independent of phi_H or phi_D; if the paper's notation is wrong, recompute cl,m with the correct azimuth fixed and report whether the frequency loss changes. (2) Implement the SH transform on the actual MERL [0, pi/2]^2 domain using two explicit extensions to S2, e.g., zero-padding and reflection, and compare Lfre and reconstruction metrics; if results change materially, the claimed frequency content is an artifact of the arbitrary extension. (3) Train the same autoencoder with eta = 0 and eta > 0, reporting PSNR, SSIM, and RMSE on the test set; if independent metrics do not improve beyond noise, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the assertion that the spherical-harmonic coefficients cl,m in Eqs. 8-9 capture the frequency content of a neural BRDF. That construction has two concrete defects. First, Sec. 3.1 states that isotropic BRDFs are independent of phi_D, but the standard Rusinkiewicz parameterization, and the paper's own Sec. 4.1 which indexes the MERL array by (theta_H, theta_D, phi_D), says isotropic BRDFs are independent of phi_H; phi_D is exactly the azimuth that carries directional reflectance variation. If Eq. 3 literally fixes phi_H = alpha, the slice still contains phi_D, so the SH coefficients do not encode what the paper claims; if the intended coordinate was phi_D, Eq. 3 discards the one azimuth that matters. Second, the remaining coordinates theta_H and theta_D only cover [0, pi/2]^2 for MERL, not S2. Eqs. 8-9 integrate over the full sphere without specifying any extension of the BRDF slice, so the coefficients cl,m are not well-defined. The dramatic Lfre drop in Tables 1-2 (6.8e-3 to 0.23e-3) is expected because Lfre is the training objective, not independent evidence. Missing k, sigma, and L_max are secondary: even with perfect hyperparameters, the transform as written is not a well-defined frequency decomposition.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FreNBRDF, a modification of the NBRDF autoencoder for measured BRDFs, adding a 'frequency-rectified' loss that penalizes differences between spherical harmonic coefficients of BRDF slices. The authors claim this improves reconstruction and editing quality over two baselines, and report quantitative results on the MERL dataset. The core technical contribution is an SH-based frequency analysis of BRDFs, obtained by fixing one angular coordinate and interpolating the tabulated MERL data onto a sphere.","tokens_in":7404,"tokens_out":4512,"duration_ms":48525,"significance":"If the frequency decomposition were well-defined, a principled frequency-domain regularizer for neural BRDFs could be a useful contribution to material modeling. The paper is clearly written and the idea of linking neural material representations to spherical harmonics is appealing. However, the central construction is mathematically defective (the isotropic co-ordinate statement is inverted and the SH transform is applied to a function that is not defined on the sphere), and the one independent rendering metric (RMSE in Table 1) contradicts the accuracy claim. The evaluation relies on the training loss as a success metric and omits all hyperparameter values. As a result, the paper's main claim is not supported.","major_comments":[{"comment":"The isotropy statement is inverted. In the Rusinkiewicz parameterization, an isotropic BRDF is independent of the half-vector azimuth φ_H, not of φ_D. This is confirmed by the paper's own Sec. 4.1, which describes the MERL array as being indexed by (θ_H, θ_D, φ_D). Fixing φ_H = α in Eq. (3) therefore leaves φ_D in the slice, and the SH coefficients do not encode the frequency content of an isotropic material as claimed. If the authors intended to fix φ_D instead, then the transform discards the one azimuthal coordinate that carries directional variation for isotropic materials; either way the construction is not a valid frequency analysis of the material.","section":"Sec. 3.1, Eq. (3)"},{"comment":"The function f_r(θ, φ) defined in Eq. (3) is not a well-defined function on the unit sphere S². The variables θ_H and θ_D each only range over [0, π/2] for the MERL dataset, so a slice in these variables covers at most a quarter of the sphere. The paper gives no extension of the BRDF to the full sphere, yet Eqs. (8)–(9) integrate over the full sphere. Consequently the spherical harmonic coefficients c_{l,m} are mathematically undefined, and the frequency-rectified loss L_fre in Eq. (10) is not a meaningful quantity. This is a load-bearing defect because the central claim rests entirely on these coefficients.","section":"Sec. 3.2, Eqs. (8)–(9)"},{"comment":"The reported RMSE for FreNBRDF in Table 1 (6.74×10⁻²) is worse than the naive NBRDF baseline (6.60×10⁻²), which directly contradicts the abstract's claim of improved accuracy. The single sentence after Table 1 that the higher RMSE is 'likely due to' the loss being designed for both spatial and frequency consistency is not a testable explanation. No error bars or statistical significance tests are provided anywhere in the paper. Furthermore, L_fre is the very loss minimized during training, so its drop from 6.80×10⁻³ to 0.23×10⁻³ in Table 1 is expected and cannot be used as independent evidence of frequency fidelity; the claim of improved frequency compliance is therefore circular.","section":"Tables 1 and 2, Sec. 4.4"},{"comment":"The method's key hyperparameters are never reported: the neighbour count k and Gaussian standard deviation σ in Eq. (4) and Eq. (6), the spherical harmonic truncation degree L_max in Eq. (8), the frequency-loss weight η in Eq. (11), and the regularization coefficients λ_1 and λ_2 in Eq. (1). The paper states that 'optimal hyperparameters are selected based on performance on the validation set' but gives no values. Without these, the experiments are not reproducible, and it is impossible to assess whether the reported results depend on an accidental choice of these parameters, especially σ and k, which determine the interpolated function that is supposedly transformed to the frequency domain.","section":"Sec. 4.2, Eqs. (4), (6), (8), (11)"}],"minor_comments":[{"comment":"The heading 'Recitification' contains a typo; it should be 'Rectification'.","section":"Sec. 3.2 title"},{"comment":"The phrase 'accessing the frequency compliance' should read 'assessing the frequency compliance'.","section":"Sec. 4.3"},{"comment":"The notation for the azimuthal coordinate is inconsistent: φ and ϕ are both used for what appears to be the same variable. Please standardize to a single symbol.","section":"Eqs. (3), (8)–(9)"},{"comment":"The captions state 'with mean (solid) and variance (dashed line)' but do not specify whether the mean and variance are taken over the test set, over samples, or over rendering viewpoints. This should be clarified.","section":"Figs. 4 and 5"},{"comment":"The phrase 'state-of-the-art baselines' is overstated: the comparison is only against one external method [6] and the authors' own naive pipeline. Please temper the 'state-of-the-art' claim or add comparisons to recent neural BRDF and frequency-aware methods.","section":"Sec. 4.4, Table 1"}],"recommendation":"reject","confidential_remarks":"The central technical construction appears to be unsound: the isotropy co-ordinate error in Sec. 3.1 and the undefined SH domain in Secs. 3.2–3.3 are not presentation issues but invalidate the proposed frequency loss. Fixing these would require redefining the method and re-running all experiments. The RMSE result in Table 1 already contradicts the paper's main claim, and the lack of hyperparameter reporting further weakens the contribution. I recommend rejection for the workshop version; a resubmission would need a corrected mathematical formulation, a valid frequency-domain metric, and a more thorough comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Chenliang and colleagues have a clear idea: take FrePolad's frequency-rectified loss and apply it to neural BRDFs. That is a genuine extension, and the experimental setup—reconstruction plus editing via latent interpolation—is a sensible way to test it. The paper is also well organized, and the qualitative results look plausible at a glance.\n\nBut the central construction does not stand up. The authors claim isotropic BRDFs are independent of phi_D and then fix phi_H. It's the opposite: in the Rusinkiewicz parameterization, isotropic materials are independent of phi_H, and the 90x90x180 MERL array the paper uses includes phi_D for a reason. So Eq. 3 fixes the wrong azimuth. Worse, the remaining coordinates theta_H and theta_D do not form a standard spherical coordinate pair; they are two polar angles on different reference axes. The SH integrals in Eqs. 8–9 over S2 are therefore not defined for the function you get. You can't just set up a 2D function on a sphere from two angles that don't parameterize a sphere. This means L_fre measures something arbitrary, not the material's frequency content. The missing values of k, sigma, and L_max are minor in comparison; no choice of those fixes the ill-posed transform.\n\nOn the experimental side, the PSNR and SSIM gains in Tables 1 and 2 are worth taking seriously, but the RMSE regression in Table 1 and the absence of error bars keep those gains from being decisive. The L_fre column is trained against itself, so a low number is expected and confirms nothing.\n\nThe right move is to send this to a reviewer who knows BRDF parameterization well. The flaw is specific and fixable—if the authors can actually define a proper frequency representation (e.g., using spherical harmonics on a correctly parameterized sphere or a manifold-aware transform), the paper could be a useful short contribution. As written, it does not support its claims. I wouldn't cite it until the math is repaired.","headline":"The frequency-rectified loss is built on a misparameterized spherical-harmonic transform; the idea is a genuine extension of FrePolad, but the central claim does not hold as written.","tokens_in":7884,"tokens_out":3464,"would_cite":false,"duration_ms":38796,"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":"FreNBRDF claims that adding a spherical-harmonic frequency error term to the training loss improves the accuracy and robustness of neural BRDF reconstruction and editing on the MERL dataset.","keywords":["neural BRDF representation","frequency-rectified loss","spherical harmonics","material reconstruction","material editing","MERL dataset","autoencoder latent space","implicit neural representation"],"falsifier":"Take a synthetic material whose true reflection function is known analytically, compute its exact spherical-harmonic coefficients, then recompute them from discretely sampled measurements using the paper's interpolation procedure. If the interpolated coefficients differ substantially from the exact ones at low frequencies, then the frequency loss is minimizing error against a distorted target rather than the material's real spectrum, and the claimed mechanism fails.","tokens_in":6843,"feed_emoji":"🎨","tokens_out":11898,"duration_ms":116771,"temperature":0.7,"pith_summary":"The paper tries to establish that neural representations of surface reflectance can be made more accurate by training them with an explicit frequency-domain error, not only a spatial-sample error. It introduces FreNBRDF, an autoencoder pipeline that takes a material's reflectance function, slices it at a fixed azimuth angle, expands each slice in spherical harmonics, and adds the mean squared difference of the resulting coefficients to the reconstruction loss. The claim is that this frequency-rectified loss helps the network learn a more structured distribution of neural material weights, which shows up as better reconstruction fidelity and, especially, more stable interpolation between materials during editing. The payoff, if true, is that implicit neural materials become more reliable for downstream appearance tasks while keeping the compact latent representation.","feed_headline":"Frequency loss sharpens neural material reconstruction and editing","feed_subtitle":"FreNBRDF adds spherical-harmonic error to BRDF training, beating baselines on MERL materials and stabilizing interpolations.","key_machinery":"The central mechanism is the frequency-rectified loss $L_{\\mathrm{fre}}$. To define it, the paper maps a BRDF to $S^2$ by fixing the azimuth argument $\\varphi_H$ to a constant $\\alpha$, interpolates the discrete measurements using normalized Gaussian weights over the $k$ nearest neighbors, and expands the result in spherical harmonics up to a truncation order $L_{\\max}$. The squared difference of the resulting coefficients $c_{l,m}$ between reconstructed and target materials is then added to the $L_1$ reconstruction loss and weight regularizers, so the autoencoder is trained to keep neural material weights consistent in both spatial and frequency domains. The mechanism carries the argument because the claimed improvements all flow from this extra term; its behavior depends on the choices of interpolation bandwidth $\\sigma$, neighbor count $k$, and truncation order $L_{\\max}$.","core_discovery":"On its own terms, the paper's central claim is that FreNBRDF, a neural material representation trained with a frequency-rectified loss, reconstructs and edits measured materials more faithfully than the hypernetwork set-encoder baseline and the naive NBRDF pipeline. The construction treats a BRDF as a function on the sphere by fixing $\\varphi_H = \\alpha$, interpolating the discrete MERL samples with normalized Gaussian $k$-nearest-neighbor weights, and computing spherical-harmonic coefficients $c_{l,m}$ per RGB channel. The frequency-rectified loss $L_{\\mathrm{fre}}$ is the mean squared error between these coefficients for reconstructed and ground-truth BRDFs, added to the spatial reconstruction loss with a weighting hyperparameter $\\eta$. With this term, the paper reports consistently lower frequency-compliance loss and better PSNR and SSIM on both reconstruction and editing, and interprets this as evidence that frequency rectification gives the latent space a more structured and interpretable organization.","pith_inferences":["The reconstruction table shows RMSE slightly worse than the naive NBRDF baseline while PSNR and SSIM improve; a plausible reading is that the frequency term trades a small amount of pixel-level intensity accuracy for structural and perceptual fidelity, a trade-off the paper does not quantify.","Because the $\\varphi_H$-fixed slice drops azimuthal dependence, the current gains on isotropic MERL data may partly reflect the simplification rather than the frequency loss itself; testing on anisotropic materials would separate these factors.","The paper reports no ablation of the weighting hyperparameter $\\eta$, so the causal role of $L_{\\mathrm{fre}}$ is inferred from comparisons against the no-frequency baseline; sweeping $\\eta$ would show whether the benefit is monotonic or saturates.","Being differentiable in the network output, the frequency loss could be dropped into other neural appearance models, such as neural scene representations, as a regularizer rather than a training objective for a dedicated material autoencoder."],"forward_implications":["If the central claim holds, FreNBRDF reconstructs unseen MERL materials with better PSNR and SSIM than the two baselines while cutting the frequency-compliance loss from about $6.80 \\times 10^{-3}$ to $0.23 \\times 10^{-3}$.","Latent-space interpolation becomes a dependable editing operation: interpolation metrics stay near reconstruction quality instead of degrading as they do for the baselines.","Because the autoencoder uses a permutation-invariant set encoder, the frequency-rectified loss applies to arbitrary-size and sparse sample sets, preserving the pipeline's generalizability.","The same frequency-rectification principle is positioned to extend to anisotropic, spatially varying, and physically-based materials, the paper's stated next targets."],"supporting_citations":[{"why":"Supplies the MERL dataset of 100 isotropic measured BRDFs used for training, reconstruction, and editing evaluation.","marker":"[3]"},{"why":"Defines NBRDF, the neural field whose weight distribution the autoencoder learns.","marker":"[4]"},{"why":"Presents the hypernetwork set-encoder baseline that FreNBRDF extends and compares against.","marker":"[6]"},{"why":"Provides the frequency-rectification idea, adapted here from point-cloud generation to BRDFs.","marker":"[7]"},{"why":"Supplies the log-relative mapping used to preprocess BRDF values before reconstruction.","marker":"[18]"},{"why":"Underlies the spherical-harmonic transform used to extract frequency coefficients for the loss.","marker":"[19]"},{"why":"Defines the half/difference vector parameterization on which the $\\varphi_H$-fixed slice construction depends.","marker":"[20]"},{"why":"Provides the Deep Sets architecture that makes the encoder permutation-invariant and size-flexible.","marker":"[21]"}],"fun_headline_variants":["Rectified frequency loss boosts neural BRDF fidelity","Spherical harmonics sharpen neural material editing","Frequency-aware loss enhances material reconstruction","Neural BRDF gains frequency discipline","Frequency rectification tightens material modeling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the frequency content of the interpolated material slices, formed by fixing one angle and averaging nearby measured samples, faithfully represents the material's true high-frequency detail, so that reducing the frequency error improves appearance instead of merely matching an artifact of the interpolation.","fun_headline_variants_meta":{"raw":{"variants":["Rectified frequency loss boosts neural BRDF fidelity","Spherical harmonics sharpen neural material editing","Frequency-aware loss enhances material reconstruction","Neural BRDF gains frequency discipline","Frequency rectification tightens material modeling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000273,"raw_usage":{"total_tokens":1615,"prompt_tokens":902,"completion_tokens":713,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":651}},"tokens_in":518,"tokens_out":713,"duration_ms":7750,"temperature":1.0,"reasoning_tokens":651,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:13:48.289201+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic material whose true reflection function is known analytically, compute its exact spherical-harmonic coefficients, then recompute them from discretely sampled measurements using the paper's interpolation procedure. If the interpolated coefficients differ substantially from the exact ones at low frequencies, then the frequency loss is minimizing error against a distorted target rather than the material's real spectrum, and the claimed mechanism fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MERL dataset of 100 isotropic measured BRDFs used for training, reconstruction, and editing evaluation."},{"cited_title":"Dataset The MERL dataset [3] is used in this study for its diversity and data-driven nature, making it suitable for both statistical and neural-network-based methods","cited_arxiv_id":null,"evidence_quote":"Defines NBRDF, the neural field whose weight distribution the autoencoder learns."},{"cited_title":"Neumadiff: Neural mate- rial synthesis via hyperdiffusion,","cited_arxiv_id":null,"evidence_quote":"Presents the hypernetwork set-encoder baseline that FreNBRDF extends and compares against."},{"cited_title":"Tech- niques for brdf evaluation,","cited_arxiv_id":null,"evidence_quote":"Provides the frequency-rectification idea, adapted here from point-cloud generation to BRDFs."},{"cited_title":"Non-linear approximation of re- flectance functions,","cited_arxiv_id":null,"evidence_quote":"Supplies the log-relative mapping used to preprocess BRDF values before reconstruction."},{"cited_title":"Measuring and modeling anisotropic re- flection,","cited_arxiv_id":null,"evidence_quote":"Underlies the spherical-harmonic transform used to extract frequency coefficients for the loss."},{"cited_title":"A reflectance model for com- puter graphics,","cited_arxiv_id":null,"evidence_quote":"Defines the half/difference vector parameterization on which the $\\varphi_H$-fixed slice construction depends."},{"cited_title":"Microfacet models for refraction through rough surfaces,","cited_arxiv_id":null,"evidence_quote":"Provides the Deep Sets architecture that makes the encoder permutation-invariant and size-flexible."}],"review_version":1}