{"id":"2c6adebe-bf20-4742-9d84-2382bba1cdf4","arxiv_id":"2508.06632","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A neural rendering method that decomposes appearance into static bases and dynamic coefficients, improving the sharpness of specular highlights in novel view synthesis.","lead":"CoDe-NeRF splits a scene's appearance into a fixed set of learned material-like basis patterns and a set of view- and light-dependent coefficients, then blends them with a small neural network to render novel views. The approach targets shiny and reflective surfaces, where standard NeRF-style models tend to produce blurry highlights.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Forward-Facing PSNR gain of ~6 dB over all baselines is an unexplained outlier; the central SOTA claim rests on this table, so the empirical claim is not established.","rationale":"The paper's central claim is explicitly empirical: 'state-of-the-art performance in rendering glossy and reflective scenes, significantly outperforming prior methods in both qualitative and quantitative comparisons.' For that claim to hold, the quantitative results in Table 1 must be trustworthy and relevant. The NeRF-Synthetic numbers are plausible (34.18 vs 33.99 for Ref-NeRF), though the SSIM inconsistency (Ref-NeRF lower than NeRF) is a warning. The Forward-Facing number is not plausible: 33.33 dB PSNR is ~6 dB above the best published baseline on LLFF, a dataset where methods cluster at 26–28 dB. The paper provides no per-scene breakdown, no error bars, no code, and no explanation for this gain. A 6 dB PSNR increase corresponds to a ~60% reduction in MSE; such an improvement would be a major breakthrough and would be visible in full-image qualitative comparisons, but Figure 2 only shows small crops. The most likely explanation is an evaluation mismatch (e.g., using the same views for training and testing, or a nonstandard split), which would invalidate the central quantitative claim. The reader's weakest_assumption about the expressiveness of 16 bases is a plausible secondary concern, but it is not needed to reject the paper; the empirical foundation is already the weak point. If the Forward-Facing result is an artifact, the SOTA claim fails regardless of whether the architecture is expressive. The concrete test—regenerating Table 1 using the standard LLFF split and per-scene PSNR—would settle this. If the number replicates, the paper would be much stronger; if not, the claim is unsupported. Hence the REJECT verdict stands.","tokens_in":10808,"tokens_out":4144,"duration_ms":40397,"concrete_test":"Reproduce Table 1 on Forward-Facing using the standard LLFF protocol: 1-in-8 test view split, downsampled 1/8 resolution, evaluate PSNR per scene and report the average. Check that test views are excluded from training. If the average drops below ~28 dB, the reported 33.33 dB is an artifact and the SOTA claim collapses. Additionally, report per-scene Shiny numbers (e.g., CD scene) to test the glossy-scene claim directly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is empirical: Table 1 reports 33.33 dB PSNR on Forward-Facing (LLFF), 6.07 dB above the best baseline (Nex, 27.26) and 6.83 dB above NeRF (26.50). LLFF is a forward-facing capture benchmark with limited view-dependent effects; published methods cluster at 26–28 dB, and a 6 dB improvement would require roughly a 60% reduction in MSE. No per-scene breakdown, error bars, or code is provided, and Figure 2 shows only selected crops, so the gain cannot be checked. The number is so far outside the expected range that it suggests a different evaluation protocol (e.g., testing on training views or a different split). Since the paper's central claim—state-of-the-art glossy rendering—depends on this quantitative comparison, the claim is not currently supported. The NeRF-Synthetic result (34.18 vs Ref-NeRF 33.99) is a modest 0.19 dB gain and within run-to-run variation; the Forward-Facing result carries the argument, and it is the least secure. Additionally, the Shiny dataset, the most direct test for glossy/reflective scenes, is only shown qualitatively; no quantitative table is provided for it.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoDe-NeRF, a neural radiance field variant that decomposes appearance into a view-invariant neural basis H (projected from TensoRF-style appearance features) and view/illumination-dependent coefficients k produced by a FiLM-modulated MLP (the \"Coefficient Network\"). A second MLP (the \"Dynamic Radiance Integrator\") nonlinearly blends k and H to produce radiance. The method is trained with photometric and TV losses, with optional per-condition illumination embeddings. The authors claim state-of-the-art rendering of specular and reflective scenes, reporting PSNR 34.18 on NeRF-Synthetic and 33.33 on Forward-Facing (Table 1), plus ablations and dense-view experiments.","tokens_in":11263,"tokens_out":4601,"duration_ms":54227,"significance":"If the reported results are reproducible, the decomposition idea is a simple and potentially useful alternative to physically-based inverse-rendering approaches, and the ablations provide a reasonable sanity check of the design. The paper is also honest about limitations: no relighting, no editable materials, and latent illumination cannot model cast shadows. However, the central empirical claim is not currently supported: the Forward-Facing PSNR is an unexplained 6 dB outlier relative to all baselines, no per-scene numbers or error bars are provided, the Shiny benchmark—the most direct test for glossy/reflective scenes—is only shown qualitatively, and there are internal inconsistencies between tables. The method's \"physics derivation\" is motivational rather than a constraint, which weakens the framing but does not by itself invalidate the architecture.","major_comments":[{"comment":"The Forward-Facing PSNR of 33.33 dB is 6.07 dB above the best baseline (Nex, 27.26) and 6.83 dB above NeRF (26.50). This is far outside the range reported by existing work on this benchmark and implies a ~60% reduction in MSE. No per-scene breakdown, error bars, or evaluation-protocol details are given, and Figure 2 shows only selected crops. Since the paper's central 'state-of-the-art' claim rests largely on this number, the claim is not established. Please provide per-scene results, the exact train/test split, and ideally code or a reproducibility statement.","section":"Table 1"},{"comment":"The Shiny dataset is the most direct benchmark for the paper's stated goal of rendering glossy and reflective scenes, yet it is only shown qualitatively. Quantitative PSNR/SSIM/LPIPS comparisons on Shiny against NeX, Ref-NeRF, and other baselines are needed to support the abstract's claim of 'several challenging benchmarks' and the SOTA claim.","section":"Figure 3 / Shiny dataset"},{"comment":"The ablation study in Table 3 reports 'Ours (Full Model)' with PSNR 32.54 and LPIPS 0.041, while Table 1 reports the same method on NeRF-Synthetic with PSNR 34.18 and LPIPS 0.037. The text does not specify which scene or subset the ablations use, making it impossible to judge the magnitude of each component's contribution. This needs to be stated clearly; otherwise the ablations are uninterpretable.","section":"Table 3 vs Table 1"},{"comment":"Table 2 reports results on the 'materials' scene of NeRF-Synthetic trained with 100, 200, and 300 views. The standard NeRF-Synthetic dataset provides 100 training views per scene. Please explain how 200- and 300-view training sets were constructed (e.g., rendered from the NeRF model, subsampled differently, or obtained from another source) and confirm no test-view contamination. This is necessary to validate the claim that the method 'scales effectively with view count.'","section":"Table 2 / Experimental protocol"},{"comment":"Equation (10) states that the learnable illumination embedding z_s approximates the lighting integral, but z_s is a per-condition vector with no dependence on surface point x or incoming direction ω_i, so it cannot represent the spatially varying integral in Eq. (5). If this is intended only as motivation, the text should say so. The phrase 'constrained by BRDF-inspired priors' in the Coefficient Network section is not backed by any explicit constraint in the loss or architecture.","section":"Eqs. (10)-(11)"}],"minor_comments":[{"comment":"Typo: 'F ortress' and 'Horn' should be 'Fortress' and 'Horns'.","section":"Figure 2 caption"},{"comment":"Grammar: 'Tables 1 present' should be 'Table 1 presents'.","section":"Page 1/Table 1"},{"comment":"Please define Δ²V and Δ²M. The denominator P as 'total number of parameters' is ambiguous given the sum is over parameters; clarify indexing.","section":"Eq. (16)"},{"comment":"H is derived from per-point appearance features T_c(x), so calling H 'static' is misleading. It is view-invariant, not spatially constant; the terminology should be clarified.","section":"Eq. (7)"},{"comment":"The text says training runs for 100k iterations, then states 'training a NeRF-Synthetic scene for 10k iterations takes approximately 2 hours.' Please reconcile these numbers or clarify that one is a typo.","section":"Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The paper has a plausible architecture and honest limitations, but the central empirical claim is not yet supported. The 6 dB Forward-Facing outlier and the missing quantitative Shiny results are serious; if the numbers are reproducible, a major revision with per-scene tables and code can resolve this, but if the baseline comparison protocol is incorrect, the SOTA claim would collapse. The author should also be asked to clarify the dense-view protocol, since the original NeRF-Synthetic benchmark only has 100 training views."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: CoDe-NeRF is a reasonable incremental architecture—static neural basis plus FiLM-modulated coefficients plus an MLP integrator—and the ablations tell a coherent story. The problem is the headline numbers. The 33.33 dB PSNR on Forward-Facing is about 6 dB above Nex, the best baseline. Published methods sit at 26–28 dB on that benchmark; a 6 dB jump means roughly a 60% reduction in MSE. That is not impossible, but nothing in the paper explains it. There is no per-scene breakdown, no error bars, no code. The Shiny benchmark, which is the most direct test for glossy reflections, appears only as qualitative crops. So the central claim—state-of-the-art glossy rendering—rests on exactly the evidence that is missing.\n\nWhat the paper does well: the decomposition is cleanly motivated, the FiLM-style coefficient network is a sensible way to condition on view direction, and the ablations isolate the contribution of each component. Removing the decomposition costs over 3 dB; replacing the nonlinear integrator with a dot product costs 1.5 dB; removing FiLM costs about 0.8 dB. Those numbers are internally consistent and make sense. The paper also honestly states its limitations: no cast shadows, no material editing.\n\nThe soft spots are real. Besides the LLFF outlier, Table 3 and Table 4 report the full model at 32.54 dB PSNR, while Table 1 reports 34.18 dB on NeRF-Synthetic. The tables don't say which dataset or scene the ablations use, so the reader cannot reconcile them. The 'physics derivation' in Eqs. (5)–(11) is motivational; no BRDF constraint is enforced, and the illumination embedding z_s is just a latent code. That is fine, but it should not be presented as a physics-based approximation. And there are no error bars anywhere, which matters when the claimed gain over Ref-NeRF on NeRF-Synthetic is 0.19 dB—well within run-to-run noise.\n\nWho this is for: someone working on view-dependent appearance modeling might find the coefficient-basis decomposition worth trying, especially the FiLM conditioning. But they should not trust the quantitative comparison until the authors supply per-scene numbers, especially for LLFF and Shiny, and ideally code.\n\nIf I were the editor, I would not desk-reject it; the architecture and ablations deserve referee time. But I would send it back with a clear request: redo the evaluation, provide per-scene results and error bars, and either explain the LLFF number or drop the SOTA claim. As it stands, the empirical case is not made.","headline":"Plausible architecture, informative ablations, but the Forward-Facing PSNR is a 6 dB outlier with no supporting detail—so the SOTA claim is not established.","tokens_in":11671,"tokens_out":3073,"would_cite":false,"duration_ms":30032,"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":"CoDe-NeRF decomposes scene appearance into a static neural basis and view-dependent coefficients, claiming sharper specular highlights than prior radiance-field methods on glossy and reflective benchmarks.","keywords":["neural radiance fields","novel view synthesis","specular reflection","dynamic coefficient decomposition","FiLM conditioning","neural basis","volume rendering","view-dependent appearance"],"falsifier":"Render a glossy scene illuminated by a small, close light that casts a sharp shadow; if CoDe-NeRF reproduces the highlight but the shadow boundary is blurry or absent while an explicit-light method renders it cleanly, the latent-illumination assumption is falsified. Equivalently, on a scene with many distinct materials, check whether PSNR keeps rising when the basis count grows beyond 16; if it plateaus or drops, the fixed basis capacity caps the method.","tokens_in":10736,"feed_emoji":"✨","tokens_out":5495,"duration_ms":57409,"temperature":0.7,"pith_summary":"The paper tries to establish that the hard problem of rendering shiny, reflective scenes from photographs can be handled without choosing between blurry entangled appearance models and unstable physics-based inverse rendering. Its proposal is to split appearance into two learned pieces: a small set of shared neural basis vectors that carry material properties, and dynamic coefficients, generated from view direction and an illumination code, that modulate those bases. A non-linear MLP then combines coefficients and basis into final radiance. If the claims hold, the payoff is sharper specular highlights and reflection details at higher image quality than NeRF, TensoRF, 3D Gaussian Splatting, and Ref-NeRF, while keeping the simple optimisation of a feed-forward radiance field. The reported results include 34.18 dB PSNR on NeRF-Synthetic and 33.33 dB on Forward-Facing.","feed_headline":"Coefficient decomposition sharpens specular rendering to 34.18 dB","feed_subtitle":"Static basis plus view-dependent coefficients renders sharper highlights than prior radiance fields—no BRDF optimization.","key_machinery":"The engine is the dynamic coefficient decomposition: appearance features $T_c$ are projected to a compact shared basis $H = W \\cdot T_c$; a Coefficient Network $F_\\theta$ applies a feature-wise linear modulation (FiLM) $k = W_x \\odot d + b_x$ using the view direction and an illumination embedding $z_s$; and a Dynamic Radiance Integrator $G_\\phi$, a small MLP, non-linearly blends $k$ and $H$ into outgoing radiance. This replaces the expensive integral over incoming light with a learned, low-rank, non-linear map, and the paper's ablations attribute most of the gain to this three-part division.","core_discovery":"The central discovery is that view-dependent appearance does not need to be baked into a monolithic color field, nor does it require solving a full BRDF inverse-rendering problem. The paper shows that decomposing color into a view-agnostic neural basis $H=\\{h_n\\}_{n=1}^{N_p}$ and view/light-dependent coefficients $k$, generated by a FiLM-style affine transform $k = W_x \\odot d + b_x$ and fused by an MLP integrator $G_\\phi(k,H) \\approx \\int_\\Omega f_r(x,\\omega_i,d)L_i(x,\\omega_i)(\\omega_i\\cdot n)\\,d\\omega_i$, produces sharper specular highlights than prior methods on the NeRF-Synthetic, Forward-Facing, and Shiny benchmarks. The decomposition acts as a learned low-rank approximation of the ren","pith_inferences":["The latent illumination embedding is a global code per lighting condition, so an untested extension is to interpolate or extrapolate these codes to approximate novel lighting; the paper does not run that experiment.","The plateau at $N_p=16$ suggests capacity is tied to scene complexity; scenes with many distinct materials may need per-scene or adaptive basis counts, and the paper's fixed choice may under- or over-allocate.","Because $G_\\phi$ approximates the BRDF integral, the learned bases could be read as a dictionary of effective BRDF lobes; that interpretation suggests a testable route to material editing even though the current framework does not expose physical parameters.","The method's gains come without explicit surface normal estimation, so it may complement or compete with normal-aware specular methods; comparing on scenes where normals are known would isolate where the coefficient shortcut fails."],"forward_implications":["Glossy and mirror-like surfaces can be rendered at higher fidelity than prior radiance-field methods without explicit BRDF decomposition, since the coefficient-basis split captures specular effects with a single forward pass.","The non-linear blending step is load-bearing: swapping the integrator for a dot product drops LPIPS sharply, so linear basis expansion alone is not enough.","Dense multi-view training improves the method's quality consistently (from 32.61 dB at 100 views to 36.27 dB at 300 views on the materials scene), while a competing method degrades.","The neural basis acts as a reusable dictionary of reflection patterns, so appearance is stored more compactly than feeding raw appearance features directly.","The design supports multiple lighting conditions by appending a learnable illumination embedding per condition, at the cost of local effects like cast shadows."],"supporting_citations":[{"why":"Supplies the NeRF baseline, the volume rendering equation, positional encodings, and the NeRF-Synthetic benchmark used for evaluation.","marker":"(Mildenhall et al. 2020)"},{"why":"Supplies TensoRF's vector-matrix tensor factorization, which provides the density and appearance feature backbone and the TV regularization; TensoRF is a primary baseline.","marker":"(Chen et al. 2022)"},{"why":"Provides Ref-NeRF, the strongest specular-focused baseline the method compares against on view-dependent appearance.","marker":"(Verbin et al. 2022)"},{"why":"Provides 3D Gaussian Splatting, an explicit representation baseline that entangles lighting and material, used as a key comparison.","marker":"(Kerbl et al. 2023)"},{"why":"Provides Nex, a neural basis expansion baseline, and the Shiny dataset used for qualitative evaluation on reflective objects.","marker":"(Wizadwongsa et al. 2021)"},{"why":"Provides the FiLM affine conditioning mechanism that the Coefficient Network adapts for view-direction modulation.","marker":"(Perez et al. 2018)"},{"why":"Provides the LLFF baseline and the Forward-Facing dataset used for evaluation.","marker":"(Mildenhall et al. 2019)"}],"fun_headline_variants":["Coefficient decomposition sharpens specular highlights in NeRF","Dynamic coefficient decomposition renders sharper highlights","View-dependent coefficients beat prior radiance fields","Decomposition sharpens speculars without BRDF optimization","Static basis plus dynamic coefficients sharpen NeRF speculars"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The method assumes that a fixed set of 16 shared neural basis vectors, modulated by a learned affine transform of the view direction and a global illumination code, is expressive enough to reproduce every view- and illumination-dependent specular effect in a scene.","fun_headline_variants_meta":{"raw":{"variants":["Coefficient decomposition sharpens specular highlights in NeRF","Dynamic coefficient decomposition renders sharper highlights","View-dependent coefficients beat prior radiance fields","Decomposition sharpens speculars without BRDF optimization","Static basis plus dynamic coefficients sharpen NeRF speculars"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001016,"raw_usage":{"total_tokens":4113,"prompt_tokens":718,"completion_tokens":3395,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":462,"completion_tokens_details":{"reasoning_tokens":3336}},"tokens_in":462,"tokens_out":3395,"duration_ms":26178,"temperature":1.0,"reasoning_tokens":3336,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:38:44.595552+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a glossy scene illuminated by a small, close light that casts a sharp shadow; if CoDe-NeRF reproduces the highlight but the shadow boundary is blurry or absent while an explicit-light method renders it cleanly, the latent-illumination assumption is falsified. Equivalently, on a scene with many distinct materials, check whether PSNR keeps rising when the basis count grows beyond 16; if it plateaus or drops, the fixed basis capacity caps the method.","supporting_citations":[{"cited_title":"P.; Tancik, M.; Barron, J","cited_arxiv_id":null,"evidence_quote":"Supplies the NeRF baseline, the volume rendering equation, positional encodings, and the NeRF-Synthetic benchmark used for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies TensoRF's vector-matrix tensor factorization, which provides the density and appearance feature backbone and the TV regularization; TensoRF is a primary baseline."},{"cited_title":"T.; and Srinivasan, P","cited_arxiv_id":null,"evidence_quote":"Provides Ref-NeRF, the strongest specular-focused baseline the method compares against on view-dependent appearance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides Nex, a neural basis expansion baseline, and the Shiny dataset used for qualitative evaluation on reflective objects."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the FiLM affine conditioning mechanism that the Coefficient Network adapts for view-direction modulation."},{"cited_title":"P.; Ortiz-Cayon, R.; Kalantari, N","cited_arxiv_id":null,"evidence_quote":"Provides the LLFF baseline and the Forward-Facing dataset used for evaluation."}],"review_version":1}