{"id":"5d2faf63-e9fa-4887-8a6c-0841a4cd604f","arxiv_id":"2501.09460","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Normal-NeRF uses transmittance gradients and dual density activations to reconstruct and render shiny scenes more accurately than prior NeRF-based methods.","lead":"Normal-NeRF introduces a transmittance-gradient method for estimating surface normals in Neural Radiance Fields, targeting the shape ambiguity that breaks reconstruction of highly reflective scenes. The authors report state-of-the-art rendering and normal accuracy on reflective synthetic benchmarks and competitive results on real captures.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central premise is unproven: for a ray-aligned ambiguous density field, Eq. (9) degenerates to the camera ray direction, so transmittance monotonicity does not guarantee a 3D normal.","rationale":"The paper is a serious empirical contribution: it has a public code link, five-run standard deviations in Table 6, and large PSNR gains on Shiny Blender and Glossy Synthetic. Those results are real evidence that the full system works, and I am not claiming fraud or that the method cannot be made to work. However, the reader's weakest assumption is the correct load-bearing point, and it can be sharpened. The derivation of the transmittance gradient proves at most that T is monotone along a ray; it never proves that translating the ray segment yields the local surface normal. A density artifact that varies only along the ray makes the transmittance gradient exactly ±d for every sample, which is the opposite of ambiguity-robustness for curved surfaces. Since the method's normal loss is the mechanism that makes the reflection-aware appearance model work, an unreliable normal estimate in precisely the ambiguous regime would undermine the central claim. The end-to-end experiments cannot rule this out because the density field is optimized jointly, and the per-scene normal MAE in Table 8 is inconsistent (toaster and ball regressions vs. Ref-NeuS and ENVIDR). I therefore endorse the reader's CONDITIONAL verdict: the paper should be accepted only with either a proof or a controlled experiment showing that Eq. (9) recovers normals in a deliberately ambiguous density field, or with the theoretical claim appropriately weakened.","tokens_in":15904,"tokens_out":10499,"duration_ms":119786,"concrete_test":"Use a frozen, deliberately ambiguous density field from a converged baseline (e.g., Ref-NeRF without orientation loss on Shiny Blender toaster) and compute Eq. (9) at all ray samples. Measure the mean angular error of nt_i versus ground-truth normals, and also measure the average |nt_i · d|. If transmittance-gradient normals are no better than density-gradient normals in this controlled setting, or if nt_i is strongly aligned with d in regions of high rendering weight, the ambiguity-robustness premise fails. For an analytic check, substitute σ(x)=f((x−o)·d) into Eqs. (7)-(9) and verify that nt_i = ±d; this shows the claimed robustness is not guaranteed by the construction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 claims that because T(t) is monotone along any ray, the transmittance gradient (Eqs. 7-9) is an ambiguity-robust surface normal. The argument only concerns the scalar T along a fixed 1D ray; monotonicity imposes no constraint on the 3D direction of ∇_x \\hat T. Worse, the exact ambiguity the paper targets admits a degenerate case: if a learned density artifact is ray-aligned, σ(x)=f((x−o)·d), then ∇σ(x)=f'(t)d and Eq. (9) yields nt_i = ±d for every sample on that ray. The estimated \"normal\" is then the camera viewing direction, not the surface normal, and the predicted-normal loss trains the normal MLP toward that view-dependent direction. T remains monotone for any nonnegative f, so the stated motivation does not prevent this. The ablations in Table 3 only report end-to-end metrics of the full pipeline, and Table 8 shows per-scene MAE is not consistently better (toaster: 8.329 vs ENVIDR 6.496, Ref-NeuS 5.560). The paper therefore does not establish that transmittance gradients are accurate normals under ambiguous shape predictions; it only shows the full system helps on the tested scenes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Normal-NeRF, a NeRF-based method for reconstructing and rendering highly reflective scenes. The core idea is to replace density-gradient normals with a transmittance-gradient normal estimate (Eqs. 7-9), which the authors argue is robust to the shape ambiguity caused by specular reflections. The method also introduces dual activated densities (softplus and exp) to reconcile smooth surface normals with sharp object boundaries, and a stop-gradient warmup for the predicted-normal loss. The system is evaluated on NeRF Synthetic, Shiny Blender, Glossy Synthetic, and real captured scenes, with per-scene tables, ablations, and released code. The paper reports large average improvements over prior work on the two reflective datasets, e.g., PSNR 39.24 vs. 35.96 for Ref-NeRF on Shiny Blender and 33.24 vs. 30.49 for Ref-NeuS on Glossy Synthetic.","tokens_in":16146,"tokens_out":6237,"duration_ms":65317,"significance":"The empirical contribution is substantial: the method achieves state-of-the-art rendering quality on two challenging specular datasets, includes detailed per-scene results (Tables 7-10), multi-seed standard deviations (Table 6), controlled ablations (Tables 3-5), and public code. If the transmittance-gradient normal estimate is accepted as a heuristic, the paper is a practical step forward for reflective-scene reconstruction. However, the theoretical justification for why transmittance gradients are robust under shape ambiguity is not proven, and the per-scene normal-error results do not uniformly support the robustness claim. The central idea is therefore defensible but needs both a more careful framing and additional analysis before the paper can be accepted.","major_comments":[{"comment":"The monotonicity of the one-dimensional transmittance T(t) along a fixed ray does not constrain the three-dimensional direction of the translation gradient ∇_x T(x;d,t), so the claim that transmittance gradients are 'ambiguity-robust' normals is not established by the argument in the text. For a ray-aligned density field of the form σ(x)=f((x−o)·d), Eq. (9) reduces to n_t^i=±d for every sample on that ray, regardless of the shape of f, while T(t) remains monotone for any nonnegative f. This is exactly the kind of degenerate, view-dependent normal estimate the method is supposed to avoid. The authors should either prove the claim under explicit assumptions about the density field (e.g., locally planar opaque surfaces) or re-frame the transmittance gradient as an empirical heuristic whose validation is purely experimental.","section":"Transmittance Gradient, Eqs. (7)-(9)"},{"comment":"Eq. (9) is introduced as a quadrature-based estimate of the transmittance gradient, but no derivation or error analysis is given. The sum over j<i of ∇σ(x_j)δ_j is a one-sided path integral whose accuracy for the translation gradient at x is unclear on curved, multi-layer, or deeply semi-transparent surfaces. This is load-bearing because the predicted-normal loss (Eq. 10) trains the normal MLP against this quantity, and the resulting normals determine the reflection direction (Eq. 12) used in the appearance model. Please provide a derivation, a bound, or an explicit statement that Eq. (9) is a numerical heuristic accepted on empirical grounds.","section":"Transmittance Gradient, Eq. (9)"},{"comment":"The per-scene MAE results in Table 8 do not consistently favor the proposed normal estimator. On toaster, Ours reports MAE 8.329 versus ENVIDR 6.496 and Ref-NeuS 5.560; on ball, Ours reports 0.529 versus ENVIDR 0.245. Since the paper's central claim is robustness under ambiguous shape prediction, the average MAE improvement alone is insufficient evidence. Please report statistical significance, discuss the scenes where the method fails, or restrict the robustness claim to the subset of scenes where it actually holds.","section":"Table 8"}],"minor_comments":[{"comment":"In the definition of T_i, the summation index is written as σ_i δ_i; this should be σ_j δ_j to match the standard quadrature and to be consistent with the surrounding text.","section":"Eq. (4)"},{"comment":"The arrow notation ← →L_n and − →L_n is used before being explained, and the decomposition in Eq. (10) is not self-explanatory. Please define the notation explicitly and clarify which term corresponds to which gradient direction.","section":"Eqs. (10)-(11)"},{"comment":"Table 3 is difficult to interpret because the rows show only checkmarks without explicit row labels; in particular, the 'Softplus' and 'Exp' columns do not uniquely indicate whether one activation is removed or both are active. Please add a clear configuration description for each row.","section":"Ablation Studies, Table 3"},{"comment":"The text repeatedly says 'batch size of 219 sample points' and 'hash table size of 219'. Given the Instant-NGP context, these likely mean 2^19; please use superscripts or the explicit form 2^19 to avoid ambiguity.","section":"Implementation Details"},{"comment":"The claim 'we are the first to identify and analyze the inherent limitations of the density gradient' is stronger than necessary, since prior work such as Ref-NeRF already discusses the unreliability of density-gradient normals on specular surfaces. Please soften the claim and cite the relevant discussion.","section":"Introduction"},{"comment":"The real-captured-scene comparison includes only Zip-NeRF and Ref-NeRF; for consistency with the other tables, please report the SDF-based baselines (ENVIDR, Ref-NeuS) on these scenes or state why they are omitted.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong empirical component and is likely salvageable, but the central theoretical claim about transmittance-gradient normals is currently overstated. I would recommend major revision rather than rejection because the method appears to work well in practice on the tested scenes; the authors need to either supply a rigorous justification for Eq. (9) or explicitly reposition it as a heuristic, and they need to address the per-scene normal-error discrepancies in Table 8. I would also encourage the editor to ask for the clarified ablation table and the batch-size typo fixes, as these affect reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What should you know about this paper? It introduces a transmittance-gradient normal estimator for NeRF that accumulates density gradients along a ray, plus a dual-activated-density scheme and a stop-gradient warmup for the predicted-normal loss. On Shiny Blender and Glossy Synthetic the reported gains are sizable (PSNR 39.24 vs Ref-NeRF 35.96; 33.24 vs Ref-NeuS 30.49), and the ablations show each component pulls its weight. The code is public. That part is solid and worth taking seriously.\n\nThe soft spot is the central claim. The paper argues that because transmittance T(t) is monotone along a 1D ray, the gradient of the transmittance with respect to translation of the ray segment is an ambiguity-robust normal. Monotonicity of a scalar function along a curve says nothing about the 3D gradient direction. And Eq. (9) is a heuristic: it sums density gradients with equal weights, not with rendering weights, and the connection to the rendering weight is asserted rather than derived. The stress-test example is not contrived: if the learned density artifact is ray-aligned, σ(x)=f((x−o)·d), then every term in the sum is parallel to d and the transmittance gradient degenerates to ±d, the camera direction, regardless of the true surface. The paper does not address this. So the theoretical motivation for the headline concept is unproven in the exact regime it targets.\n\nThe paper also overclaims priority: it says it is \"the first to identify\" the density-gradient limitation, but the related work section already credits Ref-NeRF and others with observing the foggy-artifact problem. And the SOTA comparison omits SpecNeRF and NeRF-Casting, both of which the paper cites, so the \"better than all prior work\" framing is not fully earned. Per-scene MAE on Shiny Blender is not consistently better than ENVIDR or Ref-NeuS (toaster: 8.329 vs 6.496 vs 5.560).\n\nThat said, the method appears to work empirically, and the ablations show the transmittance gradient helps end-to-end even if the per-scene normal accuracy is mixed. The core idea is novel and the engineering is careful. The right fix is to revise the paper: replace the monotonicity hand-wave with a real analysis of when Eq. (9) is a valid normal direction, add the missing baselines, and tone down the \"first to identify\" claim.\n\nBottom line: this deserves a serious referee. I would send it out, with a note that the theoretical argument needs substantial rework and the comparisons need to be complete. The empirical core is likely salvageable, and the transmittance-gradient idea is worth examining regardless.","headline":"The transmittance gradient is a genuinely new normal estimator and the reported gains look real, but the paper's theoretical story for why it works does not survive contact with the math.","tokens_in":16680,"tokens_out":2649,"would_cite":true,"duration_ms":29085,"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":"This paper claims that surface normals for NeRF should be computed from gradients of ray transmittance rather than density, and that doing so makes reconstruction of highly reflective scenes stable enough to render them faithfully.","keywords":["neural radiance fields","surface normal estimation","transmittance gradient","specular reflection","shape-radiance ambiguity","volume rendering","novel view synthesis","dual density activation"],"falsifier":"Render a synthetic semi-transparent shell with a known smooth density profile (for example, a Gaussian falloff around a sphere) and compare Eq. (9) normals against analytic sphere normals; if mean angular error grows with surface curvature or shell thickness, or if the predicted normals rotate when the sampling rate changes, the transmittance-gradient premise is falsified.","tokens_in":15663,"feed_emoji":"🪞","tokens_out":7079,"duration_ms":61458,"temperature":0.7,"pith_summary":"Neural radiance fields reconstruct shiny objects poorly because the density field develops local maxima near semi-transparent specular surfaces, so gradients of density point in erratic directions and normals are unreliable. This paper replaces density gradients with gradients of the transmittance along each ray, which stays monotonic and concentrates where the rendered color is actually determined, giving stable normal estimates even when the predicted shape is ambiguous. It also uses two density activations drawn from one shared MLP output: a sharp exponential density for rendering weights and a smooth softplus density for normal computation, plus a stop-gradient warmup that keeps randomly initialized normals from corrupting the density field early in training. On reflective benchmark scenes the reported rendering quality is higher than the compared baselines, with Shiny Blender PSNR 39.24 versus 35.96 for the Ref-NeRF baseline and Glossy Synthetic PSNR 33.24 versus 30.49 for the Ref-NeuS baseline.","feed_headline":"Transmittance gradient fixes normals in reflective NeRF scenes","feed_subtitle":"Replacing density gradients with light-transmission gradients lifts glossy-scene rendering above prior NeRF and SDF baselines.","key_machinery":"The load-bearing object is the transmittance gradient: for a ray origin $o$, direction $d$, and point $x = o + td$, the paper reparameterizes transmittance as $\\hat{T}(x; d, t) = \\exp(-\\int_0^t \\sigma(x - sd)\\,ds)$ and defines the normal estimate as the normalized gradient of this quantity with respect to $x$ (Eq. (8)), approximated in Eq. (9) by the normalized negative cumulative sum of density gradients over earlier samples along the ray. Two supporting mechanisms carry the argument: the dual activated densities module, which feeds one shared pre-activation density through $\\exp$ for rendering weights and $\\operatorname{softplus}$ for normal computation so sharp boundaries and smooth normals coexist, and a stop-gradient warmup schedule on the predicted-normal loss that gradually increases the gradient flow from normals into the density field. The reflection-aware appearance branch then uses the predicted normal to reflect the viewing direction and conditions a specular color network on the reflected direction.","core_discovery":"The central discovery is that the direction of the transmittance gradient, defined in Eqs. (7)-(9) as the normalized spatial gradient of accumulated transmittance over a ray segment with respect to translation of the segment, points at the true surface normal even when the density field is non-monotonic and the shape is ambiguous. Because transmittance is monotonically decreasing along a ray, its derivative is large where rendering weights are large, so normal estimates remain consistent near semi-transparent surfaces where density gradients vanish. The paper argues that this normal signal, refined by a spatial MLP and combined with the dual activated densities and stop-gradient warmup, resolves the shape-radiance ambiguity that reflection-aware models otherwise fight, and the experiments report the highest scores among the compared methods on Shiny Blender, Glossy Synthetic, and comparable performance on real captures.","pith_inferences":["The transmittance gradient is essentially a ray-wise accumulation of density gradients, so it should transfer to any differentiable ray-marching renderer as a drop-in normal regularizer.","Because the justification uses only ray-wise monotonicity, the same construction likely applies to thin translucent objects such as smoke, glass, and water; this could be tested on datasets with ground-truth normals.","The stop-gradient warmup may benefit other reflection-aware models that tie MLP normals to density-derived normals, since it decouples early normal supervision from density optimization."],"forward_implications":["Normals estimated from transmittance stay aligned with surfaces in semi-transparent and specular regions where density-gradient normals scatter, enabling clean normal maps (MAE 4.241 on Shiny Blender).","A single density MLP with two activations can simultaneously support sharp object boundaries in rendering and smooth normals for reflection parameterization, removing the need for separate geometry priors.","The stop-gradient warmup prevents early random normal predictions from flattening or oversmoothing the density field, preserving fine geometry such as water waves.","With these normals, a reflection-direction-based appearance model renders mirror-like scenes at higher fidelity than more complex ray-tracing or microfacet models on the tested datasets."],"supporting_citations":[{"why":"Defines the volume rendering integral and transmittance formula that the transmittance-gradient construction reparameterizes.","marker":"(Mildenhall et al. 2021)"},{"why":"Introduces reflection-direction appearance conditioning and the predicted-normal loss the paper extends, and supplies the Shiny Blender benchmark.","marker":"(Verbin et al. 2022)"},{"why":"Defines the density-gradient normal estimate (Eq. 6) that the paper identifies as unreliable on specular surfaces.","marker":"(Boss et al. 2021)"},{"why":"Shows the derivative of transmittance equals the rendering weight, the monotonicity fact motivating the transmittance gradient.","marker":"(Tagliasacchi and Mildenhall 2022)"},{"why":"Provides the microfacet reflection model and opaque-surface orientation loss used as an alternative baseline the paper compares against.","marker":"(Mai et al. 2023)"},{"why":"Represents SDF-based reflection reconstruction with reflection scores, a competing approach on specular scenes.","marker":"(Ge et al. 2023)"}],"fun_headline_variants":["Transmittance gradients resolve NeRF shape ambiguity","NeRF normals from transmittance, not density","Reflective scenes: transmission beats density for normals","Ambiguity-robust normals for shiny NeRF reconstruction","Light transmission gradients fix specular NeRF geometry"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the transmittance gradient, approximated by a cumulative sum of density gradients along the ray, gives the true surface-normal direction even when the shape is ambiguous, a claim justified only by 1D monotonicity of transmittance rather than by a proof about its 3D gradient.","fun_headline_variants_meta":{"raw":{"variants":["Transmittance gradients resolve NeRF shape ambiguity","NeRF normals from transmittance, not density","Reflective scenes: transmission beats density for normals","Ambiguity-robust normals for shiny NeRF reconstruction","Light transmission gradients fix specular NeRF geometry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1247,"prompt_tokens":893,"completion_tokens":354,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":276}},"tokens_in":509,"tokens_out":354,"duration_ms":39520,"temperature":1.0,"reasoning_tokens":276,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:01:02.176562+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a synthetic semi-transparent shell with a known smooth density profile (for example, a Gaussian falloff around a sphere) and compare Eq. (9) normals against analytic sphere normals; if mean angular error grows with surface curvature or shell thickness, or if the predicted normals rotate when the sampling rate changes, the transmittance-gradient premise is falsified.","supporting_citations":[{"cited_title":"T.; and Srinivasan, P","cited_arxiv_id":null,"evidence_quote":"Introduces reflection-direction appearance conditioning and the predicted-normal loss the paper extends, and supplies the Shiny Blender benchmark."},{"cited_title":"T.; Liu, C.; and Lensch, H","cited_arxiv_id":null,"evidence_quote":"Defines the density-gradient normal estimate (Eq. 6) that the paper identifies as unreliable on specular surfaces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the microfacet reflection model and opaque-surface orientation loss used as an alternative baseline the paper compares against."}],"review_version":1}