{"id":"09c2d01a-9a0b-4454-8f8b-b6b7d06eafda","arxiv_id":"2505.03362","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FreNeuS adds gradient-guided ray sampling and high-frequency-weighted color loss to NeuS, reducing mean Chamfer distance from 0.84 to 0.73 on DTU and from 1.97 to 1.11 on six NeRF-synthetic scenes.","lead":"A new 3D reconstruction method, FreNeuS, uses image gradients to find detailed regions and then spends more training effort on them. This sharpens surface details in neural implicit reconstruction, improving Chamfer distance on standard benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated equivalence between image gradients and geometric detail is load-bearing; the same high-frequency map drives sampling (Eq. 3) and loss weighting (Eq. 5), so a texture/geometry mismatch would directly bias the SDF, and no experiment isolates it.","rationale":"The reader's weakest_assumption identifies exactly the condition on which the paper's central claim depends: Section 3.1 equates image-gradient energy with surface detail, and the same unvalidated map is used to allocate rays and to weight the color loss. I agree that this is the most load-bearing point. The additional internal problems I noticed (the undefined threshold in Eq. (2), the ambiguous w=IH/IL in Eq. (3), and the omitted weights a and b in Eq. (4)) reinforce the need for a controlled test but are secondary. Because no experiment in the paper separates textured appearance from geometric shape, the conditional verdict is the right level of confidence; the controlled texture-vs-geometry experiment would either support the central claim or require a rejection of the proposed mechanism. No change to the reader's verdict is needed on the current evidence.","tokens_in":10860,"tokens_out":8300,"duration_ms":91841,"concrete_test":"Pick one synthetic scene geometry (e.g., Ship or Lego) and render it twice under identical cameras and lighting: once with high-frequency checkerboard albedo texture and once with uniform albedo. Train NeuS and FreNeuS on each render and compare Chamfer distance to the same ground-truth mesh. If FreNeuS's gain over NeuS is large in the textured render but near zero or negative in the uniform-albedo render, the high-frequency map is steering the SDF by appearance rather than by geometry. If the uniform-albedo render still shows comparable gain, the concern is refuted. Report the threshold used in Eq. (2) and the values of a and b used in Eq. (4).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that FreNeuS recovers geometric surface detail requires that thresholded image-gradient magnitude marks geometric surface detail. Section 3.1 explicitly groups depth changes, color changes, and illumination changes into a single high-frequency class, and Eq. (2) converts gradient intensity into a binary IH/IL partition. That partition is then the only input to both proposed mechanisms: dynamic ray allocation in Eq. (3) and the color-loss weights a and b in Eqs. (4)-(5). If strong albedo texture on a geometrically flat region falls into IH, those pixels receive more rays and a larger loss weight, so the SDF is pushed to encode appearance variation as geometry. The DTU and NeRF-synthetic benchmarks never separate texture from shape, so the reported Chamfer improvements are consistent with better photometric fitting in textured regions rather than with better geometry. The paper's own Mic result (Table 2, CD 2.7 vs HF-NeuS 0.72) shows the mechanism can hurt when high-frequency appearance is absent or misleading. The missing threshold value and the under-specified ratio w=IH/IL in Eq. (3) make the intended sampling allocation unrecoverable from the paper, so the mechanism cannot currently be checked as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes FreNeuS, an extension of the NeuS neural implicit surface reconstruction framework. The method computes a high-frequency map per input image from Gaussian-filtered gradient magnitude (Section 3.1), partitions pixels into high-frequency (IH) and low-frequency (IL) sets, and uses this partition in two ways: to allocate ray samples between the two sets (Eq. (3)) and to weight the photometric color loss (Eqs. (4)-(5)). Experiments on 15 DTU scenes and six NeRF-synthetic scenes report Chamfer-distance improvements over NeuS (0.84 -> 0.73 on DTU) and over HF-NeuS (1.12 -> 1.11 on synthetic), together with ablations (Table 3), a generalization experiment on HF-NeuS (Table 5), and runtime/memory comparisons (Tables 4 and 6).","tokens_in":11116,"tokens_out":18855,"duration_ms":164049,"significance":"The proposed modules are cheap, modular, and are demonstrated to transfer to HF-NeuS, which makes the idea potentially useful to the community if the mechanism works as described. The paper deserves credit for standard benchmark evaluation, an honest ablation study, a generalization experiment, and for reporting the negative Mic result rather than omitting it; the claims are falsifiable and the computational overhead is small (roughly 3 minutes of detection per scene). I concur with the review note that the circularity concern does not land: the high-frequency map is computed from the input images, and Chamfer distance is measured post hoc. However, the significance of the reported gains is limited by three load-bearing problems: the sampling equation as written does not implement the stated oversampling objective; the assumed equivalence between image gradients and geometric detail is unvalidated and can bias the SDF toward appearance; and key hyperparameters (detection threshold, a, b) are unreported, so the method cannot be checked or reproduced. The synthetic-benchmark headline improvement is 0.01 and is driven by one outlier scene.","major_comments":[{"comment":"The sampling allocation as written cannot deliver the stated objective. With w = IH/IL (a ratio of pixel counts), the fraction of the batch B allocated to high-frequency pixels is Hnum/B = w, while the fraction of the image that is high-frequency is IH/(IH+IL); the per-pixel oversampling factor relative to uniform random sampling is (IH+IL)/IL, which is close to 1 precisely when high-frequency pixels are rare (the regime the paper says motivates the method) and which grows only as the high-frequency fraction approaches 1/2. In the opposite regime, IH >= IL gives w >= 1 and Lnum = (1-w)*B <= 0, which is not a valid allocation. For example, a scene with 5% high-frequency pixels (w = 0.053, B = 512) yields Hnum of about 27 rays, compared with about 26 rays under uniform sampling, so the mechanism is nearly a no-op in the intended setting and cannot explain the +sampling gain in Table 3 (0.84 -> 0.75). Please state the exact allocation used in the experiments (for instance, an explicit oversampling factor beyond the population ratio) and correct Eq. (3) and the surrounding text accordingly.","section":"3.2, Eq. (3)"},{"comment":"The load-bearing assumption that thresholded image-gradient magnitude marks geometric surface detail is never validated. Section 3.1 bundles depth value changes, color changes, and illumination changes into a single high-frequency class, so albedo texture on a geometrically flat surface is labeled high-frequency and receives both more rays (Eq. (3)) and a larger color-loss weight (Eqs. (4)-(5)). Because these are the only two mechanisms added to NeuS, a texture/geometry mismatch directly biases the learned SDF toward encoding appearance as geometry. The paper's own Mic result (Table 2: CD 2.70 vs. 0.72 for HF-NeuS) is consistent with this risk, and the offered explanation (lack of high-frequency lines in the stent portion) is not obviously supported by the scene content. No experiment separates texture from shape; a concrete test would be a synthetic scene with strong albedo texture on a flat surface, or a shape with geometric detail but uniform albedo. The authors should add such an experiment or otherwise show that the high-frequency map tracks geometry rather than appearance.","section":"3.1, Eqs. (2)-(5)"},{"comment":"Reproducibility: key parameters and definitions are missing. Section 3.1 says boundary tracing is performed based on the grayscale threshold, but the threshold value is never reported; the threshold of 25 mentioned in Section 4 is the surface-extraction threshold for the reconstructed meshes, not the high-frequency detection threshold. The loss weights a and b in Eq. (4) are likewise never specified, so the weighted color loss in Eq. (5) is underspecified, and it is not stated whether a > 1 > b or some other relation is intended. In addition, Eq. (2) is circular as written (the condition defining IH refers to P in IH), and the relationship between the continuous gradient magnitude M(x,y) of Eq. (1) and the binary partition of Eq. (2) is unclear. Without these values the method cannot be reproduced, and the sensitivity of Tables 1-3 to these parameters is not assessed.","section":"3.1 and 3.3"},{"comment":"Tables 1-3 report single runs with no variance or multiple seeds, and the headline synthetic-benchmark improvement is fragile. In Table 2, the mean Chamfer-distance improvement over HF-NeuS is 1.11 vs. 1.12 (about 1%) and is driven by the Materials scene (1.08 -> 0.13), while FreNeuS is substantially worse on Lego (1.30 vs. 0.94) and Mic (2.70 vs. 0.72); no discussion of this outlier is given. In the DTU ablation (Table 3), the +sampling and +Fre-constraints rows each produce the same mean (0.75) with different per-scene patterns, which is consistent with run-to-run variation. The authors should report standard deviations or multiple seeds and discuss the Materials result before claiming consistent improvement over HF-NeuS.","section":"Tables 1-3"}],"minor_comments":[{"comment":"The Eikonal loss has mismatched parentheses; it should read (1/|I|) * sum over i of (||grad S(x_i,y_i,z_i)|| - 1).","section":"Eq. (6)"},{"comment":"The FreNeuS training-time entry reads '9' without the 'h' unit used elsewhere in the table.","section":"Table 6"},{"comment":"References [25] and [29] both cite the 2DGS paper, and the baseline list in Section 4 labels [28] as VolSDF although VolSDF is [4]; the label 'VOLSD' in Table 2 is a typo.","section":"References and Table 2"},{"comment":"The text claims that neither proposed module adds extra computational time and then states that high-frequency detection takes about 3 minutes per scene; these statements should be reconciled.","section":"4.2.3"},{"comment":"The caption says 'Second to fifth columns: NeuS, HF-NeuS, 3DGS, 2DGS, and OURS' but lists five methods for four columns.","section":"Figure 3 caption"},{"comment":"The description of Mip-NeRF and Tri-MipRF as expanding the sampling area around each ray is inaccurate; these methods integrate over cone or tri-mip regions for anti-aliasing rather than widening per-ray point sampling.","section":"Section 1"},{"comment":"The text mentions 'the first Gaussian filter G(x,y)', implying a second filter that is never introduced; the filter size and sigma are also unreported.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope, but its contribution is a training heuristic whose gains over HF-NeuS on the synthetic benchmark are marginal (0.01 mean CD) and whose main mechanism is incorrectly specified in Eq. (3). If the authors fix the sampling allocation, report all hyperparameters, and add a texture-versus-geometry isolation experiment, the paper could become a solid applied contribution; otherwise the headline claims are not supportable. I would also note that the related-work discussion does not address existing nonuniform and importance-based ray-sampling strategies in neural rendering, which is directly relevant to assessing the novelty of the dynamic-sampling contribution. The manuscript also needs thorough language editing throughout."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read on FreNeuS. The new thing is simple: take an image-gradient-based high-frequency mask and use it twice, once to allocate rays between high- and low-frequency pixels and once to weight the color loss. The combination inside NeuS is not in the cited prior work, and the generalization experiment on HF-NeuS is a genuinely nice touch — both modules improve a second NeuS-based method, which supports the modularity claim better than most papers of this kind. The gains are real but modest (DTU mean 0.73 vs 0.84 for NeuS; NeRF-synthetic 1.11 vs 1.12 for HF-NeuS), and the ablation shows both modules earn their keep.\n\nThat said, the most load-bearing equation is wrong as written. Eq. (3) sets w = IH/IL and Hnum = w*B. In any normal image IH << IL, so w < 1 and the high-frequency set gets fewer than half the rays. That is the opposite of the stated objective, which is to increase sampling in detailed regions. The intended allocation is unrecoverable from the paper; either the ratio is inverted or the formula is restructured.\n\nThe bigger conceptual soft spot is the unvalidated equation 'image gradient equals geometric detail'. Section 3.1 groups depth, color, and illumination changes into one class, and the same mask drives both sampling and loss weighting. On a geometrically flat but texture-rich surface, extra rays and extra loss weight will push the SDF to encode appearance as shape. Neither benchmark separates texture from geometry, so the reported Chamfer gains are consistent with better photometric fitting rather than better geometry. The Mic result (2.7 vs 0.72 for HF-NeuS) shows the mechanism can hurt, and the authors' explanation is speculative. This is the central claim, and no experiment isolates it.\n\nMinor but real: the high-frequency threshold, the a and b weights, and the Gaussian parameters are never given, so reproduction is impossible; six of eight NeRF-synthetic scenes are used without explanation; no code is released; and the reference list has formatting sloppiness. The circularity concern one might raise about a training heuristic does not apply here — nothing is fitted to the observed Chamfer values, and the paper measures rather than predicts them.\n\nBottom line: a serious referee can get value from this. The recipe is cheap and modular, and the HF-NeuS transfer experiment is a good sign. But the authors need to fix Eq. (3), report the missing hyperparameters, and add an experiment that decouples texture from geometry before the phrase 'reconstructs fine surface details' means what they want it to mean.","headline":"A cheap, modular NeuS improvement with a genuinely nice transfer experiment, but the key sampling equation as written does the opposite of what the paper claims, and the mapping from image gradients to geometric detail is never validated.","tokens_in":11668,"tokens_out":5020,"would_cite":false,"duration_ms":46900,"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":"FreNeuS argues that guiding a NeuS-style renderer with image-gradient high-frequency maps recovers fine surface detail that random sampling smooths away, and that the guidance transfers to any NeuS-based method.","keywords":["neural implicit surfaces","NeuS","high-frequency detail","dynamic ray sampling","gradient-weighted loss","3D surface reconstruction","volume rendering","Chamfer distance"],"falsifier":"Render a synthetic scene with a flat plane covered by a high-contrast printed texture, reconstruct it with FreNeuS and with NeuS, and compare both surfaces to the known ground-truth plane. If FreNeuS allocates extra rays and loss weight to the textured pixels yet the recovered geometry stays flat and its Chamfer distance to ground truth is no better than NeuS's, the gradient map is responding to appearance rather than surface detail.","tokens_in":10654,"feed_emoji":"📐","tokens_out":6944,"duration_ms":58842,"temperature":0.7,"pith_summary":"FreNeuS argues that the fine surface details neural implicit reconstruction tends to smooth away are concentrated in the high-frequency regions of the input images, and that those regions can be located cheaply from pixel gradients and fed back into the reconstruction loop. The method builds on the NeuS signed-distance-function renderer and adds two modules: dynamic ray sampling, which allocates more rays to high-frequency pixels, and a weighted color loss that emphasizes the same pixels. On the DTU benchmark the complete method lowers mean Chamfer distance from NeuS's 0.84 to 0.73, and on six NeRF-synthetic scenes it averages 1.11 against HF-NeuS's 1.12, with qualitative gains in roof textures, tooth gaps, and leg folds. The authors further claim the two modules transplant into any NeuS-based framework, demonstrating the transfer by improving HF-NeuS's DTU mean from 0.77 to 0.71.","feed_headline":"Gradient-guided rays sharpen 3D surface details","feed_subtitle":"FreNeuS reweights rays toward high-frequency pixels, cutting mean Chamfer distance on DTU from 0.84 to 0.73.","key_machinery":"The load-bearing object is the high-frequency map. The input image is Gaussian-denoised, its horizontal and vertical gradients $G_x(x,y)$ and $G_y(x,y)$ are combined into gradient intensity $M(x,y)=\\sqrt{G_x^2+G_y^2}$, and a grayscale threshold splits pixels into a high-frequency set $I_H$ and a low-frequency set $I_L$. This map drives two mechanisms: dynamic sampling, which draws $H_{\\mathrm{num}}=wB$ rays from $I_H$ and $L_{\\mathrm{num}}=(1-w)B$ from $I_L$ in each training step, and a high-frequency constraint, a per-pixel weight map $W_{\\mathrm{High}}$ that assigns weight $a$ to high-frequency pixels and $b$ to all others inside the color loss. The total training loss combines Eikonal regularization with $\\lambda$ times this weighted color loss, with $\\lambda=1.2$.","core_discovery":"The central claim is that a thresholded gradient-intensity map of each input view is a reliable proxy for where 3D surface detail is located, and that using this map twice—once to reallocate ray samples and once to reweight the color reconstruction loss—recovers detail that random sampling misses without adding a second network or extra training time. FreNeuS reports mean Chamfer distances of 0.73 on the DTU benchmark (baseline NeuS: 0.84) and 1.11 on six NeRF-synthetic scenes (HF-NeuS: 1.12), along with qualitative improvements such as the sail cut-outs in the Ship scene and the inner circle of the ball in Material. The paper also claims the two modules are framework-agnostic within the NeuS family, and demonstrates this by adding them to HF-NeuS, whose mean DTU error drops from 0.77 to 0.71.","pith_inferences":["Applied to depth or normal maps instead of RGB color, the same gradient-guided reweighting could separate geometric detail from painted texture, a variant the paper does not test.","The gradient map could also steer densification or loss weighting in Gaussian-splatting reconstruction, since the paper cites over-reconstruction in high-variance image regions as that family's weakness.","Using a per-scene adaptive threshold instead of a fixed grayscale threshold would stabilize the high-frequency set across datasets with different contrast, which the paper leaves unspecified."],"forward_implications":["Relying on image gradients rather than extra network capacity means fine details can be recovered at essentially NeuS's training cost, about 9 hours on an RTX A6000 and roughly 30 MB per scene.","FreNeuS reaches an average Chamfer distance of 0.73 on DTU, below NeuS (0.84), HF-NeuS (0.77), and 2DGS (0.80) on the same scenes.","On six NeRF-synthetic scenes the method averages 1.11, slightly below HF-NeuS's 1.12, with the largest gains in the Ship and Material scenes.","Adding the two modules to HF-NeuS lowers its DTU average from 0.77 to 0.71, supporting the claim that the guidance transfers to any NeuS-based framework.","Against LoD-NeuS's global enhanced sampling, FreNeuS matches its average error (0.73 vs 0.72) while using about 8.7 GB GPU memory instead of 13 GB."],"supporting_citations":[{"why":"Provides the NeuS signed-distance-function volume rendering baseline that FreNeuS builds on and compares against.","marker":"[3]"},{"why":"Supplies the NeRF volume rendering formulation and the NeRF-synthetic dataset used for the high-frequency detail experiments.","marker":"[2]"},{"why":"HF-NeuS is the main high-detail NeuS variant that FreNeuS compares with and later augments to demonstrate generalization.","marker":"[21]"},{"why":"DTU is the real-scene multi-view benchmark whose 15 scenes carry the main Chamfer distance evaluation.","marker":"[27]"},{"why":"LoD-NeuS is the global enhanced sampling method used in the efficiency comparison showing FreNeuS matches its error with less GPU memory.","marker":"[30]"},{"why":"VolSDF is a competing neural implicit surface method included as a baseline in the DTU and synthetic comparisons.","marker":"[4]"},{"why":"Eikonal regularization is the geometry-smoothness loss that FreNeuS retains alongside its weighted color loss.","marker":"[26]"}],"fun_headline_variants":["FreNeuS guides rays with gradients to sharpen surface details","Gradient maps focus neural surface reconstruction on fine detail","High-frequency weighting improves NeuS-based 3D detail","Pixel-gradient sampling recovers crisp edges in NeuS models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline rests on the assumption that pixels with large image-gradient intensity mark the locations of true 3D surface detail, even though color and illumination changes can also produce strong gradients on geometrically flat regions.","fun_headline_variants_meta":{"raw":{"variants":["FreNeuS guides rays with gradients to sharpen surface details","Gradient maps focus neural surface reconstruction on fine detail","High-frequency weighting improves NeuS-based 3D detail","Pixel-gradient sampling recovers crisp edges in NeuS models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000559,"raw_usage":{"total_tokens":2649,"prompt_tokens":929,"completion_tokens":1720,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":1666}},"tokens_in":545,"tokens_out":1720,"duration_ms":11378,"temperature":1.0,"reasoning_tokens":1666,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:52:48.168096+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a synthetic scene with a flat plane covered by a high-contrast printed texture, reconstruct it with FreNeuS and with NeuS, and compare both surfaces to the known ground-truth plane. If FreNeuS allocates extra rays and loss weight to the textured pixels yet the recovered geometry stays flat and its Chamfer distance to ground truth is no better than NeuS's, the gradient map is responding to appearance rather than surface detail.","supporting_citations":[{"cited_title":"Mildenhall, P","cited_arxiv_id":null,"evidence_quote":"Supplies the NeRF volume rendering formulation and the NeRF-synthetic dataset used for the high-frequency detail experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"HF-NeuS is the main high-detail NeuS variant that FreNeuS compares with and later augments to demonstrate generalization."},{"cited_title":"Jensen, A","cited_arxiv_id":null,"evidence_quote":"DTU is the real-scene multi-view benchmark whose 15 scenes carry the main Chamfer distance evaluation."},{"cited_title":"Zhuang, Q","cited_arxiv_id":null,"evidence_quote":"LoD-NeuS is the global enhanced sampling method used in the efficiency comparison showing FreNeuS matches its error with less GPU memory."},{"cited_title":"Yariv, J","cited_arxiv_id":null,"evidence_quote":"VolSDF is a competing neural implicit surface method included as a baseline in the DTU and synthetic comparisons."}],"review_version":1}