{"id":"88bb4e0f-44ac-4e23-afb6-e61f78dd927c","arxiv_id":"2504.20222","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FreBIS replaces VolSDF's single encoder with three frequency-band encoders and a dissimilarity-based weighting module, yielding small rendering-quality gains on 9 BlendedMVS scenes.","lead":"Neural networks that reconstruct 3D surfaces from photos often lose fine detail. FreBIS splits surface detail into three frequency bands, trains a separate encoder for each, and combines them with a weighting module, reporting slightly sharper renderings on a standard 3D dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Surface reconstruction claim rests only on rendering metrics and qualitative meshes; no quantitative geometry evaluation on a dataset with ground-truth meshes.","rationale":"The reader's weakest assumption identifies the absence of any quantitative geometry metric as the main gap, and I agree. The paper's own Section 5.1 states that ground-truth mesh is not available for BlendedMVS, so all surface-quality evidence is qualitative. Rendering metrics cannot substitute for geometric fidelity, a well-known limitation in neural surface reconstruction: appearance can be explained by view-dependent color networks even when the SDF is inaccurate. The reported rendering gains are also modest and lack statistical significance (no error bars, no multiple seeds), which weakens even the rendering-fidelity part of the claim. A secondary issue, not raised by the reader, is that the method's causal mechanism is not isolated: no ablation with the identical multi-encoder architecture but without frequency band splitting appears in the paper, so the observed gains could be due to the increased architectural capacity or the multi-branch design rather than frequency stratification per se. Both gaps point to the same conclusion: the paper is not yet ready for acceptance as-is, but the ideas are plausible and the conditional verdict is appropriate. A DTU-style geometric evaluation with error bars would settle the primary concern; if it passes, the verdict should be upgraded to accept, and if it fails, the central claim should be revised.","tokens_in":14537,"tokens_out":8987,"duration_ms":91614,"concrete_test":"Evaluate FreBIS and the matched-parameter Scaled-up VolSDF baseline on the DTU benchmark (which provides ground-truth laser-scanned meshes), computing Chamfer distance and F-score for the reconstructed surfaces over at least 8 scenes and 3 random seeds, reporting mean and standard deviation. If FreBIS does not show a statistically significant improvement on these geometric metrics, the central claim about 3D surface reconstruction quality is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that FreBIS improves 3D surface reconstruction quality is not supported by the evidence provided. Table 1 reports only rendering metrics (PSNR/SSIM/LPIPS) on BlendedMVS, and Section 5.1 explicitly notes that ground-truth meshes are unavailable for this dataset; mesh quality is assessed only through qualitative images (Fig. 5). Rendering fidelity does not imply geometric accuracy: a model can fit photometric cues while producing an incorrect zero-level set, particularly in textureless or occluded regions. The quantitative improvements are also small (PSNR +0.15 dB vs VolSDF and +0.48 dB vs the matched-parameter Scaled-up VolSDF; SSIM +0.003/+0.004) and are reported without error bars or multi-seed runs, so 'significant' is not established. Additionally, the causal role of frequency stratification is not isolated: there is no ablation with the same multi-encoder architecture where all encoders receive the full frequency input, so the gains could stem from the wider multi-branch design rather than frequency band assignment. Together these gaps leave the surface-quality part of the headline claim unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FreBIS, a modification of neural implicit surface reconstruction in which the single positional-encoding encoder of VolSDF is replaced by three parallel encoders, each dedicated to a band of positional frequencies (low, middle, high). A redundancy-aware weighting module computes feature importance from pairwise similarities and uses these weights to combine the three encodings before decoding into an SDF value and appearance feature. The model is trained with photometric and Eikonal losses and evaluated on 9 BlendedMVS scenes against VolSDF and a parameter-matched scaled-up VolSDF, reporting PSNR, SSIM, and LPIPS for rendering plus qualitative mesh comparisons.","tokens_in":14786,"tokens_out":6220,"duration_ms":60637,"significance":"If the claims are supported, FreBIS offers a simple and potentially generalizable architectural improvement for neural implicit surface methods: frequency stratification across multiple encoders, with a matched-parameter baseline to control for capacity. The inclusion of the Scaled-up VolSDF baseline is a genuine strength, as is the per-scene reporting across 9 scenes. The gains over baselines are consistent but small, and the central surface-reconstruction claim currently rests on qualitative mesh comparisons rather than quantitative geometry metrics, so the significance of the paper depends on additional validation.","major_comments":[{"comment":"The claim of improved reconstructed 3D surfaces is not supported by quantitative evidence. Section 5.1 states that ground-truth meshes are unavailable for BlendedMVS, and Table 1 reports only rendering metrics (PSNR, SSIM, LPIPS); the mesh-quality comparison is qualitative (Fig. 5). Rendering fidelity does not directly measure geometric accuracy, so the abstract's claim of improvements in the 'quality of the reconstructed 3D surfaces' goes beyond the presented evidence. Please add a quantitative geometry evaluation on a dataset with ground-truth meshes (e.g., DTU, using Chamfer distance or F-score), or restrict the claim to rendering fidelity.","section":"§5.1, §5.2, Table 1"},{"comment":"The reported improvements over baselines are small (mean PSNR 27.38 vs 27.23 for VolSDF and 26.90 for Scaled-up VolSDF; LPIPS 0.024 vs 0.025 and 0.027) and are given as single runs without error bars, number of seeds, or any significance test. The word 'significant' in the abstract is therefore not established. Please provide variance estimates across multiple runs or rephrase the claim to avoid overstatement.","section":"§5.2, Table 1"},{"comment":"The design does not isolate the effect of the frequency split as the cause of the gains. The comparison against Scaled-up VolSDF changes both parameter count and architecture, and no ablation is reported in which a three-encoder architecture receives the full set of positional frequencies in every branch. Without such a control, the improvements could stem from the multi-branch ensemble rather than from frequency stratification. Please add an ablation with identical three-encoder architecture where all encoders get the same full positional encoding.","section":"§4.1, §5.2, Table 1"},{"comment":"The paper states that the redundancy-aware weighting module 'promotes mutual dissimilarity' of the encoded features, but it does not contain a loss term enforcing dissimilarity; it only reassigns weights based on the current similarity of the features. No metric of feature redundancy or complementarity is reported, and the quantitative effect of the module on the Dog scene is a small gain (+0.25 dB PSNR over simple averaging). The band-wise mesh visualizations in Fig. 7 are expected given that each encoder receives different frequency inputs. Please provide a quantitative redundancy measure (e.g., average pairwise absolute cosine similarity of encoder features) or soften the claim about promoting dissimilarity.","section":"§4.2, Eqs. (4)–(7), Table 3"}],"minor_comments":[{"comment":"The text says 'The training loss (Eq. 4) is computed with λ = 0.1', but Eq. 4 is the similarity matrix in the redundancy module; the training loss is Eq. 8. Please correct the cross-reference.","section":"§5.1"},{"comment":"Table 2 is difficult to read because many entries appear concatenated without spaces (e.g., '26.22 27.4827.2930.52'). Please reformat the table with proper cell boundaries.","section":"Table 2"},{"comment":"In the supplementary, Table 6 uses the header 'NL,NM,NH' for what is described as a study of the number of encoder layers; since NL,NM,NH was previously used for frequency counts, please use distinct notation (e.g., LL, LM, LH) to avoid confusion.","section":"Supplementary Table 6"},{"comment":"The phrase 'gains of up to 2% on SSIM' is ambiguous about whether the percentage is relative or absolute; please specify the computation.","section":"§5.2"}],"recommendation":"major_revision","confidential_remarks":"The idea is simple and potentially useful, but the central claim about surface reconstruction quality needs quantitative geometry evaluation, and the attribution to frequency stratification needs an additional ablation. The paper is within the journal's scope. The citation and related-work coverage appear adequate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FreBIS is a clean, clearly explained architectural variant of VolSDF that splits positional frequency bands across three dedicated encoders and combines them with a dissimilarity-based weighting module. That specific combination is new relative to the cited VolSDF/NeuS literature. I believe the paper does what it claims mechanically: the rendering numbers are consistently a hair better than VolSDF and the scaled-up VolSDF baseline on most of the 9 BlendedMVS scenes, and the supplementary ablations on frequency levels and encoder depths are useful. The scaled-up baseline is a fair control for parameter count, and the weighting module's self-referential design is not a fatal circularity because it is trained against external photometric loss, not its own outputs.\n\nThe soft spots are real but not disqualifying. The biggest is that the central claim about improved 3D surface reconstruction rests on qualitative mesh visuals only. Section 5.1 explicitly says ground-truth meshes are unavailable for BlendedMVS, and the paper provides no Chamfer distance or F-score on a dataset like DTU. Rendering fidelity does not guarantee geometric accuracy, so the surface-quality part of the headline claim is under-supported. Second, the rendering improvements are small (mean PSNR 27.38 vs 27.23; SSIM 0.958 vs 0.955) and there are no error bars or multi-seed runs, so \"significant\" is not statistically established. Third, there is no ablation where the same three-encoder architecture receives the full frequency input in every branch, so the gains could partly come from the wider multi-branch design rather than the frequency-band assignment itself. To be fair, the paper does not hide the Bread scene regression, and the supplement shows the model is fairly robust to frequency-level and depth choices.\n\nThis is worth sending to peer review, not desk reject. A serious referee should ask for a quantitative geometry benchmark, error bars, and that missing full-frequency ablation. If the authors supply those and the gains hold, this becomes a useful empirical contribution. For a reader outside the neural implicit surface niche, the value is modest; for someone working on encoder design for SDFs, it is a plausible baseline to compare against.\n\nRecommendation: engage with it, but requantify the surface claim before accepting.","headline":"FreBIS is a clean frequency-stratified encoder variant of VolSDF with small rendering gains and an unverified surface-quality claim; the missing quantitative geometry evaluation is the main gap.","tokens_in":15304,"tokens_out":1629,"would_cite":false,"duration_ms":18146,"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":"FreBIS claims that splitting scene frequencies across three specialist encoders, weighted by mutual distinctness, improves reconstructed 3D surfaces and rendering for an off-the-shelf method like VolSDF on BlendedMVS.","keywords":["neural implicit surfaces","signed distance functions","frequency-based stratification","positional encoding","multi-view 3D reconstruction","neural volume rendering","redundancy-aware weighting","BlendedMVS"],"falsifier":"Run FreBIS and its baseline on a multi-view benchmark that provides ground-truth meshes, and compare Chamfer distance and F-score between the reconstructed surfaces; if FreBIS's rendering metrics improve while its geometric accuracy matches or trails the baseline, the paper's surface-reconstruction claim is refuted, and if both improve, it is confirmed.","tokens_in":14358,"feed_emoji":"🧊","tokens_out":9352,"duration_ms":82492,"temperature":0.7,"pith_summary":"The paper proposes FreBIS, a way to rebuild the input encoder of a neural implicit surface model so that low-, middle-, and high-frequency surface details are learned by three separate encoders rather than one shared network. Its central claim is that swapping this frequency-stratified encoder into a standard method, VolSDF, improves both the fidelity of rendered views and the visual quality of the reconstructed 3D mesh on the BlendedMVS dataset, at the same parameter count. A second claim is that a redundancy-aware weighting module, which up-weights each encoder's feature to the extent it differs from the other two, helps the encoders encode complementary information and is responsible for a measurable share of the gain. If the claims hold, the result suggests that one cause of detail loss in implicit surface reconstruction is encoder-capacity conflict across frequencies, and that splitting the frequency spectrum is a fix that can be dropped into existing decoders.","feed_headline":"Three frequency bands beat one encoder for 3D surfaces","feed_subtitle":"FreBIS splits a scene into low-, mid-, and high-frequency encoders, lifting VolSDF's renderings and mesh detail on BlendedMVS.","key_machinery":"The load-bearing object is the frequency-stratified encoder plus the redundancy-aware weighting rule. The input coordinate is mapped by a standard positional encoding $\\gamma(x)$ into a vector of sines and cosines at levels $2^0$ through $2^{N-1}$; FreBIS partitions these levels into bands and routes each band to its own MLP encoder, so no single network must explain both macroscopic shape and microscopic texture. The weighting module computes $F = [f_L, f_M, f_H]$, L2-normalizes each column, forms the Gram-style similarity matrix $S = \\bar{F}^T \\bar{F}$, removes the diagonal with $S' = S - I$, and sets the dissimilarity vector $d = (2I - S')\\cdot \\mathbf{1}$, then applies $w = \\mathrm{Softmax}(d/\\tau)$ with default $\\tau=0.5$. This gives higher weight to the feature column least predicted by the others and lower weight to redundant columns, which is what pushes the encoders to specialize; the weighted features $F\\cdot\\mathrm{diag}(w)$ feed a VolSDF-style decoder that outputs an SDF value and an RGB appearance feature.","core_discovery":"On the paper's own terms, the discovery is that the single-encoder design of neural implicit surface methods, not their decoder or volume-rendering machinery, is the bottleneck for scenes containing both smooth structure and fine detail. FreBIS stratifies the positional encoding of a 3D point into three non-overlapping frequency bands, two octaves per band when the total is N=6, and dedicates one 256-dimensional encoder to each band. A redundancy-aware weighting module then forms a normalized feature matrix, computes pairwise cosine similarities between the three feature columns, subtracts self-similarity, and derives a softmax weight for each encoder that grows with its dissimilarity to the other two; the weighted combination is decoded into a signed distance value and an appearance feature. The paper reports that this encoder replacement outperforms VolSDF and a parameter-matched scaled-up VolSDF on BlendedMVS on PSNR, SSIM, and LPIPS in nearly all tested scenes, and that the reconstructed meshes preserve details such as a doll's dress bands and a robot's gun that the baselines blur or lose. It also reports that the frequency bands behave as intended: decoded per-band meshes show low frequencies carrying global shape, middle frequencies rough structure, and high frequencies fine carving, while the weighting visualizations place high-frequency features on detailed regions.","pith_inferences":["If the frequency-specialization story is right, the same stratification should transfer to radiance-field-only models and to multi-resolution grid encoders, where frequency bands are already separated spatially rather than by octave; a natural test is whether per-octave feature decorrelation improves those baselines as well.","The weighting module is an implicit diversity regularizer: by discouraging redundancy among feature columns, it resembles decorrelation penalties, and one could push further by adding an explicit loss that penalizes mutual information between band features, then check whether the gains grow.","The paper's evidence for geometric improvement is indirect, so the clean way to confirm the central claim is to measure Chamfer distance and F-score against ground-truth meshes on a dataset that provides them; if those metrics do not improve with rendering, the claim would need to be narrowed to rendering fidelity.","The demonstrated robustness to N=12 suggests a practical recipe: raise the number of frequency levels once stratification is in place, since the low-frequency encoder is protected from the instability that high octaves introduce in single-encoder training."],"forward_implications":["Replacing a single implicit-surface encoder with frequency-band specialists improves rendered image quality and mesh detail on BlendedMVS, and the paper's scaled-up VolSDF baseline shows the gain is not simply extra parameters.","Assigning the six frequency levels evenly across the three encoders is the most stable configuration across scenes, whereas uneven splits help some scenes but hurt others.","FreBIS remains usable at higher total frequency levels (N=9, 12) where a scaled-up single encoder degrades or fails to produce a mesh, because the bands shield low-frequency shape information from high-frequency noise.","Ablating the redundancy-aware weighting and averaging the features instead lowers PSNR, SSIM, and LPIPS on the Dog scene, so the dissimilarity weighting contributes beyond mere stratification.","Since the decoder is off-the-shelf, the encoder design is a drop-in component: any method that consumes point features can adopt the stratification, and the paper suggests combining with object-compositional methods as future work."],"supporting_citations":[{"why":"Supplies the baseline network that FreBIS modifies, the decoder it reuses, and the SDF-to-density conversion used for rendering.","marker":"[52]"},{"why":"Provides the BlendedMVS dataset used for all quantitative and qualitative comparisons.","marker":"[50]"},{"why":"Grounds the positional-encoding and volume-rendering equations that FreBIS stratifies and optimizes.","marker":"[29]"},{"why":"Establishes the alternative SDF-plus-volume-rendering formulation against which FreBIS's design is motivated.","marker":"[43]"},{"why":"Defines LPIPS, one of the three rendering-fidelity metrics used in the comparisons.","marker":"[55]"}],"fun_headline_variants":["Split frequency bands sharpen implicit 3D surfaces","FreBIS: dedicated encoders per frequency band improve 3D","Three bands beat one for neural surface quality","Why single-encoder neural surfaces miss fine detail"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that better rendering metrics and visual mesh inspection are enough to show the reconstructed 3D surface is better, because BlendedMVS has no ground-truth mesh to measure geometry directly.","fun_headline_variants_meta":{"raw":{"variants":["Split frequency bands sharpen implicit 3D surfaces","FreBIS: dedicated encoders per frequency band improve 3D","Three bands beat one for neural surface quality","Why single-encoder neural surfaces miss fine detail"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000784,"raw_usage":{"total_tokens":3522,"prompt_tokens":1069,"completion_tokens":2453,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":685,"completion_tokens_details":{"reasoning_tokens":2399}},"tokens_in":685,"tokens_out":2453,"duration_ms":21063,"temperature":1.0,"reasoning_tokens":2399,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:33:48.997474+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FreBIS and its baseline on a multi-view benchmark that provides ground-truth meshes, and compare Chamfer distance and F-score between the reconstructed surfaces; if FreBIS's rendering metrics improve while its geometric accuracy matches or trails the baseline, the paper's surface-reconstruction claim is refuted, and if both improve, it is confirmed.","supporting_citations":[{"cited_title":"Srinivasan, Matthew Tancik, Jonathan T","cited_arxiv_id":null,"evidence_quote":"Grounds the positional-encoding and volume-rendering equations that FreBIS stratifies and optimizes."},{"cited_title":"Efros, Eli Shecht- man, and Oliver Wang","cited_arxiv_id":null,"evidence_quote":"Defines LPIPS, one of the three rendering-fidelity metrics used in the comparisons."}],"review_version":1}