{"id":"14bbbf3a-a816-44c2-87d0-9b5c4358fa33","arxiv_id":"2411.11691","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A 3D-degradation-aware feature extractor, combining depth-based warping and a restoration head, improves GNeRF reconstruction from degraded source images, and the Objaverse Blur Dataset provides a large-scale benchmark for this task.","lead":"This paper introduces a training-time module that makes generalizable NeRF models more robust to blur, noise, and other degradations in input images, plus a new synthetic blur dataset for training and evaluation. It reports consistent but modest PSNR improvements across several GNeRF models and degradation types.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Depth pseudo-supervision loop is never validated against ground-truth depth the authors could have obtained from their Blender renderer, so the reported gains may come from the restoration head rather than from accurate geometry.","rationale":"The reader identified the self-referential depth supervision as the weakest assumption, and my stress-test confirms this is the most load-bearing concern. The paper's central mechanism is explicitly motivated by NeRFool's finding that depth perturbation is the leading cause of reconstruction degradation; the proposed module therefore claims to improve robustness by stabilizing geometry through a depth estimator. But the depth estimator is trained entirely against the GNeRF's own depth prediction, and the paper offers no external validation of depth quality. Because the authors render the dataset themselves, ground-truth depth is free—its absence from the evaluation is a clear gap. A simple depth-accuracy check (AbsRel/RMSE against Blender ground truth) plus a GT-depth warping ablation would settle whether the gains are attributable to the depth module or to the auxiliary restoration head. The reported quantitative gains are consistent (Tables 1-4, 7-11) and the dataset is a plausible contribution, so I do not see grounds to reject; but the unvalidated core mechanism justifies the CONDITIONAL verdict. The secondary uncertainty about test-split construction (whether the 52 test settings share 3D models with training) is also worth checking but is less damaging than the missing depth validation, because even a perfect object-disjoint split would not address the circularity of the depth supervision. I therefore agree with the reader's weakest_assumption and recommend leaving the verdict unchanged.","tokens_in":18426,"tokens_out":4290,"duration_ms":43009,"concrete_test":"Using the Blender renderer for the Objaverse Blur Dataset, export ground-truth depth maps for the test views and the 52 test settings. Compare the predicted depth D_tar against ground-truth depth with standard metrics (AbsRel, RMSE) across blur levels, for both clean and degraded inputs. Additionally, run an inference-time ablation: replace the predicted depth in the differentiable warping (Eq. 5) with the ground-truth depth map and measure rendered PSNR for the 'Proposed' model. If using ground-truth depth yields a large PSNR gain over the predicted depth, the depth estimator is a bottleneck and its self-supervised loop is not providing reliable geometry. If the gain is small, the module's improvement is driven by the restoration head rather than by accurate depth, weakening the paper's geometry-based motivation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the 3D-degradation-aware feature extractor improves GNeRF robustness by stabilizing geometry. The depth estimator D is supervised only by the GNeRF's own fine depth prediction D_hat_tar (Eq. 7, Sec. 4.3), creating a feedback loop: D's predicted depth defines the warping (Eq. 5), which defines the features that the GNeRF uses to produce D_hat_tar, which then supervises D. If the GNeRF's depth is biased under degradation, D learns that bias, and the loop can reinforce inaccurate geometry. The paper provides no analysis of when this pseudo-supervision is reliable. Table 5 measures only stability of depth against the clean-input prediction, not accuracy; it never compares to ground-truth depth. Crucially, the Objaverse Blur Dataset is rendered from 3D models with Blender, so the authors have immediate access to ground-truth depth maps for every rendered view, yet they do not report depth accuracy. The self-referential supervision could be benign if the photometric loss and the clean-source restoration loss (Eq. 8) keep the loop anchored, but without a depth-accuracy check the claimed geometry-based mechanism is unsupported. The paper's own limitation section (Appendix E) concedes failure under larger degradation but does not examine the depth loop. If the restoration head alone is responsible for the PSNR gains, the module's specific design (depth estimation and warping) is not justified as the cause of robustness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies the robustness of generalizable NeRF (GNeRF) models when source images are degraded by blur, noise, adversarial perturbations, or blur combined with noise, and makes two contributions. First, it introduces the Objaverse Blur Dataset, a Blender-rendered dataset of over 50,000 motion-blurred images generated from 250 Objaverse models at four blur levels via an algorithm that averages latent images along sampled camera trajectories for 3D-consistent degradation. Second, it proposes a model-agnostic 3D-degradation-aware feature extractor: a self-supervised depth estimator (PatchMatchNet-style), a differentiable depth-based warping that aligns nearby views (Eq. 5), and a Restormer-based restoration head producing degradation-robust features for the downstream GNeRF aggregator. Training adds two auxiliary losses — depth pseudo-supervision against the GNeRF's own fine depth prediction (Eq. 7) and a restoration loss against clean source images (Eq. 8). Experiments with NAN, GeoNeRF, and GNT report mostly consistent PSNR/SSIM/LPIPS gains across four degradation settings, a depth-stability analysis (Table 5), and an inference-speed comparison (Table 6).","tokens_in":18794,"tokens_out":14342,"duration_ms":124384,"significance":"The empirical pattern is the paper's main strength: the module improves three structurally different GNeRF backbones (image-based NAN, cost-volume-based GeoNeRF, attention-based GNT) across blur, noise, adversarial, and combined blur-and-noise degradations, and the main metrics are evaluated on held-out target views, so the reported reconstruction gains are not fitted quantities. The dataset construction is specified in enough detail (Algorithm 1, camera and trajectory sampling, scene-dependent blur weights) to be reproducible, and the paper honestly concedes in Appendix E that the method may fail under larger degradation or inaccurate poses, consistent with the diminishing gains at high blur/noise levels in Table 4. If these results hold, the dataset and the plug-in module are practically useful resources for the GNeRF community. The principal weakness is that the stated mechanism — robustness through geometry stabilization via depth estimation and warping — is not directly validated: depth is only reported for stability against the clean-input prediction, never for accuracy, and no ablation isolates the depth/warping component from the restoration head.","major_comments":[{"comment":"The depth estimator D is supervised only by the GNeRF's own fine depth prediction D_hat_tar (Eq. 7), while the features consumed by that same GNeRF are produced by a warping that depends on D's output (Eq. 5); this self-referential loop is never validated against ground-truth depth, although the Objaverse Blur Dataset is rendered from 3D models and ground-truth depth is available at essentially zero cost. Table 5 measures only the deviation of predicted depth from the clean-input prediction (stability), not accuracy, so the paper's central mechanistic claim — that the robustness gain stems from correcting inaccurate geometry (Sec. 4) — remains unsupported. The photometric and restoration losses may anchor the loop, but this is not demonstrated. Please (i) report depth accuracy (absolute and relative error against rendered ground-truth depth) for D and for the GNeRF's fine depth with and without the module on the Objaverse Blur Dataset, and (ii) add an ablation that isolates the depth/warping path from the restoration head, for example by feeding oracle ground-truth depth to the warping or by removing the warping and keeping only the restoration head.","section":"§4.3, Eq. (7); §6.6, Table 5"},{"comment":"The evaluation protocol on the newly introduced dataset is under-specified. The paper states that 52 settings were randomly chosen and that results are averaged over 13 scenes (Appendix D), but it does not state whether these test 3D models are held out from the training set, nor does it state the training protocol for the blur experiments — that is, whether NAN and GeoNeRF were trained from scratch on the Objaverse Blur Dataset, pre-trained elsewhere and fine-tuned, or evaluated with only the proposed module trained. Without this information the reported improvements cannot be interpreted as generalization to unseen scenes. Please specify the exact train/test model split, the number of training settings, and the training protocol, and reconcile the dataset size claims (over 1000 settings, 50,000 images) with the rendering recipe in Sec. 5, since 250 models times 4 levels times 34 images per viewpoint would require roughly one to two viewpoints per model to reach 50,000 images.","section":"§6.1; Appendix D"},{"comment":"The claim of consistent improvement is made without any measure of variance. Several deltas are very small — +0.05 dB at gain 4 in Table 2 and +0.08 dB at blur-4/gain-4 in Table 4 — and the per-scene adversarial results in Table 10 include negative deltas (Fortress −0.41 in the Adversarial-Clean block, Orchids −0.19 and −0.10 in the adversarial blocks), which contradicts the unqualified statement in Sec. 6.3 that the module is effective across all training and test combinations. Please report multiple seeds or per-setting standard deviations, and either qualify the adversarial claim or explain the negative per-scene cases.","section":"§6.3–6.4, Tables 2, 4, 10"}],"minor_comments":[{"comment":"GeoNeRF* at blur level 3 is 23.78 dB in Table 1 but 23.79 dB in Table 8; duplicated numbers should be consistent across the main text and appendix.","section":"Table 1 vs. Table 8"},{"comment":"The loss weights are called λdepth and λrec in the text but λ1 and λ2 in Eq. (9), and the sentence setting them to 1.0 and 0.01 does not say which weight is which; please clarify the mapping.","section":"§6.1 and Eq. (9)"},{"comment":"The text refers to 'Tabs. 8, 9, and 10' when summarizing results, but the main-text tables are numbered 1–6 and the cited tables appear only in the appendix; cross-references should be consistent.","section":"§6.2"},{"comment":"Ldepth is defined only for the target view, while §4.1 states that depths are estimated for all N source images; please state how source-view depth estimates are supervised, if at all.","section":"§4.3 and §4.1"},{"comment":"The summation index i = 1..K is unclear because the restoration head is described as operating on a single source view; please clarify the indexing and define I^GT_i as the clean counterpart of the i-th source image.","section":"Eq. (8)"},{"comment":"The figure caption uses the symbol S^{src} while Eq. (2) defines the feature set S_x = {s_i}; please unify the notation.","section":"Fig. 2 and Eq. (2)"},{"comment":"The relationship between the Objaverse Blur Dataset and the 3D-consistent corruption benchmark of [30] is mentioned only in passing; a sentence clarifying the difference (single corruption type, multi-level blur, GNeRF training scale) would strengthen the dataset positioning.","section":"Sec. 2 (Related Work)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an empirical systems paper with a genuinely useful dataset, and I did not find signs of dishonesty in the experiments; the main risk is over-claiming the mechanism. I recommend requiring (a) ground-truth depth evaluation, which is essentially free given the Blender renderer, (b) a component ablation of depth/warping versus the restoration head, and (c) a precise statement of the train/test split and training protocol for the Objaverse Blur Dataset. The self-referential depth supervision in Eq. (7) is the design point most worth pressing on before publication. A scope note for the editor: the method itself is incremental, so the impact of the paper will depend substantially on the actual release of the dataset and code, including the license and the exact rendering scripts; this should be verified at the revision stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid empirical paper that deserves a serious referee. The new thing is the Objaverse Blur Dataset, a large motion-blur 3D dataset rendered from 250 Objaverse models, and a model-agnostic module that combines a PatchMatchNet-style depth estimator, differentiable warping, and a Restormer-based restoration head to feed 3D-aware features into any GNeRF. The experiments are broader than most: three very different GNeRF models (NAN, GeoNeRF, GNT), four degradation settings (blur, noise, adversarial, blur+noise), plus a real low-light capture. Gains are modest (roughly 0.5–1.2 PSNR for blur, 0.05–0.63 for noise) but consistent across models and levels, and the module beats both Restormer pre-processing and training directly on degraded images. That is a believable package.\n\nThe soft spots are real but not fatal. The depth estimator is supervised only by the GNeRF's own fine depth prediction (Eq. 7), which creates a feedback loop that could reinforce biased geometry. The paper's Table 5 measures depth stability against clean-input predictions, not accuracy, and since the dataset is Blender-rendered they have ground-truth depth available for every view. They should have reported depth error against that. The stress-test note is right that this missing check leaves the 'geometry stabilization' mechanism unproven; the gains might come mostly from the restoration head. That is my main technical complaint, and it is fixable in a revision.\n\nTwo smaller issues. The test set is described as 'randomly chose 52 settings,' but the paper doesn't say whether those settings come from 3D models held out from training. If not, the generalization claim is weaker than stated. And there are no error bars or significance tests anywhere; for a paper with sub-PSNR improvements that matters, though it is the norm in this subfield.\n\nOverall: the dataset is a useful contribution by itself, and the module is a reasonable baseline that the GNeRF-robustness community will want to compare against. I'd send it out. A good referee should ask for the test-split details, a ground-truth depth comparison, and ideally an ablation that keeps the depth warping but removes the restoration head, to see how much each part contributes.","headline":"A useful dataset and a simple, consistently effective plug-in module for degradation-robust GNeRF; the depth-supervision loop is under-validated but fixable.","tokens_in":19304,"tokens_out":3073,"would_cite":true,"duration_ms":28650,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Generalizable NeRFs can be made robust to blur, noise, and adversarial corruption by replacing the image-feature extractor with a depth-aware module.","keywords":["generalizable neural radiance fields","degradation robustness","motion blur dataset","depth estimation","differentiable warping","novel view synthesis","image restoration","self-supervised depth"],"falsifier":"Train the module with an independent depth signal, such as depths estimated from clean views of the same training scenes, instead of the GNeRF's fine depth as pseudo-ground truth. If the robustness gains on blurry or noisy test views disappear, or if the estimated depths drift from the independent depths, the self-referential depth loop is carrying the result.","tokens_in":1968,"feed_emoji":"🛡️","tokens_out":3337,"duration_ms":89705,"temperature":0.7,"pith_summary":"The paper argues that generalizable neural radiance fields (GNeRFs), which synthesize new views from a handful of source images without per-scene training, lose accuracy when those source images are blurry, noisy, or adversarially perturbed. It proposes a model-agnostic replacement for the standard image-feature extractor: first estimate depth from nearby degraded views, warp those views onto each source view using the predicted depth, then restore a clean latent representation with an auxiliary denoiser. On three popular GNeRF designs with different aggregation mechanisms, the module consistently improves novel-view PSNR, SSIM, and LPIPS under blur, noise, adversarial attacks, and blur-plus-noise, and it also steadies depth prediction as blur increases. The paper also introduces the Objaverse Blur Dataset, over 50,000 images rendered from hundreds of 3D models along simulated camera trajectories, as a large-scale benchmark for training and evaluating degradation-robust GNeRFs.","feed_headline":"Depth-aware fix lifts blurred 3D reconstruction","feed_subtitle":"A plug-in feature extractor improves blurry, noisy, and adversarial source views across three NeRF variants.","key_machinery":"The central object is the 3D-degradation-aware feature extractor $\\mathcal{F}$, a two-stage module consisting of a self-supervised depth estimator $\\mathcal{D}$ and an auxiliary restoration head $\\mathcal{R}$. The load-bearing operation is the differentiable homography warping of Eq. (5), which uses estimated depth and relative camera pose to align nearby views to each source view; the concatenated aligned stack $J_i$ of Eq. (6) is what lets the restoration head exploit multi-view, depth-consistent evidence. The depth loss of Eq. (7) closes a self-supervised loop from the GNeRF's own fine depth prediction, and the restoration loss of Eq. (8) provides clean-image supervision; together with the photometric loss they form the total loss of Eq. (9). Because the module outputs a feature tensor identical in shape to the original CNN feature map, it plugs into any GNeRF without changing the volume-rendering stage.","core_discovery":"The paper's central claim is that the degradation vulnerability of GNeRF models is largely a geometry problem: when source images are degraded, the features sampled along a ray become inconsistent, depth prediction drifts, and rendering quality collapses. To counter this, the proposed 3D-degradation-aware feature extractor first estimates a coarse depth map for each source view from its nearest neighboring views using a multi-scale patch-match stereo cost volume, then applies differentiable homography warping, $\\pi^k_i(p) = K_k \\cdot (R^k_i \\cdot (K_i^{-1} \\cdot p \\cdot D_i[p]) + t^k_i)$, to align the nearby views and their depths onto the source view. The aligned stack $J_i = \\text{Concat}(I_i, D_i, \\tilde{I}_1, \\ldots, \\tilde{I}_K, \\tilde{D}_1, \\ldots, \\tilde{D}_K)$ is fed to a restoration head built from multi-Dconv head transposed attention blocks, whose output features have the same shape as a conventional CNN feature map, so any existing GNeRF can use them without altering its renderer. The depth estimator is trained without ground-truth depth by regressing the GNeRF's own fine depth prediction through a self-supervised loss, $\\mathcal{L}_{\\mathrm{depth}} = \\text{SmoothL1}(\\hat{D}_{\\mathrm{tar}}, D_{\\mathrm{tar}})$, while the restoration head is supervised by clean source images through an auxiliary restoration loss. Combined with the original photometric loss, these terms train the module end-to-end along with the GNeRF; the paper reports consistent quantitative and visual improvements across blur levels, noise gain levels, adversarial perturbations, and combined blur-plus-noise degradation.","pith_inferences":["The self-supervised depth loop of Eq. (7) is the most fragile link; swapping the GNeRF's fine depth for an externally validated depth map, or masking the loss where the GNeRF depth is uncertain, could push gains beyond what the paper reports.","Because the new dataset renders blur by averaging frames along a camera trajectory, the degradation is depth-dependent; a direct comparison against image-space uniform blur on the same views would test whether the geometry-alignment mechanism is truly what carries the improvement.","The same aligned-feature design could extend to other structured degradations such as haze, rain, or low light by exchanging the restoration head, since the warping stage only requires poses and estimated depth.","If the module's benefit grows with view count, it could combine naturally with inference-time selection of additional source views, an extension the paper leaves unexplored."],"forward_implications":["The robustness gain transfers across GNeRF architectures, including attention-based, image-based, and cost-volume-based models, so the mechanism is not tied to one rendering formula.","Stronger degradation sees larger relative gains: noise experiments show improvements up to +0.63 PSNR at gain 20, adversarial evaluation gains reach +1.22 PSNR, and blur gains remain positive across all four blur levels.","Depth prediction becomes more stable: the absolute and relative deviation of predicted depth from the clean-image prediction drops by roughly 28 to 35 percent when the module is added under blur.","The module also improves rendering on clean source images, indicating that the 3D-aligned features help representation quality beyond robustness alone.","More source views amplify the benefit, which suggests the alignment mechanism increasingly exploits multi-view redundancy as the number of degraded inputs grows."],"supporting_citations":[{"why":"Supplies the noise-robust GNeRF baseline, the LLFF-N evaluation protocol, and the pose-aware denoising motivation.","marker":"[49]"},{"why":"Provides the cost-volume-based GeoNeRF baseline used in blur experiments.","marker":"[28]"},{"why":"Provides the attention-based GNT baseline used for adversarial robustness evaluation.","marker":"[53]"},{"why":"Establishes that depth perturbation is the main cause of GNeRF failure under adversarial degradation and supplies the adversarial training and testing scheme.","marker":"[16]"},{"why":"Supplies the multi-scale patch-match stereo approach used for the depth estimator.","marker":"[57]"},{"why":"Supplies the multi-Dconv head transposed attention blocks used in the restoration head.","marker":"[68]"},{"why":"Provides the convention of rendering multiple frames and averaging them to synthesize a blurry image.","marker":"[42]"},{"why":"Supplies the diverse 3D models used to construct the new blur dataset.","marker":"[12]"},{"why":"Justifies the requirement for 3D-consistent degradation in training data for real-world robustness.","marker":"[30]"}],"fun_headline_variants":["Depth-aware module hardens NeRF against blur and noise","Plug-in depth features stabilize NeRF on degraded views","Blur-resistant NeRF with depth-guided feature extraction","Degradation-robust NeRF via depth and denoising features"],"cache_read_input_tokens":21376,"weakest_assumption_plain":"The depth estimator is trained to imitate the same GNeRF depth prediction it helps produce, so if that prediction is already corrupted by degradation, the module may learn to lock in wrong geometry instead of correcting it.","fun_headline_variants_meta":{"raw":{"variants":["Depth-aware module hardens NeRF against blur and noise","Plug-in depth features stabilize NeRF on degraded views","Blur-resistant NeRF with depth-guided feature extraction","Degradation-robust NeRF via depth and denoising features"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00072,"raw_usage":{"total_tokens":3312,"prompt_tokens":1103,"completion_tokens":2209,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":719,"completion_tokens_details":{"reasoning_tokens":2141}},"tokens_in":719,"tokens_out":2209,"duration_ms":16772,"temperature":1.0,"reasoning_tokens":2141,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:14:06.273818+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the module with an independent depth signal, such as depths estimated from clean views of the same training scenes, instead of the GNeRF's fine depth as pseudo-ground truth. If the robustness gains on blurry or noisy test views disappear, or if the estimated depths drift from the independent depths, the self-referential depth loop is carrying the result.","supporting_citations":[{"cited_title":"NAN: noise- aware nerfs for burst-denoising","cited_arxiv_id":null,"evidence_quote":"Supplies the noise-robust GNeRF baseline, the LLFF-N evaluation protocol, and the pose-aware denoising motivation."},{"cited_title":"Geonerf: Generalizing nerf with geometry priors","cited_arxiv_id":null,"evidence_quote":"Provides the cost-volume-based GeoNeRF baseline used in blur experiments."},{"cited_title":"Is attention all that nerf needs? In Proceedings of ICLR, 2023","cited_arxiv_id":null,"evidence_quote":"Provides the attention-based GNT baseline used for adversarial robustness evaluation."},{"cited_title":"Nerfool: Uncovering the vulnerability of generalizable neural radiance fields against adversarial perturbations","cited_arxiv_id":null,"evidence_quote":"Establishes that depth perturbation is the main cause of GNeRF failure under adversarial degradation and supplies the adversarial training and testing scheme."},{"cited_title":"Patchmatchnet: Learned multi- view patchmatch stereo","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-scale patch-match stereo approach used for the depth estimator."},{"cited_title":"Restormer: Efficient transformer for high-resolution image restoration","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-Dconv head transposed attention blocks used in the restoration head."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the convention of rendering multiple frames and averaging them to synthesize a blurry image."},{"cited_title":"Objaverse: A universe of annotated 3d objects","cited_arxiv_id":null,"evidence_quote":"Supplies the diverse 3D models used to construct the new blur dataset."},{"cited_title":"3d common corruptions and data augmentation","cited_arxiv_id":null,"evidence_quote":"Justifies the requirement for 3D-consistent degradation in training data for real-world robustness."}],"review_version":1}