{"id":"076928be-44ab-409a-91a9-e51df5f8fda6","arxiv_id":"2505.20473","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Adding Gaussian noise to query coordinates during neural field training, with a decaying amplitude, acts as stochastic preconditioning that improves optimization quality across SDF fitting, surface reconstruction, and radiance fields.","lead":"Training neural fields with slightly randomized query positions, with the noise level annealed from large to zero, improves reconstruction quality across many tasks and representations. The technique is simple enough to add in a few lines of code and often matches or beats custom-designed coarse-to-fine hierarchies.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own blur-based mechanism is not what is optimized: under nonlinear losses the single-sample procedure minimizes a biased objective, and the Section 6 acknowledgment is the only treatment.","rationale":"The reader's weakest assumption correctly identifies the unsupported bridge between the blur identity and the actual training objective. I considered the limited hierarchy-replacement evidence and the per-task alpha tuning as alternative concerns; both are real, but the bias issue is more load-bearing because it touches the stated theoretical foundation and is explicitly conceded by the authors. The empirical results are extensive, consistent across tasks and representations, and credible enough to justify a conditional rather than rejecting verdict. The proposed experiment would settle whether the method works because of the claimed blur preconditioner or despite it, and would give the authors a concrete way to convert the Section 6 acknowledgment into analysis.","tokens_in":21409,"tokens_out":7464,"duration_ms":87176,"concrete_test":"On the hashgrid SDF-from-point-cloud task of Section 5.1.1 (Table 1), fix the alpha schedule and train four variants: baseline, single-sample SP, K=32-sample SP that averages f_theta(x+eps_i) before computing the loss (an approximation to the true blurred-field objective), and a control that adds the explicit variance penalty Var(f_theta(x+eps)) to the baseline loss with coefficient 1. If the K=32 variant retains the Chamfer-distance gains of single-sample SP, the bias is not load-bearing and the blur mechanism is approximately correct; if it does not, or if the variance-penalty control alone reproduces the gains, the claimed preconditioning mechanism is contradicted and the paper should be revised to describe the actual biased objective.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The motivating identity (2) holds for field values, but the actual losses in Sections 5.1-5.2 are nonlinear (L2, MAPE, rendering loss). For an L2 term, E_eps[(f_theta(x+eps)-y)^2] = (E[f_theta(x+eps)]-y)^2 + Var(f_theta(x+eps)), so the single-sample procedure is not a stochastic approximation to optimizing L(Blur[f_theta](x), y); it adds a variance penalty. The same structural bias applies to the rendering losses, compounded by volume rendering's nonlinear composition. Section 6 acknowledges this but explicitly defers analysis ('we do not observe this to cause problems in practice, but there is opportunity for deeper analysis'). Because the central claim is presented as 'querying the blurred field during optimization' (Section 4), the mechanism is load-bearing; the empirical gains could instead come from the variance penalty, from stochastic gradient noise, or from standard input-noise regularization (Bishop 1995), none of which require the blur interpretation. The single-vs-multiple-samples experiment (Section 5.3, Table 6) does not settle this: averaging f before the loss reduces but does not eliminate the bias, is run with K<=4 in one task only, and K=4 still leaves one quarter of the variance bias for quadratic losses.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes stochastic preconditioning (SP): during neural field training, query locations are perturbed by Gaussian noise x + eps, eps ~ N(0, alpha), with alpha annealed from an initial value to zero. The method is motivated by the identity Blur_alpha[f](x) = E[f(x + eps)] and is presented as an implicit, stochastic way to optimize through a blurred version of the field. The paper derives boundary handling via reflection, proposes a spatially varying alpha map optimized as an additional parameter, and reports experiments on SDF fitting, surface reconstruction from images, sparse-view NeRF, ReLU fields, and image fitting across several representations. The main claims are that SP improves quality and robustness, matches or outperforms custom hierarchies in several settings, and is simple to integrate, requiring only a few lines of code.","tokens_in":21651,"tokens_out":5415,"duration_ms":55279,"significance":"If the empirical gains hold, the paper describes a genuinely simple and broadly applicable technique that could be adopted across many neural field pipelines. The experimental coverage is extensive, with per-scene tables for DTU, Tanks and Temples, and synthetic NeRF scenes, and the method is tested on coordinate MLPs, hashgrids, triplanes, and ReLU fields. The paper is also admirably honest about its limitations, explicitly acknowledging the bias of single-sample nonlinear loss estimation in Section 6. However, the central theoretical framing is not currently matched to the objective actually optimized: the blur identity applies to field values, while the real losses are nonlinear in the field, so the single-sample procedure is a biased estimator of the blurred-field loss. The paper's own acknowledgment defers analysis, and the single-vs-multiple-samples experiment is too limited to resolve the question. Because the abstract and Section 4 attribute the gains to optimizing through the blurred field, this gap is load-bearing and should be addressed by either additional theoretical analysis or a careful reframing of the claims.","major_comments":[{"comment":"The central theoretical justification is not connected to the objective actually optimized. Equation (2) shows E[f_theta(x+epsilon)] = Blur_alpha[f_theta](x) for the field value, but the losses in Sections 5.1 and 5.2 are nonlinear in f_theta, including the SDF loss terms in Eq. (5) and the volumetric rendering loss in Section 5.2. For a quadratic loss, E_epsilon[(f_theta(x+epsilon)-y)^2] = (E[f_theta(x+epsilon)]-y)^2 + Var(f_theta(x+epsilon)), so the single-sample procedure minimizes a biased objective containing a variance penalty, not the blurred-field loss. The paper acknowledges this in Section 6 but only defers analysis ('we do not observe this to cause problems in practice, but there is opportunity for deeper analysis'). Because the abstract and Section 4 attribute the gains to 'querying the blurred field during optimization,' this gap is load-bearing: the empirical improvement could arise from variance regularization, standard input-noise regularization (Bishop 1995), or stochastic gradient noise. I ask for a theoretical analysis of the bias for the concrete losses used, and for experiments that target the mechanism directly, such as comparing single-sample training against training on a multi-sample estimate of E[f] with a matched cumulative query budget.","section":"Section 4, Eq. (2), Section 6"},{"comment":"The single-sample versus multiple-samples experiment does not resolve the bias concern. It is run only in the ReLU-field NeRF setting and only up to K=4 samples; for a quadratic loss, K=4 still leaves about one quarter of the variance term, so the comparison does not approach the true blurred-field objective. Moreover, averaging f before the loss reduces but does not eliminate the bias for nonlinear losses, and the reported PSNR differences (33.72, 33.74, 33.79) are small and likely within noise. A convincing control would use a larger K, or a closed-form estimate of E[f], and confirm whether the benefits of stochastic preconditioning persist when the variance bias is suppressed.","section":"Section 5.3, Table 6"},{"comment":"The main text states that stochastic preconditioning yields 'high quality results on-par with FreeNeRF' in the sparse-view NeRF setting, and the abstract claims matching or outperforming custom-designed hierarchies, but Supplement Table 5 shows that MipNeRF with stochastic preconditioning generally underperforms FreeNeRF on the reported scenes (e.g., 0.582 vs. 0.698 on scene 21, 0.699 vs. 0.746 on scene 55, 0.758 vs. 0.812 on scene 103). The supplement itself states 'Stochastic preconditioning improves upon the baseline but generally underperforms FreeNeRF.' The comparative claims in the main text and abstract should be revised to reflect the full results, and the settings in which the method matches, exceeds, or falls short should be stated explicitly.","section":"Section 5.2 vs. Supplement Table 5"}],"minor_comments":[{"comment":"The phrase 'See Table 2 and Table 3 for full tables in the supplemental document' is confusing because the main text already contains a Table 2; renumber or rename the supplement tables to avoid ambiguity.","section":"Section 5.1.2"},{"comment":"The text says 'Please see Section A.1.3 in main paper' but the discussion of fully-supervised SDF fitting is in the supplement; correct the cross-reference.","section":"Supplement A.1.3"},{"comment":"The label 'Ground Truth α-Map' is misleading because the optimized alpha map is not a ground-truth frequency map; consider using 'Optimized α-Map' instead.","section":"Figure 1"},{"comment":"The loss-landscape inset is presented as qualitative evidence of more isotropic level sets; this is fine as intuition but should be explicitly labeled as illustrative rather than a quantitative conditioning analysis.","section":"Section 3.2"},{"comment":"Figure 8 reports PSNR values in the figure but not in a table; adding a small table with per-scene PSNR/LPIPS numbers would improve reproducibility and readability.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is unusually transparent about its limitations, and the experimental effort is substantial. The main issue is the disconnect between the blur-based theoretical framing and the biased single-sample training objective; this is not a fatal flaw, but it should be resolved by analysis or by reframing the central claims. The discrepancy between the abstract/main-text claim of matching or outperforming custom hierarchies and the supplement's 'generally underperforms FreeNeRF' statement in the sparse-view experiment should be fixed. I see no reason to doubt the empirical effectiveness of the technique; the revision should focus on precision of claims and on either supplying a bias analysis or clearly positioning the method as an empirical regularizer."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is a clean, useful method paper. Input noise injection is classical, but the systematic application to spatial neural fields — with annealed scale, boundary reflection, and spatially varying alpha — is new, and the experiments are unusually thorough. The central empirical claim holds up: adding Gaussian query noise during training improves quality across MLPs, hashgrids, triplanes, and ReLU fields, and in several cases matches or replaces hand-built coarse-to-fine schemes. The per-scene tables and the hyperparameter robustness sweep give me real confidence that the effect is not a fluke of one configuration. The authors also deserve credit for shipping a transparent limitation section instead of burying the biased-estimator issue.\n\nThe soft spots are real but not fatal. The motivating identity E[f(x+eps)] = Blur[f](x) holds for field values, but the losses are nonlinear — L2, MAPE, rendering loss — so a single noisy sample is a biased estimator of the blurred-field loss, and the training objective is not actually \"optimizing through the blurred field.\" The paper acknowledges this in Section 6 but does not analyze it. The stress-test note is right that the empirical gains could come from the variance penalty or from ordinary input-noise regularization, not from the blur mechanism. That does not undermine the practical result, but it means the theoretical framing oversells the mechanism. I would ask the authors to either analyze the bias or soften the claim. Two smaller concerns: alpha schedules are tuned per task with no sensitivity study of alpha itself, and there is no variance across seeds for a stochastic method. The \"replace or outperform hierarchies\" headline rests on one ReLU-field experiment and a preliminary Neuralangelo study; it should be tempered.\n\nWho is this for? Anyone working on neural field optimization — graphics, vision, inverse problems. It is a strong candidate for serious peer review, not a desk reject. My recommendation: send it to referees, but ask them to focus on the bias analysis, seed variance, and whether the hierarchy-replacement claim is overbroad. If those are addressed, this is a solid SIGGRAPH/TOG-level contribution.","headline":"A simple, well-tested empirical trick for neural field optimization; the blur framing is loose but the contribution stands.","tokens_in":22185,"tokens_out":1216,"would_cite":true,"duration_ms":14632,"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":"Adding Gaussian noise to query coordinates makes neural field training behave like optimizing a blurred field, which avoids spurious minima and often beats hand-built coarse-to-fine schemes.","keywords":["neural fields","stochastic preconditioning","Gaussian blur","query perturbation","coarse-to-fine optimization","signed distance fields","neural radiance fields","hashgrids"],"falsifier":"Train the same field and loss on a 2D signal with a known spurious local minimum under three configurations: single-sample Gaussian query noise exactly as proposed, an exact blurred-field loss computed by numerical quadrature or an analytic blur network with the same alpha schedule, and no noise. If the exact blurred loss does not reproduce the improved convergence of the single-sample version, the explanation that stochastic preconditioning optimizes the blurred field is wrong; if no-noise training matches the noisy version, the benefit comes from something other than blur.","tokens_in":21165,"feed_emoji":"🎲","tokens_out":9327,"duration_ms":93242,"temperature":0.7,"pith_summary":"Neural field training is a hard nonlinear optimization problem whose spurious local minima produce artifacts. This paper proposes adding Gaussian noise to every query coordinate during training, evaluating the field at $x+\\varepsilon$ instead of $x$, and annealing the noise to zero over time. Because a Gaussian-blurred field is the expectation of the field at noisy locations, the training process effectively optimizes through a blurred field, suppressing high-frequency failures while the field is coarse, and the final model is an ordinary field with no extra machinery. The paper shows this 'stochastic preconditioning' improves quality on SDF fitting, surface reconstruction from images, and radiance fields across MLP, hashgrid, and triplane representations, and in several settings matches or improves on custom coarse-to-fine and hierarchy methods.","feed_headline":"Noisy queries replace custom hierarchies for neural fields","feed_subtitle":"Adding Gaussian offsets to query coordinates smooths the loss landscape, cutting artifacts in SDF and radiance field fitting.","key_machinery":"The load-bearing object is the identity $\\operatorname{Blur}_{\\alpha}[f](x) = \\mathbb{E}_{\\varepsilon}[f(x+\\varepsilon)]$ with $\\varepsilon \\sim \\mathcal{N}(0,\\alpha)$, which turns an intractable convolution over a neural field into a one-line sampling procedure. The method uses a single Monte Carlo sample per query so the computational cost is unchanged, and it adds two practical mechanisms: reflecting perturbed samples across the domain boundary to keep the sampling distribution uniform, and annealing $\\alpha$ from an initial scale down to zero so training finishes on the exact unblurred field. An optional extension stores $\\alpha(x)$ on a grid and optimizes it through the reparameterization trick, letting the noise scale itself become a learned spatially varying frequency map.","core_discovery":"The central claim is that stochastic evaluation is a preconditioner for neural field optimization. If $\\varepsilon \\sim \\mathcal{N}(0,\\alpha)$, then $\\operatorname{Blur}_{\\alpha}[f](x) = \\mathbb{E}_{\\varepsilon}[f(x+\\varepsilon)]$, so replacing every query $x$ by $x+\\varepsilon$ during training is, in expectation, querying a blurred low-pass version of the field. The optimizer therefore sees a smoother landscape with fewer spurious local minima, which the paper describes as a stochastic approximation to optimizing through the blurred field. The blur scale starts near two percent of the domain diagonal and is annealed to zero by the first third of training, boundary samples are reflected rather than clamped to preserve a uniform sampling distribution, and a single noisy sample per query is sufficient in the experiments. Across several representations and tasks the paper finds consistent quality gains, and an optional learned spatially varying noise field produces a frequency map of the target signal.","pith_inferences":["One testable extension the paper leaves open is to treat the initial blur scale as a per-region learned quantity from the start, turning the hand-annealed schedule into a fully automatic coarse-to-fine process.","Because the blur identity is dimension-agnostic, the same query-perturbation trick should apply to non-spatial inputs such as time in dynamic fields or view direction in appearance models, though the paper does not experiment with these.","The paper's finding that additional samples per query do not help suggests the benefit comes from stochastic gradient smoothing rather than from variance reduction; a direct comparison with deterministic blur would separate these mechanisms.","If the learned $\\alpha$ maps reflect local frequency content, they could be reused outside training, for example to set per-region hashgrid resolutions or to guide mesh extraction, though the paper only demonstrates sampling and feature unlocking."],"forward_implications":["Any queryable field, including coordinate MLPs, hashgrids, and triplanes, can obtain coarse-to-fine behavior with one added line of code, giving an immediate quality boost in settings without existing hierarchies.","Hashgrid SDF fitting with stochastic preconditioning reduces Chamfer error dramatically and can make geometric initialization unnecessary, although combining both still works well.","Removing the custom hierarchy from ReLU fields and using stochastic preconditioning instead produces comparable or better novel-view synthesis quality, at the cost of longer training time because all iterations run at full resolution.","Sparse-view NeRF training with stochastic preconditioning moves a standard MipNeRF baseline close to a purpose-built frequency-regularization method.","An optimized spatially varying noise field encodes a level-of-detail map of the target signal and can be used for adaptive sampling and hashgrid feature masking."],"supporting_citations":[{"why":"Supplies the classical result that input noise injection is equivalent to Tikhonov regularization, grounding the perturbation-as-regularizer interpretation the paper extends.","marker":"Bishop [1995]"},{"why":"Connects Gaussian input noise to heat-kernel diffusion, the theoretical link the paper uses to justify the blur interpretation.","marker":"Grandvalet et al. [1997]"},{"why":"Provides the hashgrid representation and the direct SDF fitting baseline that stochastic preconditioning most strongly improves.","marker":"Müller et al. [2022]"},{"why":"Defines ReLU fields and the four-stage hierarchy that the paper replaces with stochastic preconditioning.","marker":"Karnewar et al. [2022]"},{"why":"Defines the NeRF radiance field formulation used in the sparse-supervision experiments.","marker":"Mildenhall et al. [2021]"},{"why":"Provides the NeuS SDF-from-images baseline augmented with stochastic preconditioning in the DTU experiments.","marker":"Wang et al. [2021]"},{"why":"Provides Neuralangelo, the coarse-to-fine surface reconstruction method that stochastic preconditioning improves and can partially replace.","marker":"Li et al. [2023]"},{"why":"Provides MipNeRF, the baseline in the sparse-view NeRF experiment where stochastic preconditioning closes much of the gap to a specialized frequency-regularization approach.","marker":"Barron et al. [2021]"},{"why":"Defines the frequency-regularization method whose sparse-view results stochastic preconditioning approximates without representation-specific changes.","marker":"Yang et al. [2023]"}],"fun_headline_variants":["Stochastic queries precondition neural field optimization","Add Gaussian offsets to neural field queries for smoother training","Sampling offsets replace custom hierarchies in neural fields","Gaussian jitter preconditions neural field optimization","Blur in expectation: stochastic preconditioner for neural fields"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that using a single noisy field evaluation through a nonlinear loss, which is a biased estimate of the true blurred-field loss, preserves the optimization benefits of true blurring.","fun_headline_variants_meta":{"raw":{"variants":["Stochastic queries precondition neural field optimization","Add Gaussian offsets to neural field queries for smoother training","Sampling offsets replace custom hierarchies in neural fields","Gaussian jitter preconditions neural field optimization","Blur in expectation: stochastic preconditioner for neural fields"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000266,"raw_usage":{"total_tokens":1603,"prompt_tokens":928,"completion_tokens":675,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":603}},"tokens_in":544,"tokens_out":675,"duration_ms":6546,"temperature":1.0,"reasoning_tokens":603,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:54:38.416071+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same field and loss on a 2D signal with a known spurious local minimum under three configurations: single-sample Gaussian query noise exactly as proposed, an exact blurred-field loss computed by numerical quadrature or an analytic blur network with the same alpha schedule, and no noise. If the exact blurred loss does not reproduce the improved convergence of the single-sample version, the explanation that stochastic preconditioning optimizes the blurred field is wrong; if no-noise training matches the noisy version, the benefit comes from something other than blur.","supporting_citations":[],"review_version":1}