{"id":"203f63a8-af3f-48c1-8c80-0a94a00ab1db","arxiv_id":"2504.16389","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A self-supervised normalization and zero-event regularization scheme suppresses artifacts in event-based NeRF reconstruction, improving novel-view synthesis quality over EventNeRF and E-NeRF.","lead":"SaENeRF is a method for turning event-camera streams into 3D scene reconstructions with fewer visual artifacts. It adds two new training losses, a normalized event-loss and zero-event regularizers, that help neural radiance fields learn cleaner geometry and color from event data alone.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The training loss is invariant to global per-channel scaling of rendered radiance, so absolute color/brightness is unconstrained and reported PSNR may reflect initialization rather than the method.","rationale":"The reader identified Eq. (3), the accumulated-polarity event model, as the weakest assumption. That is a reasonable robustness concern, but it is not the most load-bearing issue for the paper's central quantitative claim. Even if Eq. (3) were exact and noise-free, the proposed normalization losses in Eqs. (8) and (10) are invariant to per-channel global scaling of rendered radiance, so absolute brightness, contrast, and color balance are not identifiable from the stated objective. Since the paper evaluates with PSNR/SSIM/LPIPS against target views and claims 'photorealistic' color reproduction, this invariance directly undermines the meaning of those numbers unless the authors specify an implicit anchor such as network initialization, activation bounds, or a separate calibration term. The concern is not about disagreement with community consensus; it is an internal property of the loss formulation. The paper does provide code, which makes the proposed test feasible, and the qualitative artifact-suppression comparisons may survive even if absolute color is not fully constrained. I therefore keep the reader's CONDITIONAL verdict, but the condition should include demonstrating that the reported metrics are stable under global color-scale perturbations or adding an explicit anchor to the loss. This is a partial agreement with the reader because both concerns target the supervision signal, but the scale-invariance problem is more fundamental and more specific to this paper's normalization contribution.","tokens_in":11962,"tokens_out":10543,"duration_ms":127411,"concrete_test":"Retrain the synthetic Lego and Chair sequences with the color MLP output multiplied by three different global per-channel factors, e.g., s=(1,1,1), s=(0.7,0.8,0.9), and s=(1.3,1.2,1.1), applied immediately after initialization and before computing ΔL_hat. Because Eqs. (8)–(10) are invariant to such scaling, the three runs should converge to the same training loss; if their final PSNR differs by more than 1 dB, the reported photometric accuracy is an artifact of initialization rather than a property of the loss. Alternatively, add a threshold-anchored term using C from Eq. (3), such as ||ΔL_hat(u) − E(u)C|| on positive pixels, and verify that the spread across the three runs vanishes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The composite loss is scale-invariant in a way that removes absolute photometry from the supervision. In Eq. (8), both predicted and event terms are divided by L1 norms over the same positive-sample set, so a global multiplicative factor on predicted log-intensity differences cancels. In Eq. (10), the positive zero-events loss is a ratio of L1 norms and is also scale-invariant; Eq. (9), when used, depends only on log-intensity differences. Since ΔL_hat(u) = log I(u,t) − log I(u,t0), applying a per-channel global scale to the rendered radiance field, I_c -> s_c I_c, leaves every ΔL_hat(u) unchanged and hence leaves the entire loss of Eqs. (8)–(12) unchanged. With λ0 = 0 for SaENeRF (Sec. III-E), the main model has no term that breaks this invariance at all; even with λ0 > 0, Lzero− penalizes differences, not absolute levels. The event threshold C from Eq. (3), which is the only quantity carrying physical contrast information, is explicitly divided out by the normalization in Eq. (8). Consequently, the objective has a continuous family of global per-channel gain solutions with identical training loss but different rendered colors. The paper reports PSNR/SSIM/LPIPS against target views, but those metrics are not invariant to this family. Therefore the quantitative evidence for 'photorealistic' reconstruction and the claimed color accuracy is not determined by the proposed losses; it depends on the arbitrary initialization and activation range of the color MLP. This is a more fundamental issue than the sensor-noise approximation in Eq. (3): even with noise-free events, the loss cannot select absolute color or contrast.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SaENeRF, a self-supervised method for training a Neural Radiance Field (NeRF) from event streams alone. The core ideas are a normalization loss (Eq. 8) that restricts the L1 normalization to pixels with non-zero accumulated event polarity, and two zero-events regularization losses (Eqs. 9-10) that suppress photometric changes in no-event regions while promoting contrast in event regions. The method is evaluated on the EventNeRF dataset for novel view synthesis, reporting PSNR/SSIM/LPIPS improvements over EventNeRF and an E-NeRF reimplementation, together with a real-time Instant-NGP variant. The code is publicly available.","tokens_in":12274,"tokens_out":8579,"duration_ms":84724,"significance":"The proposed losses are simple, clearly motivated, and easy to implement; the paper also ships code, which is a strength. If the identifiability concern discussed below were resolved, the method would constitute an incremental but potentially useful improvement for event-based NeRF training. However, the central quantitative claim of 'photorealistic' reconstruction is undermined by the scale-invariance of the loss, and the reported gains over the strongest baseline are small and lack statistical backing. As it stands, the evidence does not support the abstract's claim of 'significantly reduces artifacts'.","major_comments":[{"comment":"The loss function is invariant under a global per-channel multiplicative rescaling of the rendered radiance I_c -> s_c I_c, because every loss term depends only on log-intensity differences ΔL_hat(u) = log I(u,t) - log I(u,t0), and the normalization in Eqs. (8) and (10) divides by L1 norms that also scale. Consequently, the absolute brightness and color of the reconstruction are unconstrained by the event supervision. The reported PSNR/SSIM/LPIPS values against target views therefore do not measure the method's ability to recover photometry; they reflect the MLP's initialization, bias, and activation range. This directly affects the central claim of 'photorealistic' reconstruction in the Abstract and Sec. I, and it limits the interpretation of Tables I and II. Please (i) acknowledge this identifiability issue explicitly, (ii) demonstrate that the metrics are stable under changes to the color MLP's initialization/biasing (e.g., multiple seeds, adding a constant offset to the output color, or including a scale parameter), or (iii) add a constraint that fixes the absolute scale (e.g., a gray-world prior or a penalty on the mean rendered color). Without such evidence, the quantitative comparison is not well grounded.","section":"Sec. III-C/D, Eqs. (8)-(12)"},{"comment":"The claim that SaENeRF 'significantly reduces artifacts and achieves superior reconstruction quality' is not supported by the reported statistics. The comparison is based on a single run per method with no error bars or multiple seeds. The average PSNR gain over E-NeRF† is only 0.24 dB (30.39 vs 30.15), and SaENeRF is worse than E-NeRF† on three of the seven synthetic scenes (Lego, Ficus, Materials). The ablation study in Table III is performed on a single scene (Drums, misspelled as 'Drams' in Sec. IV-C), and the difference between norm+&zero+ and norm-&zero+ is 0.02 dB PSNR, which is well within typical run-to-run variation for NeRF training. I recommend reporting multiple seeds with standard deviations, or at least softening the language from 'significantly' to 'marginally' or 'on the tested sequence'.","section":"Sec. IV-A, Tables I and II"}],"minor_comments":[{"comment":"There are typos: 'changes in logarithmic brightness changes' should be 'changes in logarithmic brightness'; 'threshould' in Sec. I should be 'threshold'.","section":"Abstract and Sec. I"},{"comment":"The equation 'pC = log I(u,t) - log I(u,tprev)' is confusing because p is defined as the polarity in {+1,-1}; it would be clearer to write 'ΔL = pC' or 'each event contributes pC to the log-intensity difference'.","section":"Sec. III-B, Eq. (2)"},{"comment":"λ0 is set to 0 for SaENeRF, so the main model does not use Lzero-; the paper should justify this choice and state whether the results in Table I rely on Lzero+ alone. A sensitivity analysis for λ and λ0 would strengthen the claims.","section":"Sec. III-E"},{"comment":"The comparison with EventNeRF NGP uses different negative sampling ratios (0.1 for EventNeRF NGP vs. 0.05 for the other methods), which the text acknowledges but still constitutes a confound. Please at least run EventNeRF NGP with the same ratio for a fairer comparison.","section":"Sec. IV-A"},{"comment":"The paper does not discuss the case where the positive-sample set is empty (i.e., no non-zero accumulated polarities in an event window), which would make the denominators in Eqs. (8) and (10) zero. Please clarify how this is handled in practice.","section":"Sec. III-C/D, Eqs. (8) and (10)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal, but the contribution is incremental relative to E-NeRF and EventNeRF. The scale-invariance issue is fundamental and should be resolved or explicitly acknowledged before publication. The reviewer suggests that the editor ask for a revised version that addresses the identifiability problem and provides more robust statistical evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a modest but sensible contribution to event-based NeRF. It adds two loss terms to the EventNeRF/E-NeRF training scheme: a positive-only normalized L1 loss (Eq. 8) and a ratio-based zero-event loss (Eq. 10). These are new—I don't see them in EventNeRF or E-NeRF—and they address a real issue: balancing artifact suppression against contrast loss. The ablations support the design choices, and the code is available. That's the good news.\n\nThe soft spots are more than cosmetic. The most serious is a scale-invariance problem in the loss. Eqs. (8)–(10) are all built from log-light differences, so multiplying the rendered radiance by a global per-channel factor changes nothing in the loss. The only term that could break this, Lzero− (Eq. 9), is set to zero for SaENeRF (λ0=0). That means the objective cannot distinguish between reconstructions that differ by a global per-channel scale. The reported PSNR, SSIM, and LPIPS numbers are not invariant to this, so they depend partly on the MLP initialization and output activation range, not just on the proposed losses. This is not a minor sensor-noise issue; it is a degeneracy in the training objective. The paper never acknowledges it, yet the claims of 'photorealistic' and 'accurate color reproduction' rest on it.\n\nThe experiments are also thinner than the prose suggests. One dataset, no error bars or multiple seeds, only two quantitative event baselines (one reimplemented in a different framework), and an average PSNR gain of 0.24 dB over E-NeRF. The word 'significantly' is doing too much work. That said, the framework is sound in its main elements: the normalization idea is plausible, the equations are consistent, and the qualitative results do show cleaner backgrounds and edges on the scenes tested. The limitations section even notes the known-pose and static-scene restrictions, which is honest.\n\nThis is an incremental paper that yields some insight into loss design for event-based NeRF. It deserves a serious referee—not because it's transformative, but because the method is clearly specified, reproducible, and the artifact-suppression problem is real. But the review should demand: multiple seeds with error bars, better baselines (preferably official code), and a direct treatment of the scale ambiguity—either a color calibration step or a clear statement that only relative radiance is recovered. The authors should also soften the 'significantly' and 'photorealistic' language.\n\nIf I were editor, I'd send it out, but prepare the reviewers for a heavy revision. A workshop paper at this stage; a conference paper after the scale issue is addressed.","headline":"A modest, clearly specified loss-design contribution for event-based NeRF that deserves a real referee, but the objective has a scale-invariance gap the paper never acknowledges and the empirical support is thin.","tokens_in":12864,"tokens_out":3349,"would_cite":true,"duration_ms":33329,"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":"Two new losses let event streams alone train clean, photorealistic neural radiance fields.","keywords":["event camera","neural radiance fields","event-based 3D reconstruction","novel view synthesis","self-supervised learning","artifact suppression","volumetric rendering"],"falsifier":"Record a static scene with a real event camera and a synchronized high-frame-rate intensity camera, then compare per-pixel accumulated event polarity times the fixed threshold against the actual log-intensity difference over the same window. If the mismatch is large precisely where SaENeRF still renders artifacts, or if training on synthetic events that satisfy Eq. (3) exactly removes all artifacts while training on real events does not, the paper's linear event model is the simplification that fails.","tokens_in":11756,"feed_emoji":"🎥","tokens_out":8285,"duration_ms":75208,"temperature":0.7,"pith_summary":"This paper claims that the artifacts in event-based NeRF reconstruction come from two fixable sources: early aggressive learning of brightness changes before geometry is settled, and noise at pixels that trigger no events. SaENeRF replaces the raw event-supervision loss with a positive-only normalization between predicted radiance changes and accumulated event polarities, plus two zero-event regularization terms that suppress artifacts without flattening contrast. If correct, a static scene can be reconstructed from an event camera alone, without RGB frames, at quality close to frame-based methods, with cleaner backgrounds and sharper edges than EventNeRF and E-NeRF on the reported dataset. This matters because event cameras offer low latency, high dynamic range, and low power, and this is a step toward using them on their own for photorealistic 3D reconstruction.","feed_headline":"Event-camera 3D views cleaned by two loss tweaks","feed_subtitle":"New normalization and zero-event penalties let a NeRF train from events alone with sharper edges and fewer artifacts.","key_machinery":"The load-bearing mechanism is the loss family built on the Event Generation Model, where accumulated event polarities over a time window, scaled by a fixed threshold, act as the supervision signal for rendered brightness change. The first component is positive-only normalization: both the predicted radiance difference and the accumulated polarity are divided by their magnitudes computed only over pixels whose predicted sign matches the event polarity, so zero-event pixels cannot inject noise into the normalization and geometry is learned before large brightness variations. The second component is zero-event regularization: one term penalizes predicted brightness change at zero-event pixels, and a second term divides that penalty by the positive-pixel variation so artifact suppression does not wash out image contrast. Two monitoring quantities, the Temporal Average of Predicted Event Thresholds and the Proportion of Appropriate Pixels, select the consistent pixels used in the normalized loss.","core_discovery":"The paper's central claim is that event-only NeRF training can be made artifact-resistant by reshaping the supervision signal. Instead of matching raw rendered brightness changes to raw accumulated event polarity, SaENeRF normalizes both quantities by their magnitude computed only over positive, photometrically consistent pixels, then adds penalties that drive predicted brightness change toward zero on zero-event pixels while dividing that penalty by the variation on positive pixels to preserve contrast. The authors report that this removes the aggressive early brightness learning that distorts geometry and suppresses the background artifacts caused by event noise, yielding higher PSNR and SSIM and lower LPIPS than EventNeRF and E-NeRF on synthetic scenes and visibly cleaner reconstructions on real low-light sequences.","pith_inferences":["The linear event model in Eq. (3) is the obvious stress point; if per-event threshold noise matters, a natural extension is to estimate a per-pixel or time-varying threshold (the paper's Temporal Average of Predicted Event Thresholds could seed such an estimate) instead of treating the threshold as a fixed constant.","The same normalization and zero-event penalties should transfer to other event-supervised inverse rendering losses, such as event-based 3D Gaussian Splatting or event-based SLAM, since those also supervise from accumulated polarities and should inherit the artifact suppression.","A stronger test than the reported low-light static scenes would be fast camera motion or high scene contrast, where longer windows accumulate more polarity and the linear approximation in Eq. (3) is most strained; artifacts in those regimes would show where the method's premise ends.","A confidence-weighted version of the positive-only normalization, replacing the hard sign-consistency mask, might recover signal from small noisy polarities that the current selection discards."],"forward_implications":["If the paper is right, event-only NeRF training no longer needs RGB frames to suppress background artifacts, so the low latency and high dynamic range of event cameras can be used directly for photorealistic novel view synthesis.","The reported real-time variant converges in about one minute on a single 3090 GPU, suggesting event-based scene reconstruction could become an interactive capture-and-model tool rather than an offline process.","Because the losses work at a low zero-event sampling ratio, they reduce artifact counts and sampling cost at the same time.","The normalization changes the order of learning: geometry converges before large brightness variations are fitted, which is the mechanism the paper credits for suppressing early-training artifacts."],"supporting_citations":[{"why":"Supplies the Event Generation Model that connects event polarity to log-brightness change.","marker":"[1]"},{"why":"Baseline E2VID+NeRF pipeline that the paper compares against.","marker":"[8]"},{"why":"Base NeRF scene representation whose rendering equation and MLP are reused.","marker":"[10]"},{"why":"Provides the dataset, the EventNeRF baseline, and the negative-sampling strategy the paper builds on.","marker":"[11]"},{"why":"E-NeRF baseline and the normalization idea the paper refines with positive-only normalization.","marker":"[12]"},{"why":"Additional source of the Event Generation Model linking events to brightness changes.","marker":"[15]"},{"why":"Instant-NGP implementation basis for the real-time SaENeRF NGP variant.","marker":"[34]"}],"fun_headline_variants":["Event-only NeRF training cleans up with smart loss","New loss scheme fixes artifacts in event NeRFs","SaENeRF: sharpening event-camera 3D without clean frames","Normalized event losses banish NeRF artifacts","Event NeRF artifact fix: normalize, penalize, repeat"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole training signal assumes that each pixel's accumulated event polarity times one fixed threshold equals the true log-brightness change, while real event sensors add per-event threshold noise, refractory behavior, and asynchronous triggering.","fun_headline_variants_meta":{"raw":{"variants":["Event-only NeRF training cleans up with smart loss","New loss scheme fixes artifacts in event NeRFs","SaENeRF: sharpening event-camera 3D without clean frames","Normalized event losses banish NeRF artifacts","Event NeRF artifact fix: normalize, penalize, repeat"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000236,"raw_usage":{"total_tokens":1495,"prompt_tokens":925,"completion_tokens":570,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":487}},"tokens_in":541,"tokens_out":570,"duration_ms":5145,"temperature":1.0,"reasoning_tokens":487,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:03:46.969388+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a static scene with a real event camera and a synchronized high-frame-rate intensity camera, then compare per-pixel accumulated event polarity times the fixed threshold against the actual log-intensity difference over the same window. If the mismatch is large precisely where SaENeRF still renders artifacts, or if training on synthetic events that satisfy Eq. (3) exactly removes all artifacts while training on real events does not, the paper's linear event model is the simplification that fails.","supporting_citations":[{"cited_title":"High speed and high dynamic range video with an event camera,","cited_arxiv_id":null,"evidence_quote":"Baseline E2VID+NeRF pipeline that the paper compares against."},{"cited_title":"Eventnerf: Neural radiance fields from a single colour event camera,","cited_arxiv_id":null,"evidence_quote":"Provides the dataset, the EventNeRF baseline, and the negative-sampling strategy the paper builds on."},{"cited_title":"E-nerf: Neural radiance fields from a moving event camera,","cited_arxiv_id":null,"evidence_quote":"E-NeRF baseline and the normalization idea the paper refines with positive-only normalization."},{"cited_title":"Event-aided direct sparse odometry,","cited_arxiv_id":null,"evidence_quote":"Additional source of the Event Generation Model linking events to brightness changes."}],"review_version":1}