{"id":"72c2a6bd-5900-4a02-95b2-4627be5fa6fd","arxiv_id":"2412.18862","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"WeatherGS combines a diffusion-based weather filter and a lens occlusion detector with masked 3D Gaussian Splatting training to reconstruct clean scenes from snowy and rainy multi-view images.","lead":"WeatherGS cleans up 3D scene reconstructions taken in rain and snow by first removing weather particles and lens stains from the photos, then training 3D Gaussians on the cleaned images. If it works as claimed, autonomous driving and AR systems could get clean 3D models from cameras that currently see through bad weather.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim depends on AEF outputs being faithful, multi-view-consistent clean images; the paper never validates this, and SSIM regressions suggest structure loss.","rationale":"The reader's weakest assumption identified the AEF fidelity question, and I agree that this is the most load-bearing point. The reason it is load-bearing is structural: 3DGS optimizes a photometric loss against the preprocessed images, so any per-image error or cross-view inconsistency in the AEF output is directly baked into the learned Gaussians. The paper's own text admits possible structural detail loss, and the SSIM results corroborate that the preprocessed targets are not strictly faithful to the clean scene. I considered the alternative concern that lens-occlusion masks simply omit supervised pixels, but this is less fatal because camera motion means masked pixels typically correspond to different world rays across views, so 3DGS can still constrain those regions. The proposed test is cheap and decisive: it uses only the existing synthetic benchmark and checks the intermediate representation that the whole method depends on. This does not warrant rejecting the paper, since the synthetic PSNR and LPIPS gains are real evidence of practical value, but it does strengthen the need for the code, thresholds, and intermediate fidelity analysis already requested by the reader. The CONDITIONAL verdict remains appropriate.","tokens_in":10037,"tokens_out":8337,"duration_ms":89912,"concrete_test":"Using the synthetic benchmark's Blender scenes, compute the AEF preprocessed images i' for all training views and compare each directly to the corresponding ground-truth clean image. Report PSNR/SSIM/LPIPS separately in (1) regions affected by weather particles and (2) clean background regions that should be untouched. Also, for a set of SfM-verified 3D points visible in multiple views, measure the color/texture variance of their projections in i' across views against the same variance in the original clean images. If AEF does not improve fidelity in region (2), or if it increases cross-view variance substantially, then 3DGS is being trained on corrupted or inconsistent targets, and the clean-scene claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the AEF-produced images i' used as 3DGS training targets. For WeatherGS to recover the true clean scene, i' must both remove weather particles without altering scene geometry or appearance, and remain consistent across views, because 3DGS fuses multi-view observations into one radiance field. The paper only asserts that Diff-Plugin preserves fine details, then concedes in Section IV-C that the diffusion process 'may introduce minor impacts to structural details due to the diffusion process.' No experiment measures the fidelity of i' against the known clean ground truth on the synthetic benchmark, nor the cross-view consistency of i'. If AEF hallucinates texture or suppresses genuine scene content, those errors become the supervision signal for 3DGS: final renderings can look clean while being a smoothed or invented version of the scene. The SSIM regressions in Table I (WeatherGS 0.787 vs 3DGS 0.821 on snowy scenes; Factory-snow 0.700 vs 0.768) are consistent with structure loss in the preprocessed targets. The abstract's claim of 'accurately recover the underlying clear scene' therefore rests on an unverified property of the preprocessing stage.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WeatherGS, a 3D Gaussian Splatting framework for reconstructing clean 3D scenes from multi-view images captured under rainy or snowy conditions. The method decomposes weather artifacts into dense particles and sparse lens occlusions, removes particles with a diffusion-based Atmospheric Effect Filter (AEF) using Diff-Plugin task plugins, detects occlusion masks with a Lens Effect Detector (LED) based on the AttGAN detector, and trains 3DGS on the preprocessed images while masking out occluded regions. The authors introduce a synthetic benchmark derived from Deblur-NeRF scenes with ground truth, and report quantitative comparisons against NeRF, DerainNeRF, 3DGS, and GS-W, plus qualitative results on two real-world scenes. The main claims are that WeatherGS outperforms existing methods in PSNR and LPIPS on average and recovers the underlying clear scene.","tokens_in":10216,"tokens_out":7321,"duration_ms":59006,"significance":"If validated, WeatherGS would offer a practical and modular recipe for extending 3DGS to adverse weather by leveraging off-the-shelf 2D restoration and detection models. The synthetic benchmark with ground truth is a useful resource for future work, and the dense-to-sparse decomposition is a sensible design principle. The paper is not circular: the external pretrained components (Diff-Plugin, Stable Diffusion, AttGAN) are used as black boxes, and the only self-citation [28] is in related work, not in the method. However, the central claim that the AEF outputs faithfully preserve the underlying scene is not directly validated, and the SSIM regressions in the snowy scenes weaken the 'consistently outperforming' statement. The significance is therefore conditional on additional evidence about the fidelity and multi-view consistency of the preprocessed images.","major_comments":[{"comment":"The load-bearing step of the pipeline is the AEF-produced image i' used as the 3DGS training target; for the final radiance field to represent the true clear scene, i' must remove weather particles without altering scene geometry or appearance and must be consistent across views. The paper does not measure the fidelity of i' against the known clean ground truth available in the synthetic benchmark, nor does it measure cross-view consistency. Section IV-C concedes that integrating both components 'may introduce minor impacts to structural details due to the diffusion process,' and the SSIM results in Table I (e.g., Factory-snow: WeatherGS 0.700 vs 3DGS 0.768; average snowy: 0.787 vs 0.821) are consistent with structural loss in the preprocessed targets. Please add a direct evaluation of i' against the clean images (PSNR/SSIM/LPIPS) and a cross-view consistency metric, or the claim of 'accurately recover the underlying clear scene' is not supported.","section":"Section III-B, Section IV-C, Table I"},{"comment":"The abstract states that WeatherGS 'consistently produces high-quality, clean scenes across various weather scenarios, outperforming existing state-of-the-art methods,' but the quantitative results are more mixed. In Table I, WeatherGS has lower SSIM than vanilla 3DGS on all snowy scenes (average 0.787 vs 0.821), and in Table II it is lower on Pool-rain (0.805 vs 0.817) while higher on Tanabata-rain and Factory-rain. The explanation in Section IV-B that sparse weather artifacts 'do not significantly impact the SSIM metric' does not address the systematic SSIM drop on snowy scenes. The claim should either be tempered or supported by additional evidence (e.g., a perceptual user study, or a discussion of which metric matters for the target application).","section":"Tables I and II, Abstract"},{"comment":"The method has at least three free parameters that are not reported: the plugin-selection threshold θ in Eq. (4), the mask binarization threshold t in Eq. (5), and the loss balancing weight λ in Eq. (8). No sensitivity analysis or error bars are provided for any experiment. Since the final quality likely depends on t (which determines how much of the lens occlusion is masked) and λ, the absence of these values and of a robustness study prevents reproduction and makes it hard to assess whether the reported PSNR/LPIPS gains are stable.","section":"Section III-B/C, Eqns (4), (5), (8)"},{"comment":"The real-world evaluation is purely qualitative because ground truth is unavailable; this is an acceptable limitation, but it means the claim of generalization to 'various weather scenarios' is not quantitatively supported for real data. Please either state this limitation more explicitly in the conclusions or add a quantitative proxy such as a user study, no-reference metrics, or evaluation on a publicly available real dataset with known 3D scans.","section":"Section IV-A"}],"minor_comments":[{"comment":"Section III-A contains the typo 'Prelimineries' (should be 'Preliminaries'), and Section IV-B contains 'dense-to-spare' (should be 'dense-to-sparse').","section":"General"},{"comment":"The equations are poorly typeset: Eq. (2) shows '|N |X' instead of a summation, and Eq. (6) has garbled notation that appears to mix L1 and L2 norms. Please correct the mathematical formatting throughout.","section":"Eqns (2), (6)"},{"comment":"References [6] and [41] are both the original NeRF paper and should be merged; the reference list also contains incomplete URLs (e.g., [39], [40] are only 'Video link.' without working hyperlinks in the text).","section":"References"},{"comment":"In Table III, the full method (Ours) has lower SSIM (0.7395) than w/o AEF (0.7452); this is related to the major comment about AEF fidelity and should be discussed in the ablation section.","section":"Table III"},{"comment":"The description of the real-world dataset extraction is incomplete: the paper should state how many keyframes were used, how camera poses were obtained (e.g., SfM), and whether the scenes are static or contain dynamic content.","section":"Section IV-A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the idea is timely. The main concern is the missing validation of the preprocessing target fidelity, which I believe is fixable with additional experiments on the existing synthetic benchmark. I do not see research integrity issues. The related work section appropriately cites prior work, including the authors' own [28], which is not used in the method. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is a re-combination of existing pieces (Diff-Plugin AEF, AttGAN detector as LED, vanilla 3DGS) into one pipeline, and the paper is upfront about that. What's new is the dense-vs-sparse artifact categorization and the masked-loss integration, plus a small synthetic benchmark with ground truth that the field doesn't otherwise have. The average PSNR and LPIPS gains over vanilla 3DGS and GS-W are real on that benchmark. I believe the central claim is plausible, not demonstrated.\n\nThe clearest thing the paper does well is to separate dense airborne particles from lens occlusions and to treat them with different mechanisms. The qualitative figures show the failure modes of NeRF and 3DGS clearly, and Table IV shows a big efficiency win for the 3DGS backend. The authors also concede in IV-C that the diffusion step may damage structural details, which is the right instinct.\n\nThe soft spots are empirical, not logical. No code, no released benchmark data, and the three free parameters theta, t, lambda are never reported. There are no repeated trials or error bars. Real-world validation is purely qualitative, which matters because the benchmark is synthetic. More importantly, the paper never measures the fidelity of the AEF outputs against the known clean ground truth, nor their cross-view consistency. That is the load-bearing step: if the diffusion model hallucinates or erases structure, the 3DGS just learns to reconstruct the hallucination, and final renderings can look clean while being smoothed or invented. The SSIM regressions on snowy scenes (0.787 vs 0.821 for vanilla 3DGS; 0.700 vs 0.768 on Factory-snow) are consistent with exactly that worry. It may be a minor structural loss in a metric that favors retaining weather particles, as the authors argue, but they haven't shown it.\n\nThe self-citation to AllWeather-Net appears only in related work and is not load-bearing, so I wouldn't flag it. Nothing here is circular.\n\nWho is this for? Anyone building outdoor 3DGS pipelines under weather. It deserves a serious referee, not a desk reject. I'd ask for code and benchmark release, all thresholds, repeated trials, and a direct check of AEF fidelity against GT with a multi-view consistency metric. If those come back clean, this is a useful recipe and a reusable benchmark.\n\nMy recommendation: send it to review, conditional on the fidelity verification being added.","headline":"WeatherGS is a plausible, well-scoped recipe for rain/snow 3D reconstruction whose load-bearing assumption—diffusion-preprocessed images are faithful enough to supervise 3DGS—is never actually tested.","tokens_in":10821,"tokens_out":2429,"would_cite":true,"duration_ms":22515,"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":"The paper claims that preprocessing multi-view images by removing dense weather particles and masking lens occlusions lets 3D Gaussian Splatting reconstruct the underlying clean scene, and verifies this on new synthetic and real-world…","keywords":["3D Gaussian Splatting","adverse weather","rain removal","snow removal","lens occlusion","diffusion model","scene reconstruction","benchmark"],"falsifier":"Take a fixed scene, render it clean, then add synthetic snow or rain over it; run WeatherGS on the degraded views and compare the reconstruction to the clean render. If the restored images or reconstructed geometry are not closer to the clean ground truth than the unprocessed weather-corrupted input in non-occluded regions, the AEF is adding distortion rather than removing weather.","tokens_in":9762,"feed_emoji":"🌧️","tokens_out":7904,"duration_ms":64161,"temperature":0.7,"pith_summary":"The paper's central claim is that 3D Gaussian Splatting fails under rain and snow because it treats weather artifacts as part of the scene, and that a preprocessing stage can fix this without changing the rendering backend. WeatherGS sorts the artifacts into dense small particles (snowflakes, raindrops in the air) and sparse larger occlusions (precipitation on the lens), removes the first with a diffusion-based Atmospheric Effect Filter and detects the second with a Lens Effect Detector, then trains 3D Gaussians on the cleaned images while masking occluded pixels out of the loss. On a new benchmark of synthetic scenes and real-world footage, the paper reports that this pipeline beats vanilla NeRF, DerainNeRF, 3DGS, and GS-W on average PSNR and LPIPS, at a fraction of the training and rendering cost of a NeRF backend. If right, the contribution is a practical preprocessing recipe that makes real-time 3D reconstruction usable outdoors in bad weather.","feed_headline":"Rain and snow no longer break 3D Gaussian Splatting","feed_subtitle":"Dense-to-sparse preprocessing clears weather particles and lens drops so 3DGS reconstructs the clean scene.","key_machinery":"The dense-to-sparse preprocessing pipeline is the load-bearing mechanism: the Atmospheric Effect Filter (AEF) removes dense weather particles with a text-guided diffusion model, and then the Lens Effect Detector (LED) thresholds a confidence map to produce binary masks $M$. These masks enter the 3DGS training loss as a multiplication by $(1-M)$ in both the L1 term $L_1 = \\sum_t |\\hat{I}(t)-I(t)| \\circ (1-M)$ and the D-SSIM term, so pixels judged as lens occlusions do not supervise the Gaussians. The AEF also helps the LED, because removing the dense particles makes the sparser lens occlusions easier to localize. Everything else is standard 3DGS, including initialization from structure-from-motion points, $\\alpha$-compositing rendering, and the masked L1/D-SSIM objective.","core_discovery":"WeatherGS's discovery, on its own terms, is that the two weather artifact types have opposite statistics and should be treated asymmetrically. Dense weather particles are small but numerous and vary across views, so 3DGS currently averages them into blur or reconstructs them as solid content; lens occlusions are sparser but larger and are stable enough across views to be modeled as scene geometry. The proposed remedy is to remove the dense particles first with an Atmospheric Effect Filter (AEF), a text-guided diffusion model that selects a derain or desnow task plugin by comparing a text prompt with image features, and then to generate a per-image occlusion mask with a Lens Effect Detector (LED). Training 3DGS with the L1 and D-SSIM losses masked by $(1-M)$ makes the occluded lens regions irrelevant to the optimization, so the Gaussians model the clear scene. The paper reports that this consistently improves average PSNR and LPIPS over the baselines, while noting that SSIM sometimes favors vanilla 3DGS because the baselines train on the original, unprocessed images.","pith_inferences":["An implication the paper leaves implicit is that the text-driven plugin selection in the AEF generalizes beyond rain and snow: the same cosine-similarity switch could route fog, haze, dust, or spray to their own task plugins, making the pipeline a single framework for many weather types.","A testable extension is to validate against true clean geometry: capture or render the same scene both clear and weather-degraded, run WeatherGS on the degraded views, and compare the reconstructed 3D geometry (not just rendered images) to the clear capture; if the diffusion filter hallucinates detail, image metrics can look good while geometry drifts.","The masked-loss idea also suggests a pipeline-level improvement: instead of accepting the LED mask as fixed, the masks could be refined jointly with the 3DGS optimization, since occluded pixels are exactly those where the rendered clear scene disagrees with the input in a spatially coherent way."],"forward_implications":["WeatherGS makes 3DGS usable in outdoor rain and snow without architectural changes to the splatting itself, so existing 3DGS tools and pipelines can adopt weather robustness as a preprocessing step.","With a 3DGS backend the pipeline renders a frame in about 0.02 seconds after roughly 28 minutes of training on one RTX 3090, whereas the same preprocessing with a NeRF backend takes hundreds of minutes to train and over 30 seconds per render; the paper presents this as making real-time applications feasible.","The benchmark introduces three synthetic scenes (Tanabata, Factory, Pool) in snowy and rainy versions plus two real-world video-derived scenes, giving the community a common testbed for weather-degraded 3D reconstruction.","The ablation results imply that neither component alone suffices: the AEF alone leaves lens occlusions, the LED alone leaves dense particles, and the combination gives the best LPIPS."],"supporting_citations":[{"why":"Supplies the 3DGS backend and its training objective that WeatherGS builds on and modifies with masks.","marker":"[7]"},{"why":"Establishes the NeRF-based waterdrop-removal baseline and the lens-occlusion detection idea that LED adapts.","marker":"[5]"},{"why":"Provides the latent diffusion model that the AEF guides to remove dense weather particles.","marker":"[31]"},{"why":"Supplies the weather-specific task plugins and the text-image similarity mechanism used to select derain or desnow behavior.","marker":"[32]"},{"why":"Provides the raindrop dataset and detection module used to train the Lens Effect Detector.","marker":"[24]"},{"why":"Serves as the wild-image 3DGS baseline that WeatherGS compares against in quantitative and qualitative tests.","marker":"[12]"},{"why":"Provides the vanilla NeRF baseline that shows the blur artifacts WeatherGS avoids.","marker":"[6]"},{"why":"Supplies the base scenes from which the synthetic snowy and rainy benchmark images are generated.","marker":"[37]"}],"fun_headline_variants":["WeatherGS splits weather into particles and lens drops for clean 3D","Two-step preprocessing removes rain and lens drops in 3DGS","Dense-to-sparse masking clears weather from 3D scenes","WeatherGS: clear 3D reconstruction under adverse weather","Filter particles, mask lens drops: 3DGS works in rain and snow"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline hinges on the Atmospheric Effect Filter removing weather particles while leaving the underlying scene untouched, because if the diffusion model distorts or invents content, 3DGS is trained on corrupted targets and the 'clean' reconstruction is not the real scene.","fun_headline_variants_meta":{"raw":{"variants":["WeatherGS splits weather into particles and lens drops for clean 3D","Two-step preprocessing removes rain and lens drops in 3DGS","Dense-to-sparse masking clears weather from 3D scenes","WeatherGS: clear 3D reconstruction under adverse weather","Filter particles, mask lens drops: 3DGS works in rain and snow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000704,"raw_usage":{"total_tokens":3229,"prompt_tokens":1052,"completion_tokens":2177,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":2084}},"tokens_in":668,"tokens_out":2177,"duration_ms":14759,"temperature":1.0,"reasoning_tokens":2084,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:23:08.559708+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed scene, render it clean, then add synthetic snow or rain over it; run WeatherGS on the degraded views and compare the reconstruction to the clean render. If the restored images or reconstructed geometry are not closer to the clean ground truth than the unprocessed weather-corrupted input in non-occluded regions, the AEF is adding distortion rather than removing weather.","supporting_citations":[{"cited_title":"Derainnerf: 3d scene estimation with adhesive waterdrop removal,","cited_arxiv_id":null,"evidence_quote":"Establishes the NeRF-based waterdrop-removal baseline and the lens-occlusion detection idea that LED adapts."},{"cited_title":"Deblur-nerf: Neural radiance fields from blurry images,","cited_arxiv_id":null,"evidence_quote":"Supplies the base scenes from which the synthetic snowy and rainy benchmark images are generated."}],"review_version":1}