{"id":"9c4e547a-da2e-4e48-b394-b27067d0a283","arxiv_id":"2504.13561","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"WeatherGen is a unified diffusion framework that generates diverse adverse-weather LiDAR data and improves 3D object detection in fog, snow, and rain.","lead":"WeatherGen generates synthetic LiDAR point clouds for rain, fog, and snow using a diffusion model whose Mamba backbone scans along LiDAR beam lines. If the results hold up, self-driving perception systems could train on cheap synthetic adverse-weather data instead of costly real-world collection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MDP is the load-bearing pretraining component, but Eq. 3 leaves r_w, Rn, and Me1/Me2 unspecified and defers settings to a missing appendix, so the source of the claimed fidelity gain cannot be checked from the submitted material.","rationale":"I agree with the reader that the map-based data producer is the least secure link in the argument. The pretraining corpus is not a peripheral detail; it is the entire basis for claiming that WeatherGen can learn diverse weather without enough real data. The ablation in Table 3 shows that MDP changes FPD by roughly 25 points, so if MDP's construction is arbitrary or its r_w values are tuned to the evaluation set, the quantitative advantage could be a property of hand-set simulator thresholds rather than of the learned spider mamba, latent feature aligner, or contrastive controller. I considered making the absence of error bars the primary concern, but that is a real yet more general issue affecting many tables, whereas the MDP underspecification is specific to the architecture's central pretraining mechanism and directly blocks reproducibility. I also considered the overstated claim that prior methods are single-weather, but that is a presentation issue and not a load-bearing technical flaw. A focused re-run of the MDP ablation with disclosed parameter values and confidence intervals is the minimal experiment that would convert the current CONDITIONAL assessment into a fully supported one.","tokens_in":14647,"tokens_out":14065,"duration_ms":134773,"concrete_test":"Obtain the code from the provided repository and (1) extract the actual r_w values, the Rn sampling code, and the Me1/Me2 definitions from the MDP implementation, then report them; (2) run the Table 3 MDP on/off ablation on the snow-heavy split with at least 5 random seeds, computing FPD/FRD/MMD/JSD with bootstrap 95% confidence intervals under both conditions. If the no-MDP condition's FPD interval overlaps the full model's interval, then MDP is not load-bearing and the missing numbers are a reporting issue. If the full model is significantly better, and the r_w values are narrow or appear tuned to the Seeing Through Fog test splits, then the missing settings are critical and must be disclosed before the fidelity claim can be evaluated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 3.2's MDP is the foundation of the pretraining strategy: it is the only source of abundant diverse-weather data, and the ablation in Table 3 shows that removing MDP changes FPD from 59.28 to 84.58 and FRD from 1241.66 to 1501.79. However, Eq. 3 is not fully specified: the masks Me1 and Me2 are never defined (only a single Me is described in the text), Rn is introduced as \"random noise related to r_w\" without a distribution or scale, and r_w is described as an adjustable parameter whose settings are deferred to an appendix that is absent from the submission. Since Md is said to be learned through Eq. 9, but the total loss is applied to the denoising objective, it is also unclear what supervision signal actually shapes Md during pretraining. Consequently, a reader cannot determine whether the reported fidelity advantage comes from the proposed SMG/LFA/CLC components, from the MDP's hand-set r_w values, or from an artifact of the missing implementation. The central claim of significantly improving fidelity therefore rests on a pretraining premise that is currently unverifiable from the submitted paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WeatherGen, a diffusion-based framework for generating LiDAR point clouds under multiple adverse weather conditions. It combines a map-based data producer (MDP) for pretraining, a Spider Mamba generator (SMG) that scans range-map rows and columns, a latent feature aligner (LFA) that transfers real-world knowledge, and a contrastive learning-based controller (CLC) for weather control. Experiments on KITTI-360 and Seeing Through Fog report improved fidelity metrics over prior simulators and generative models, and a constructed mini-weather dataset is shown to improve PointPillars detection under fog, snow, and rain.","tokens_in":14840,"tokens_out":4514,"duration_ms":38176,"significance":"If the fidelity claims hold, WeatherGen would be a practical path to scaling adverse-weather LiDAR data without costly collection, and the beam/ray scanning design of the SMG is well motivated by the LiDAR imaging process. The paper ships code, includes ablations for each component, and evaluates downstream detection utility. However, the central fidelity comparison rests on under-specified pretraining details, small-sample point estimates, and unreplicated detector results, so the significance is conditional on those issues being resolved.","major_comments":[{"comment":"The MDP is a load-bearing component, as Table 3 shows that removing it raises FPD from 59.28 to 84.58, yet Eq. (3) is not fully specified: the masks Me1 and Me2 appear in the equation but are never separately defined, Rn is introduced only as 'random noise related to r_w' with no distribution or scale, and the adjustable parameter r_w is deferred to an appendix that is absent from the submitted version. Without these settings the pretraining premise cannot be audited, and the source of the reported fidelity gain cannot be attributed to the proposed learned components.","section":"Section 3.2, Eq. (3)"},{"comment":"The text states that Md is a mask 'learned through Equation 9,' but Eq. (9) defines the total loss L = LSMG + LLFA + LCLC, and none of these terms provides an explicit supervision signal for Md. Unless a separate objective for Md exists and is described, the claim that MDP 'adaptively align[s] with real-world distribution' is not supported by the stated loss.","section":"Section 3.2, Eq. (9)"},{"comment":"Weather-conditional generation is evaluated with only 200 generated samples per condition, and Tables 1–3 report single point estimates for FPD, FRD, MMD, and JSD without error bars or repeated sampling. Since diffusion sampling is stochastic, the reported margins, such as WeatherGen(Snow) FPD 59.28 versus LSS(Snow) 106.37, cannot be judged as significant without variance information; the authors should provide standard deviations or confidence intervals over multiple generation runs.","section":"Section 4.3, Table 2"},{"comment":"The downstream detection improvements are small and unreplicated: for example, Table 4 shows +1.39 and +2.08 AP over the 'None' baseline in the 0–80 m and 0–30 m dense-fog rows, and Table 5 shows only +0.25 AP for the 50–80 m snow row. No variance over detector training is reported, so it is unclear whether these gains are beyond training stochasticity; at minimum, multiple seeded runs or a significance test are needed to support the mini-weather claim.","section":"Section 4.5, Tables 4–6"}],"minor_comments":[{"comment":"The projection in Eq. (4) is written as [F1_i W; ...; Fj_i W], but the superscript indexing of patches and the relationship among J, C, and the number of patches is not defined; please clarify the reshaping step.","section":"Section 3.3, Eq. (4)"},{"comment":"The label 'Per tubed' in Figure 2 appears to be a typo, and the flow arrows for CLC and LFA are not described in the caption, making the figure hard to follow.","section":"Figure 2"},{"comment":"In Table 3, the checkmark patterns for the first and last rows are visually inconsistent (e.g., rows with '✓ ✓ ✓- -' and '✓ ✓ ✓- ✓' are difficult to parse); please use explicit Yes/No columns or a clearer legend.","section":"Table 3"},{"comment":"The evaluation description says 10,000 generated samples are used for unconditional generation and 200 for weather-conditional generation, but it does not state whether the same real-world test set is used for both, nor how many real-world samples are used in the weather-conditional metrics; please specify the exact protocol.","section":"Section 4.1"},{"comment":"The percentages in Figure 8, such as '-7.38% FRD' and '+0.51%', are not defined in the text; please state the reference baseline for each percentage.","section":"Section 4.6, Figure 8"},{"comment":"The paper refers to 'the Appendix' multiple times (for r_w settings, visual structures, densification details), but no appendix is included in the arXiv v1 submission; this is a presentation issue that should be fixed by including the supplementary material.","section":"Appendix references"},{"comment":"The KL divergence in Eq. (6) is notationally ambiguous because the two distributions are defined on different latent variables (zreal and zgen); please specify how the alignment is computed across the two latent spaces.","section":"Section 3.4, Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The central concern is that the pretraining MDP, which is the key enabler of the claimed fidelity, is under-specified in the submitted manuscript and depends on a missing appendix. I do not see circularity: fidelity metrics are computed on held-out test splits, and the LFA uses KL alignment during training, which is a standard transfer mechanism. The 'first unified' novelty claim is reasonable given the absence of prior diverse-weather generative frameworks, though a comparison with a weather-conditioned variant of an existing diffusion LiDAR model would strengthen the positioning."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"WeatherGen is worth a serious look. The core novelty is real: it is the first learned, unified generator for snow, rain, and fog LiDAR point clouds, and the spider-mamba scanning along beam circles and central rays is a sensible adaptation of SSMs to range maps. The MDP pretraining, latent feature aligner, and CLIP-based controller are coherent components, and the ablation study gives each of them credit. The comparisons against physical simulators on Seeing Through Fog show consistent gains on FPD/FRD, and the downstream detection experiment with 7.4% generated data is a reasonable practical test.\n\nThat said, the paper has some soft spots that keep me from calling it fully supported. The biggest is reproducibility: Equation 3 leaves Me1, Me2, Rn, and r_w effectively unspecified, and the appendix where r_w settings are supposed to live is absent from the submission. Since the ablation in Table 3 shows MDP has a large effect, a reader cannot tell whether the fidelity gain comes from the proposed learning machinery or from hand-tuned simulator parameters. This is a fixable problem—the authors clearly have the details—but it is load-bearing for the central claim.\n\nThe other issues are more minor: no error bars anywhere, weather-conditioned metrics on only 200 generated samples, and a code URL with no verifiable contents. The related-work claim that previous methods simulate only a single weather is a slight overstatement, since LISA covers rain, snow, and fog—though as a fixed physical simulator rather than a learned unified model, so the novelty claim survives.\n\nOverall, the architecture is plausible, the experiments are extensive enough for a first pass, and the direction is useful to the autonomous-driving and LiDAR-perception communities. I would send this to peer review, with the clear expectation that the authors provide the missing appendix, specify MDP precisely, and add error bars. If they do that, the paper could be a solid contribution.","headline":"A genuinely new unified learned generator for adverse-weather LiDAR with a sensible spider-mamba scan, but the missing appendix and underspecified MDP equation make the reported gains hard to verify.","tokens_in":15385,"tokens_out":1705,"would_cite":true,"duration_ms":16101,"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":"WeatherGen generates fog, rain, and snow LiDAR point clouds in one diffusion framework, and the synthetic scans improve 3D detection when they replace part of real training data.","keywords":["LiDAR point cloud generation","adverse weather simulation","diffusion models","state space models","range map generation","3D object detection","controllable generation"],"falsifier":"Run the same pre-train-then-fine-tune recipe with $r_w$ varied across a wide range, for example one-tenth and ten times the values used in the paper, while changing nothing else. If the distributional metrics and the dense-fog and heavy-snow detection AP stay roughly constant, the map-based producer's specific threshold is not the source of WeatherGen's gain; if the metrics swing sharply, the producer is load-bearing.","tokens_in":14429,"feed_emoji":"🌨️","tokens_out":8456,"duration_ms":78169,"temperature":0.7,"pith_summary":"WeatherGen sets out to establish that diverse-weather LiDAR data does not require separate physics simulators or costly real collection: one diffusion model, pre-trained on synthetic maps and fine-tuned on a small real set, can generate fog, rain, snow, and clear scans at usable fidelity. The paper's contribution is a full pipeline consisting of a learnable map-based data producer, a spider mamba generator that scans along LiDAR beam circles, a latent feature aligner, and a contrastive weather controller, with ablations claiming that each piece earns its keep. A reader should care because adverse-weather LiDAR is scarce and expensive; if the claim holds, replacing only 7.4% of a clear-weather training set with generated adverse-weather frames can measurably improve a 3D detector under fog, snow, and rain. The paper supports the claim with distributional metrics in three representations and with detection experiments on a real fog, rain, and snow LiDAR benchmark.","feed_headline":"One network now generates fog, rain, and snow LiDAR scans","feed_subtitle":"Synthetic adverse-weather scans replace part of real training data and lift 3D detection scores.","key_machinery":"The load-bearing object is the range map together with the spider mamba scan. LiDAR scans are projected to a range map whose rows are beam circles and whose columns are central rays; the spider mamba generator runs a state-space scan along those rows and columns at point level rather than over image patches, so long-range sensor structure is preserved during denoising. The other essential mechanism is the map-based data producer of Equation (3), which converts a clear range map into a weather map using a hand-set distance threshold $r_w$, a Bernoulli drop function, additive noise, and a learned mask, and is then frozen during fine-tuning. The latent feature aligner and the contrastive weather controller supply real-data alignment and weather-discriminative conditioning, and the total training loss is the sum of their three losses.","core_discovery":"WeatherGen's central claim is that LiDAR point-cloud generation under multiple adverse weathers should be treated as one learnable distribution rather than as separate physical models. Starting from a clear-weather range map, the map-based data producer turns that map into fog, rain, and snow range maps using a hand-set distance threshold, a learned mask, Bernoulli point dropout, and random noise, thereby supplying enough pre-training data. A diffusion backbone then denoises these maps: the spider mamba generator scans range-map rows and columns, which correspond to LiDAR beam circles and central rays, so the annular structure of the sensor is preserved; the latent feature aligner pushes the one-step denoised output toward real adverse-weather latents; and the contrastive weather controller ties the weather code to text embeddings of weather names so the four modes stay distinct. After fine-tuning on a small real dataset, the model produces scans whose distributional distances to real data are lower than those of single-weather simulators, and a 256-frame-per-weather mini-dataset improves a 3D detector under dense fog, heavy snow, and rain when it supplies 7.4% of the training data.","pith_inferences":["A direct comparison against a non-learnable version of the map-based producer, with fixed masks and no learned component, would show how much of the reported gain comes from learning rather than from the hand-set threshold and dropout; the reported ablations only remove the producer entirely.","The spider mamba's row-and-column scanning prior is not specific to generation and could be transplanted to range-map completion, densification, or semantic segmentation, where preserving beam circles also matters.","Because the weather control is anchored to language embeddings, adding new weather descriptions such as light drizzle or blizzard may require only new text prompts rather than new training data.","The mini-weather gain is measured with one detector on one fog, rain, and snow benchmark; checking whether the substitution effect transfers to other detectors and to mixed-weather training sets would settle how general the improvement is."],"forward_implications":["A single WeatherGen model can output clear, fog, rain, and snow scans from one set of weights, removing the need for per-weather physics simulators.","The pre-training-plus-fine-tuning recipe means a few hundred real adverse-weather frames can be enough to specialize a generator, because the map-based producer supplies the bulk of the training signal.","Replacing only 7.4% of a clear-weather training set with generated adverse-weather frames can raise detector performance on dense fog, heavy snow, and rain, making generated data a scalable augmentation channel.","The mini-weather dataset can be constructed in days rather than the years cited for real collection, making adverse-weather evaluation accessible without new field campaigns.","Because WeatherGen also runs as an unconditional clear-weather generator, the same architecture serves both standard LiDAR generation and weather-conditioned generation.","pith_inferences: These are the editor's own extensions, not the paper's claims."],"supporting_citations":[{"why":"Supplies the real fog, rain, and snow LiDAR data used for fine-tuning and as the test benchmark.","marker":"[1]"},{"why":"Provides the fog simulation baseline and the physics-inspired fog model that WeatherGen learns to surpass.","marker":"[9]"},{"why":"Provides the snowfall simulation baseline and the wet-ground snow model used for comparison.","marker":"[10]"},{"why":"Provides the rain, snow, and fog simulation baseline and the physics simulator compared in the detection tables.","marker":"[15]"},{"why":"Supplies the denoising diffusion probabilistic model formalism and training objective that the generator builds on.","marker":"[11]"},{"why":"Provides the range-map diffusion baseline whose U-Net is used as the ablation replacement when the spider mamba generator is removed.","marker":"[25]"},{"why":"Supplies the first LiDAR diffusion generation method and the Frechet range-map distance evaluation protocol.","marker":"[54]"},{"why":"Provides the text-guided LiDAR generation baseline and comparison for speed, parameters, and fidelity.","marker":"[46]"},{"why":"Supplies the frozen text encoder whose weather-name embeddings anchor the contrastive controller's semantic knowledge.","marker":"[32]"},{"why":"Supplies the vision mamba block design that the spider mamba adapts from image patches to LiDAR beam circles and central rays.","marker":"[52]"}],"fun_headline_variants":["Spider Mamba diffusion generates realistic fog, rain, and snow LiDAR","Unified diffusion model brings adverse weather to LiDAR training data","WeatherGen: one model, four weathers, realistic LiDAR scans","From clear to stormy: diffusion produces diverse LiDAR weather scenes","Synthetic adverse weather LiDAR data boosts 3D detection accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-set distance threshold $r_w$ in Equation (3), along with Bernoulli dropout and the learned mask, captures fog, rain, and snow well enough that pre-training on those maps helps rather than hurts; the paper defers the actual $r_w$ values to an appendix that is not present in the submitted text.","fun_headline_variants_meta":{"raw":{"variants":["Spider Mamba diffusion generates realistic fog, rain, and snow LiDAR","Unified diffusion model brings adverse weather to LiDAR training data","WeatherGen: one model, four weathers, realistic LiDAR scans","From clear to stormy: diffusion produces diverse LiDAR weather scenes","Synthetic adverse weather LiDAR data boosts 3D detection accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000316,"raw_usage":{"total_tokens":1841,"prompt_tokens":1052,"completion_tokens":789,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":696}},"tokens_in":668,"tokens_out":789,"duration_ms":6414,"temperature":1.0,"reasoning_tokens":696,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:05:23.801531+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same pre-train-then-fine-tune recipe with $r_w$ varied across a wide range, for example one-tenth and ten times the values used in the paper, while changing nothing else. If the distributional metrics and the dense-fog and heavy-snow detection AP stay roughly constant, the map-based producer's specific threshold is not the source of WeatherGen's gain; if the metrics swing sharply, the producer is load-bearing.","supporting_citations":[{"cited_title":"See- ing through fog without seeing fog: Deep multimodal sen- sor fusion in unseen adverse weather","cited_arxiv_id":null,"evidence_quote":"Supplies the real fog, rain, and snow LiDAR data used for fine-tuning and as the test benchmark."},{"cited_title":"Fog simulation on real lidar point clouds for 3d object detection in adverse weather","cited_arxiv_id":null,"evidence_quote":"Provides the fog simulation baseline and the physics-inspired fog model that WeatherGen learns to surpass."},{"cited_title":"Lidar snowfall simulation for robust 3d object detection","cited_arxiv_id":null,"evidence_quote":"Provides the snowfall simulation baseline and the wet-ground snow model used for comparison."},{"cited_title":"Learning to generate realistic lidar point clouds","cited_arxiv_id":null,"evidence_quote":"Supplies the first LiDAR diffusion generation method and the Frechet range-map distance evaluation protocol."},{"cited_title":"Text2lidar: Text-guided lidar point cloud generation via equirectangular transformer","cited_arxiv_id":null,"evidence_quote":"Provides the text-guided LiDAR generation baseline and comparison for speed, parameters, and fidelity."}],"review_version":1}