Pith. sign in

REVIEW 5 major objections 5 minor 48 references

WeatherCity claims a single 4D urban reconstruction can be re-rendered as rain, snow, or fog by splitting each Gaussian's appearance into a shared feature and per-weather color decoders, with physics-driven particles supplying the motion.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

WeatherCity turns a driving video into an editable 4D scene that can be re-rendered in consistent, controllable rain, snow, and fog with stable geometry.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection WeatherCity is a well-built system for multi-weather 4D editing, but the SOTA claim rests on weak evaluation and the fixed-geometry representation can't express some effects it claims. the 5 major comments →

arxiv 2602.22096 v2 pith:TSO6D6OF submitted 2026-02-25 cs.CV

WeatherCity: Urban Scene Reconstruction with Controllable Multi-Weather Transformation

Wenhua Wu , Huai Guan , Zhe Liu , Hesheng Wang This is my paper
classification cs.CV
keywords 4D scene reconstructionweather editing3D Gaussian Splattingautonomous driving simulationphysics-driven particlesmulti-weather consistencycontent consistency lossurban scene graph
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

WeatherCity aims to turn raw dash-cam-style video of an urban street into a single reconstructed 4D scene that can be re-rendered under rain, snow, or fog with the same cars, lanes, and buildings staying in place. It claims that the key is to split each Gaussian primitive's appearance into a shared feature vector and weather-specific color decoders, so geometry is identical across weathers while color changes. A text-guided image editor produces per-frame weather versions used as training supervision, and a consistency loss anchored to the original frames suppresses the editor's flicker and hallucinations. Dynamic rain and snow are added as physics-driven Gaussian particles, and fog as depth-dependent Beer–Lambert blending. If the claims hold, the result is a controllable, real-time weather simulator for autonomous-driving training and closed-loop testing that 2D editors and static 3D baselines do not provide.

Core claim

The paper's central discovery is that multi-weather 4D editing can be reduced to a disentangled appearance problem: instead of learning a separate scene per weather, every Gaussian carries one intrinsic feature f_i, and each weather condition has a small MLP that maps that feature to RGB. Geometry—position, scale, rotation, opacity—is shared and never changes across weathers. Training alternates between raw frames and text-edited weather frames, with an L1+SSIM RGB loss, a content-consistency loss anchored to the raw frames in a pre-trained feature space, and a LiDAR depth loss. In evaluation on two public driving benchmarks over rain, snow, and fog prompts, the method reports higher content

What carries the argument

The load-bearing object is the Weather Gaussian: each Gaussian stores a shared appearance feature (dimension 32) instead of a color, and weather-specific MLP decoders turn that feature into per-weather colors while position, scale, rotation, and opacity stay fixed. This makes structure identical across weathers by construction. Around it sit three supporting mechanisms: a content-consistency loss that pulls every rendered weather frame toward the raw frame in a pre-trained feature space, a dynamic Gaussian scene graph that allows object nodes to be edited or removed, and a physics layer that injects rain and snow particle Gaussians and applies depth-aware Beer–Lambert fog blending in the sam

Load-bearing premise

The bet is that weather acts only on appearance: every weather effect can be captured by repainting fixed scene geometry with per-weather colors plus overlaid particles and fog, so snow cover, rain sheen, or flood water never change the shape of the scene itself.

What would settle it

Take a parked car and a road region, reconstruct with the same video under light-snow and heavy-snow prompts, and compare the rendered depth maps and Gaussian centers on those static regions. If the heavy-snow reconstruction moves the road surface or car contours by more than LiDAR noise, the fixed-geometry assumption has failed and the claimed weather identity is enforced rather than learned.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • One captured drive can be re-rendered in multiple weathers without re-capturing, enabling closed-loop simulation and rare-weather training data for autonomous driving.
  • Weather intensity becomes a dial: particle counts, densities, and fog parameters are user-set, so light versus heavy rain, snow, and fog are explicit controls.
  • Because weather decoders are independent of geometry, switching weathers at render time does not require re-optimizing the scene, and the system reports near-real-time frame rates.
  • Object-level manipulation can coexist with weather editing, such as removing vehicles and snowing the scene in a single operation.
  • The per-frame inconsistencies of the underlying 2D editor are repaired by the consistency optimization rather than baked into the reconstructed scene.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension implied but not run: compare light- versus heavy-snow edits and inspect reconstructed Gaussian centers and rendered depth on static ground; if they shift, the fixed-geometry assumption is silently suppressing real accumulation, and snow or water would need per-weather geometry offsets.
  • The shared-feature scheme suggests a cheap interpolation route the paper does not explore: blend or condition the weather-specific decoder weights to synthesize intermediate weather intensities or unseen conditions between the trained decoders.
  • Because the consistency loss anchors every weather to the raw capture, systematic biases of the editing model could propagate into the decoders; anchoring to a canonical clear reconstruction instead would decouple weather appearance from one camera pass.
  • Since all weather particles are ordinary Gaussians in the same scene graph, the same representation could extend to other transient phenomena—leaves, dust, smoke—by swapping particle shapes and motion models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes WeatherCity, a framework for 4D urban scene reconstruction with controllable multi-weather editing. It builds on a dynamic Gaussian scene graph (following OmniRe) and introduces a Weather Gaussian representation: each Gaussian stores a shared appearance feature vector, and a per-weather MLP decoder maps that feature to a color for each target weather (rainy, snowy, foggy). Multi-weather edited images, obtained from a text-guided image editing model (Qwen-Image), serve as pseudo-ground-truth supervision alongside the original clear-weather frames. A content-consistency loss aligns rendered weather views with the raw scene in VGG feature space, and a physics-driven module adds dynamic rain/snow particles and Beer-Lambert-based fog. Experiments on Waymo and nuScenes compare against image/video editors (ControlNet, TurboEdit, FRESCO, Qwen-Image) and the static 3D editor ClimateNeRF, using CLIP-S, CLIP-DS, and Sem-CS metrics, plus runtime FPS. The paper claims state-of-the-art performance across all metrics, with additional demonstrations of object-level editing and temporal consistency.

Significance. If validated, WeatherCity would be a useful integration of 4D reconstruction and weather simulation for autonomous driving, combining a simple shared-feature/decoder decomposition with a unified particle-based weather rendering. The open-source release and the attempt to unify background appearance editing with foreground dynamic weather are valuable. However, the evaluation currently does not establish the central claim: metrics are limited to CLIP-based and segmentation-based scores on a small, hand-picked set of training-like sequences, with no reconstruction-quality metrics, no error bars, and no held-out-view generalization tests. The representational limitations (fixed geometry, view-independent color) also conflict with some of the qualitative claims (snow accumulation, wet-road specularity). The core idea is promising, but the evidence is insufficient to support the paper's strong SOTA statement.

major comments (5)
  1. [Sec. 4.1, Tab. 1] The quantitative evaluation is limited to five 30-frame scenes per dataset, with all metrics reported as point estimates with no error bars, standard deviations, or significance tests. Moreover, no reconstruction metrics (PSNR, SSIM, LPIPS) are provided for the raw-scene reconstruction, and there is no held-out-view split. The model is thus evaluated only on the same distribution it was trained on, and the 'state-of-the-art across all metrics' claim is not statistically grounded.
  2. [Sec. 3.2, Eq. (1)-(2)] The Weather Gaussian representation fixes geometry (μ, s, r, o) across weathers and decodes colors via φ_w(f_i), which has no view-direction input. This cannot represent view-dependent wet-road specular reflections or geometric snow accumulation, yet the paper claims these effects in Sec. 4.2 and Supp. Figs. 8-13. Because the optimization fits Qwen-Image edited frames that may imply such geometric/view-dependent changes, the model must compromise between L_rgb and L_cc, likely producing artifacts that CLIP-S/Sem-CS do not penalize. A concrete test: render from held-out views or with camera viewpoint changes and check whether reflections and snow cover behave physically.
  3. [Sec. 3.3, Eq. (6)-(7)] The content-consistency loss L_cc aligns rendered weather frames with the raw clear frame in VGG feature space. While the authors argue this repairs per-frame editing inconsistencies, it may also suppress the weather effect itself, yielding washed-out or under-edited results. The ablation (Tab. 3) only shows that L_cc improves CLIP-S and Sem-CS, which measure content preservation, not weather effect strength or realism. The paper does not provide any metric (e.g., FID, weather-specific fidelity, or user study) to confirm that the weather effect remains strong and realistic when L_cc is applied.
  4. [Tab. 2] The runtime comparison is not like-for-like. For baselines, the reported FPS is per-frame inference speed for editing; for WeatherCity, it is the rendering speed after training. The full WeatherCity pipeline includes a 2D editing step (Qwen-Image) and 30k training iterations, none of which are counted. Thus the statement that WeatherCity meets real-time requirements for simulation is misleading as an end-to-end performance claim.
  5. [Sec. 4.2 / Related Work] The set of 3D weather-editing baselines is incomplete and inconsistently evaluated. WeatherGS [27], RainyGS [8], and Fiebelman et al. [11] are cited in Related Work but never compared quantitatively. ClimateNeRF, the only 3D baseline included, is static, tested on a single scene in the Supplementary, and cannot handle dynamic objects. To substantiate the 'state-of-the-art' claim among 3D-level weather editing methods, the authors should include a modern dynamic 3D Gaussian weather baseline, even if only in the supplementary material.
minor comments (5)
  1. [Eq. (5)] 'di is the depth of the depth of the i-th Gaussian' — duplicated phrase.
  2. [Sec. 3.1] The text says Qwen-Image yields 'temporally consistent multi-weather image sequences,' but this is contradicted by the need for L_cc to fix per-frame inconsistencies (Sec. 4.3). The wording should be softened to 'per-frame edited sequences' or similar.
  3. [Fig. 2 caption] The caption says 'text-guided video editing foundation model,' but the method uses an image editing model (Qwen-Image). This is inconsistent with the main text.
  4. [Tab. 1 vs. Supp. Tabs. 5/6] It is unclear whether the aggregate CLIP-S/CLIP-DS/Sem-CS values in Tab. 1 are averaged over rainy/snowy/foggy conditions. The Supplementary shows per-weather breakdowns and notes that foggy results only use CLIP-DS. This should be clarified in the main text.
  5. [Sec. 3.2] The notation alternates between G_i = {μ_i, s_i, r_i, o_i, f_i} and G_i^w = {μ_i, Σ_i, o_i, c_i^w}. Define Σ_i explicitly in terms of s_i and r_i in the main text for clarity.

Circularity Check

0 steps flagged

No significant circularity: WeatherCity's weather decoders are trained, not derived, and no load-bearing result reduces to its inputs.

full rationale

The paper's central claim is empirical: a Weather Gaussian representation (Eq. 1–2) with shared features and weather-specific decoders is optimized by RGB and content losses (Eq. 6–7) against raw and Qwen-Image-edited frames, then compared with baselines. None of the reported metrics (CLIP-S, CLIP-DS, Sem-CS) is identical to a training loss or a fitted parameter. L_cc uses VGG features (Eq. 7), while CLIP-S uses CLIP embeddings and Sem-CS uses ConvNeXt segmentation; these are correlated content-preservation measures but not the same function, so there is no equation-level identity that makes a 'prediction' equal to a fit. The weather decoders are trained parameters, not derived predictions, and the paper does not claim to derive weather effects from first principles. Self-citations (EMIE-Map, BEV-GS) appear only in the Related Work overview and are not load-bearing for the proposed representation or the physics module; the cited OmniRe scene-graph structure is external. The physics-driven rain/snow/fog module is an explicit modeling choice (particle ellipsoids, Beer–Lambert transmittance), and the paper acknowledges manual tuning as a limitation, which is an assumption, not circularity. The main caveat—that L_cc's content-preservation objective aligns with the content-preservation metrics—is a metric/training correlation concern and a correctness risk, but it does not reduce the derivation to its inputs. Therefore no specific circular step can be quoted and exhibited, and the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 8 axioms · 2 invented entities

The central claim rests on one architectural hypothesis (weather = per-weather color decodes of shared features plus overlay particles/fog) and one supervisory assumption (Qwen-Image edits are good enough to train the decoders). Many simulation parameters are hand-chosen, and the paper itself admits manual tuning in the Limitations section. No new physical entities are introduced beyond simulated weather particles, whose realism is visually checkable.

free parameters (7)
  • Fog density d_f = 0.2
    Hand-set in §3.4/Supp. §6.1; controls transmittance exp(-d_f·depth) in Eq. 11. The Limitations section admits manual tuning of weather parameters.
  • Fog color c_fog = [0.8, 0.8, 0.85]
    Hand-chosen global fog tint in Eq. 11 blending.
  • Rain particle parameters (count, color, scale, opacity) = 40,000; [0.7,0.7,0.8]; [0.0025,0.0025,0.075]; 0.13
    Manually tuned to produce visible streaks (Supp. §6.1); part of the manual tuning the Limitations section flags.
  • Snow particle parameters (count, color, scale, opacity) = 16,000; [0.9,0.9,0.95]; [0.0064,0.004,0.004]; 0.2
    Manually tuned flake size, color, and visibility (Supp. §6.1).
  • Loss weights λ_cc, λ_depth, λ_opacity, λ_SSIM = 1.0, 0.01, 0.05, 0.2
    Balancing weights (§3.3, Supp. §6.2) that determine how strongly content consistency, geometry, and opacity are enforced.
  • Gaussian feature dimension / decoder width = d=32, hidden=64
    Architecture hyperparameters determining decoder capacity (§4.1, Supp. §6.1).
  • Weather decoder MLP weights = learned
    Fit to Qwen-Image edited images; each weather's appearance is trained against the generative model's outputs, so weather-appearance quality is bounded by Qwen-Image.
axioms (8)
  • standard math 3D Gaussian Splatting provides a differentiable, high-fidelity scene model (covariance projection and alpha blending)
    Adopted wholesale from Kerbl et al. [18]; Eqs. 3–5 rely on its projection, sorting, and blending.
  • domain assumption Urban dynamic scenes decompose into sky/static/rigid/non-rigid Gaussian nodes as in OmniRe [4]
    The scene graph G = {N_sky, N_bg, N_rigid, N_nonrigid} in §3.2 is taken from OmniRe; object-editing claims depend on this node structure.
  • domain assumption Qwen-Image edits are trustworthy enough to supervise weather appearance (content-preserving and weather-realistic)
    Edited sequences I^w_t are the only weather-appearance supervision (§3.1); the paper itself documents Qwen-Image per-frame inconsistencies (§4.3) that L_cc must repair.
  • ad hoc to paper Weather appearance is fully expressible as per-weather MLP decodings c^w_i = φ_w(f_i) of a shared feature, plus overlay particles/fog (no geometry-level weather changes)
    Core representational bet in Eq. 2; snow accumulation and wet reflections must be captured by color only, not by changing μ, s, r, o.
  • domain assumption VGG-19 relu4_1 features quantify cross-weather content consistency
    L_cc in Eq. 7 assumes perceptual feature distance to the raw frame enforces the right notion of scene fidelity.
  • domain assumption LiDAR projection gives correct sparse depth for geometry supervision
    L_depth in Eq. 8 treats projected LiDAR as ground truth for all scenes.
  • standard math Beer–Lambert blending models fog sufficiently (Eq. 11)
    Exponential transmittance attenuation by depth, with hand-set parameters.
  • domain assumption CLIP-S, CLIP-DS, and Sem-CS are adequate proxies for editing quality
    All state-of-the-art claims rest on these three metrics (§4.1, Supp. §6.4); no human study is reported.
invented entities (2)
  • Weather Gaussian representation (shared feature vector f_i + weather-specific decoders φ_w) no independent evidence
    purpose: Disentangle scene structure/texture from weather-dependent color so geometry stays consistent across weather conditions
    Architectural construct (§3.2, Eq. 2). No external falsifiable handle; supported only by the internal ablation (w/o WGS in Tab. 3).
  • Weather particle Gaussian nodes for rain and snow (elongated streaks; three-lobe crystal) independent evidence
    purpose: Render dynamic falling rain/snow integrated into the same rasterization pass
    Simulated entities with a falsifiable visual handle: rendered particles can be compared against real rain/snow footage, and their motion is governed by stated velocity equations (Eq. 10).

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of WeatherCity: Urban Scene Reconstruction with Controllable Multi-Weather Transformation." pith.science (2026). https://pith.science/paper/TSO6D6OF

@misc{pith2026260222096,
  author       = {Pith},
  title        = {Pith review of: WeatherCity: Urban Scene Reconstruction with Controllable Multi-Weather Transformation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TSO6D6OF}},
  note         = {Machine review of arXiv:2602.22096}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Editable high-fidelity 4D scenes are crucial for autonomous driving, as they can be applied to end-to-end training and closed-loop simulation. However, existing reconstruction methods are primarily limited to replicating observed scenes and lack the capability for diverse weather simulation. While image-level weather editing methods tend to introduce scene artifacts and offer poor controllability over the weather effects. To address these limitations, we propose \textbf{WeatherCity}, a novel framework for 4D urban scene reconstruction and weather editing. Specifically, we leverage a text-guided image editing model to achieve flexible editing of image weather backgrounds. To tackle the challenge of multi-weather modeling, we introduce a novel weather Gaussian representation based on shared scene features and dedicated weather-specific decoders. This representation is further enhanced with a content consistency optimization, ensuring coherent modeling across different weather conditions. Additionally, we design a physics-driven model that simulates dynamic weather effects through particles and motion patterns. Extensive experiments on multiple datasets and various scenes demonstrate that WeatherCity achieves flexible controllability, high fidelity, and temporal consistency in 4D reconstruction and weather editing. Our framework not only enables fine-grained control over weather conditions (e.g., light rain and heavy snow) but also supports object-level manipulation within the scene. Codes are released at https://github.com/IRMVLab/WeatherCity.

Figures

Figures reproduced from arXiv: 2602.22096 by Hesheng Wang, Huai Guan, Wenhua Wu, Zhe Liu.

Figure 1
Figure 1. Figure 1: We present WeatherCity, a novel framework for dynamic urban scene reconstruction and controllable weather editing. Given a sequence of raw images, WeatherCity seamlessly integrates 4D reconstruction with flexible weather manipulation, producing highly consistent, photorealistic, and versatile multi-weather rendering results. Abstract Editable high-fidelity 4D scenes are crucial for au￾tonomous driving, as … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of WeatherCity. Our framework comprises four main modules. First, the image editing module employs a text-guided [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative results of multi-weather editing on the Waymo Open Dataset. Our method produces realistic and consistent weather [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results of multi-weather editing on the nuScenes dataset. Our method produces realistic and consistent weather [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of object editing results. The text prompt [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of ablation study. of all methods in Tab 2. For image and video editing models, we report the average inference speed (in FPS). For WeatherCity, We provide the average rendering speed. WeatherCity achieves a rendering speed of 25.67 FPS, which is sufficient to meet the requirements for real-time simulation. 4.3. Ablation Study To validate the effectiveness of each core component in WeatherCit… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison of rainy weather on the Waymo Open Dataset. Our model excels at capturing complex lighting interactions and environmental changes induced by rain, providing a higher level of physical realism. Raw images ControlNet FRESCO TurboEdit WeatherCity (Ours) [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison of snowy weather on the Waymo Open Dataset. As shown, WeatherCity generates highly realistic snow accumulation on vehicles and roads while avoiding the texture distortions common in other approaches. where nc is the number of pixels belonging to class c in the ground-truth segmentation of the original image. We note that fog synthesis substantially reduces scene visibility, which con… view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison of foggy weather on the Waymo Open Dataset. WeatherCity demonstrates superior depth-consistent haze rendering, effectively preserving the semantic layout of the original scene compared to baseline methods. Raw images ControlNet FRESCO TurboEdit WeatherCity (Ours) [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparison of rainy weather on the nuScenes Dataset. Our method accurately simulates realistic wet surface reflections and rain streaks, significantly outperforming ControlNet and FRESCO in terms of visual fidelity. between the road and the environment, our method lever￾ages 3D scene representations to ensure that reflections are geometrically consistent with the camera view. Snowy: In snowy w… view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative comparison of snowy weather on the nuScenes Dataset. Our approach produces the most visually plausible winter scenes with natural white-out effects, surpassing the consistency and quality of competing editing frameworks. Raw images ControlNet FRESCO TurboEdit WeatherCity (Ours) [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative comparison of foggy weather on the nuScenes Dataset. WeatherCity achieves a natural atmospheric haze that smoothly obscures distant objects, exhibiting fewer artifacts than TurboEdit or other baselines. weather particles (such as falling snowflakes) and achieves significantly higher rendering efficiency than ClimateNeRF, conclusively validating its superiority. 6 [PITH_FULL_IMAGE:figures/full… view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative comparison of rainy weather on temporal consistency. The baselines suffer from scene deformation and erratic fluc [PITH_FULL_IMAGE:figures/full_fig_p017_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative comparison of snowy weather on temporal consistency. The baselines suffer from scene deformation and erratic [PITH_FULL_IMAGE:figures/full_fig_p017_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative comparison of foggy weather on temporal consistency. The baselines suffer from scene deformation and erratic [PITH_FULL_IMAGE:figures/full_fig_p018_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative comparison with ClimateNeRF on Waymo Open Dataset. ClimateNeRF is limited to static modeling and static [PITH_FULL_IMAGE:figures/full_fig_p019_17.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 8 linked inside Pith

  1. [1]

    In- structpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. In- structpix2pix: Learning to follow image editing instructions. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 18392–18402, 2023. 3

    2023

  2. [2]

    nuscenes: A multi- modal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020. 6

    2020

  3. [3]

    Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering.arXiv preprint arXiv:2311.18561, 2023

    Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering.arXiv preprint arXiv:2311.18561, 2023. 2

    Pith/arXiv arXiv 2023

  4. [4]

    Omnire: Omni urban scene reconstruction

    Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Goj- cic, Sanja Fidler, Marco Pavone, et al. Omnire: Omni urban scene reconstruction. InThe Thirteenth International Con- ference on Learning Representations. 2, 3, 1

  5. [5]

    Stargan v2: Diverse image synthesis for multiple domains

    Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 8188–8197, 2020. 3

    2020

  6. [6]

    Survey on 3d recon- struction techniques: Large-scale urban city reconstruction and requirements.IEEE Transactions on Visualization and Computer Graphics, 2025

    Andreas Christodoulides, Gary KL Tam, James Clarke, Richard Smith, Jon Horgan, Nicholas Micallef, Jeremy Mor- ley, Nelly Villamizar, and Sean Walton. Survey on 3d recon- struction techniques: Large-scale urban city reconstruction and requirements.IEEE Transactions on Visualization and Computer Graphics, 2025. 1

    2025

  7. [7]

    Streetsurfgs: Scalable ur- ban street surface reconstruction with planar-based gaussian splatting.IEEE Transactions on Circuits and Systems for Video Technology, 2025

    Xiao Cui, Weicai Ye, Yifan Wang, Guofeng Zhang, Wengang Zhou, Tong He, and Houqiang Li. Streetsurfgs: Scalable ur- ban street surface reconstruction with planar-based gaussian splatting.IEEE Transactions on Circuits and Systems for Video Technology, 2025. 2

    2025

  8. [8]

    Rainygs: Efficient rain synthe- sis with physically-based gaussian splatting

    Qiyu Dai, Xingyu Ni, Qianfan Shen, Wenzheng Chen, Bao- quan Chen, and Mengyu Chu. Rainygs: Efficient rain synthe- sis with physically-based gaussian splatting. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 16153–16162, 2025. 3

    2025

  9. [9]

    Turboedit: Text-based image editing using few-step diffusion models

    Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. InSIGGRAPH Asia 2024 Con- ference Papers, pages 1–12, 2024. 3, 7, 8

    2024

  10. [10]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. InForty-first international conference on machine learning,

  11. [11]

    Let it snow! animating static gaussian scenes with dynamic weather effects.arXiv preprint arXiv:2504.05296, 2025

    Gal Fiebelman, Hadar Averbuch-Elor, and Sagie Benaim. Let it snow! animating static gaussian scenes with dynamic weather effects.arXiv preprint arXiv:2504.05296, 2025. 3

    arXiv 2025

  12. [12]

    Stylegan-nada: Clip-guided domain adap- tation of image generators.CoRR, abs/2108.00946, 2021

    Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adap- tation of image generators.CoRR, abs/2108.00946, 2021. 6, 3

    Pith/arXiv arXiv 2021

  13. [13]

    Sgsst: Scaling gaussian splatting style transfer

    Bruno Galerne, Jianling Wang, Lara Raad, and Jean-Michel Morel. Sgsst: Scaling gaussian splatting style transfer. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 26535–26544, 2025. 3

    2025

  14. [14]

    Tokenflow: Consistent diffusion features for consistent video editing.arXiv preprint arXiv:2307.10373, 2023

    Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing.arXiv preprint arXiv:2307.10373, 2023. 3

    Pith/arXiv arXiv 2023

  15. [15]

    Generative adversarial networks.Commu- nications of the ACM, 63(11):139–144, 2020

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Commu- nications of the ACM, 63(11):139–144, 2020. 2

    2020

  16. [16]

    S2st: Image-to-image translation in the seed space of latent dif- fusion.arXiv preprint arXiv:2312.00116, 2023

    Or Greenberg, Eran Kishon, and Dani Lischinski. S2st: Image-to-image translation in the seed space of latent dif- fusion.arXiv preprint arXiv:2312.00116, 2023. 2

    Pith/arXiv arXiv 2023

  17. [17]

    Clipscore: A reference-free evaluation met- ric for image captioning, 2022

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation met- ric for image captioning, 2022. 6, 3

    2022

  18. [18]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1,

  19. [19]

    Weather trans- lation via weather-cue transferring.IEEE Transactions on Neural Networks and Learning Systems, 35(6):7988–7998,

    Xuelong Li, Chen Li, Kai Kou, and Bin Zhao. Weather trans- lation via weather-cue transferring.IEEE Transactions on Neural Networks and Learning Systems, 35(6):7988–7998,

  20. [20]

    Climatenerf: Extreme weather synthesis in neural radiance field

    Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Extreme weather synthesis in neural radiance field. InProceedings of the ieee/cvf inter- national conference on computer vision, pages 3227–3238,

  21. [21]

    De- rainnerf: 3d scene estimation with adhesive waterdrop re- moval

    Yunhao Li, Jing Wu, Lingzhe Zhao, and Peidong Liu. De- rainnerf: 3d scene estimation with adhesive waterdrop re- moval. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2787–2793. IEEE, 2024. 2

    2024

  22. [22]

    Stylegaussian: Instant 3d style transfer with gaussian splatting

    Kunhao Liu, Fangneng Zhan, Muyu Xu, Christian Theobalt, Ling Shao, and Shijian Lu. Stylegaussian: Instant 3d style transfer with gaussian splatting. InSIGGRAPH Asia 2024 Technical Communications, pages 1–4. 2024. 3

    2024

  23. [23]

    A convnet for the 2020s, 2022

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s, 2022. 7, 3

    2022

  24. [24]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021. 2

    2021

  25. [25]

    Neural scene graphs for dynamic scenes

    Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 2856–2865, 2021. 2

    2021

  26. [26]

    A survey of structure from motion*.Acta Numerica, 26:305–364, 2017

    Onur ¨Ozyes ¸il, Vladislav V oroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*.Acta Numerica, 26:305–364, 2017. 2 9

    2017

  27. [27]

    Weathergs: 3d scene reconstruction in adverse weather conditions via gaussian splatting

    Chenghao Qian, Yuhu Guo, Wenjing Li, and Gustav Markkula. Weathergs: 3d scene reconstruction in adverse weather conditions via gaussian splatting. In2025 IEEE In- ternational Conference on Robotics and Automation (ICRA), pages 185–191. IEEE, 2025. 2, 3

    2025

  28. [28]

    Climategan: Raising climate change awareness by generating images of floods

    Victor Schmidt, Alexandra Luccioni, M ´elisande Teng, Tianyu Zhang, Alexia Reynaud, Sunand Raghupathi, Gautier Cosne, Adrien Juraver, Vahe Vardanyan, Alex Hern ´andez- Garc´ıa, et al. Climategan: Raising climate change awareness by generating images of floods. InInternational Conference on Learning Representations. 3

  29. [29]

    Emu edit: Precise image editing via recognition and gen- eration tasks

    Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and gen- eration tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871– 8879, 2024. 3

    2024

  30. [30]

    Very deep convo- lutional networks for large-scale image recognition.CoRR, abs/1409.1556, 2014

    Karen Simonyan and Andrew Zisserman. Very deep convo- lutional networks for large-scale image recognition.CoRR, abs/1409.1556, 2014. 5, 7

    Pith/arXiv arXiv 2014

  31. [31]

    Scalability in perception for autonomous driving: Waymo open dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446–2454, 2020. 6

    2020

  32. [32]

    The beer-lambert law.Journal of chem- ical education, 39(7):333, 1962

    Donald F Swinehart. The beer-lambert law.Journal of chem- ical education, 39(7):333, 1962. 6

    1962

  33. [33]

    Multi-view stereo in the deep learning era: A comprehensive review.Displays, 70: 102102, 2021

    Xiang Wang, Chen Wang, Bing Liu, Xiaoqing Zhou, Liang Zhang, Jin Zheng, and Xiao Bai. Multi-view stereo in the deep learning era: A comprehensive review.Displays, 70: 102102, 2021. 2

    2021

  34. [34]

    Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025. 3, 7, 8

    Pith/arXiv arXiv 2025

  35. [35]

    Emie-map: Large-scale road surface recon- struction based on explicit mesh and implicit encoding

    Wenhua Wu, Qi Wang, Guangming Wang, Junping Wang, Tiankun Zhao, Yang Liu, Dongchao Gao, Zhe Liu, and Hes- heng Wang. Emie-map: Large-scale road surface recon- struction based on explicit mesh and implicit encoding. In European Conference on Computer Vision, pages 370–386. Springer, 2024. 2

    2024

  36. [36]

    Bev-gs: Feed-forward gaussian splatting in bird’s-eye-view for road reconstruction

    Wenhua Wu, Tong Zhao, Chensheng Peng, Lei Yang, Yintao Wei, Zhe Liu, and Hesheng Wang. Bev-gs: Feed-forward gaussian splatting in bird’s-eye-view for road reconstruction. arXiv preprint arXiv:2504.13207, 2025. 2

    Pith/arXiv arXiv 2025

  37. [37]

    Segformer: Simple and efficient design for semantic segmentation with transform- ers.Advances in neural information processing systems, 34: 12077–12090, 2021

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transform- ers.Advances in neural information processing systems, 34: 12077–12090, 2021. 2

    2021

  38. [38]

    Street gaussians: Modeling dynamic urban scenes with gaussian splatting

    Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians: Modeling dynamic urban scenes with gaussian splatting. InEuropean Conference on Computer Vision, pages 156–173. Springer, 2024. 2

    2024

  39. [39]

    Emernerf: Emergent spatial-temporal scene decomposition via self-supervision.arXiv preprint arXiv:2311.02077, 2023

    Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Se- ung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision.arXiv preprint arXiv:2311.02077, 2023. 2

    Pith/arXiv arXiv 2023

  40. [40]

    Fresco: Spatial-temporal correspondence for zero-shot video translation

    Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Fresco: Spatial-temporal correspondence for zero-shot video translation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8703– 8712, 2024. 7, 3, 8

    2024

  41. [41]

    Absgs: Recovering fine details for 3d gaussian splat- ting, 2024

    Zongxin Ye, Wenyu Li, Sidun Liu, Peng Qiao, and Yong Dou. Absgs: Recovering fine details for 3d gaussian splat- ting, 2024. 1

    2024

  42. [42]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836–3847, 2023. 2, 3, 7, 8

    2023

  43. [43]

    Tpsence: Towards artifact-free realistic rain generation for deraining and object detection in rain

    Shen Zheng, Changjie Lu, and Srinivasa G Narasimhan. Tpsence: Towards artifact-free realistic rain generation for deraining and object detection in rain. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5394–5403, 2024. 2, 3

    2024

  44. [44]

    Scene parsing through ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3

    2017

  45. [45]

    Drivinggaussian: Composite gaussian splatting for surrounding dynamic au- tonomous driving scenes

    Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic au- tonomous driving scenes. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21634–21643, 2024. 2 10 WeatherCity: Urban Scene Reconstruction with Controllable...

    2024

  46. [46]

    Implementation Details We build WeatherCity upon a dynamic Gaussian scene graph following the node design in OmniRe [4], contain- ing a sky node, a static background node, and multiple rigid and non-rigid object nodes for vehicles and pedestrians, re- spectively, each represented by 3D Gaussian primitives with learnable position, scale, rotation, opacity,...

  47. [47]

    editing direction

    generates target weather effects based on text prompts. However, experimental results indicate that it is prone to scene content distortion (e.g., vehicle deformation, incor- rect lane markings), lacks fine-grained control over weather intensity, and has a slow inference speed (only 0.033 FPS), making it difficult to meet the real-time and consistency re-...

  48. [48]

    Detailed Quantitative and Qualitative Results Table 5 and Table 6 present the complete quantitative com- parison results for the Waymo and nuScenes datasets, re- spectively

    Additional Results and Analysis 7.1. Detailed Quantitative and Qualitative Results Table 5 and Table 6 present the complete quantitative com- parison results for the Waymo and nuScenes datasets, re- spectively. Our method significantly outperforms all base- line approaches (ControlNet, FRESCO, and TurboEdit) across all metrics. Specifically, higher CLIP-S...

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.