WeatherEdit: Controllable Weather Editing with 4D Gaussian Field

Chenghao Qian; Gustav Markkula; Wenjing Li; Yuhu Guo

REVIEW 4 major objections 6 minor 2 cited by

WeatherEdit: Controllable Weather Editing with 4D Gaussian Field

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims a single pipeline can turn sunny driving footage into controllable snow, rain, and fog with adjustable severity and multi-view consistency.

desk verdict Solid engineering paper with a strong downstream result; the central attention equation is too ambiguous to reproduce, and the physics language oversells a constant-velocity model. read the letter →

arxiv 2505.20471 v3 pith:NI3BDSFW submitted 2025-05-26 cs.CV cs.AIcs.ETcs.LGcs.RO

Chenghao Qian , Wenjing Li , Yuhu Guo , Gustav Markkula This is my paper · ORCID

classification cs.CVcs.AIcs.ETcs.LGcs.RO

keywords weatherediting4DGaussianfielddiffusionmodeltemporal-viewattentionmulti-viewconsistencycontrollableseveritydrivingscenesimulation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that normal multi-view driving footage can be edited into a full 3D weather scene with user control over weather type, whether the effect is static or ongoing, and severity. Its argument is a two-stage pipeline: one diffusion model carries all the 2D background weather tones, and a separate dynamic 4D Gaussian field supplies the falling particles. The experiments compare against 2D and 3D baselines, ablate the attention and particle modules, and show that the simulated weather frames improve a downstream semantic segmentation task. If the central claim holds, adverse-weather data for autonomous driving simulation can be produced from sunny data instead of being collected or hand-simulated.

What carries the argument

Three mechanisms carry the argument. (1) An all-in-one low-rank adapter: each weather style has a low-rank matrix $L_i$ added to the pretrained weights $W_0$, so a single model's forward pass is $h = W_0x + L_ix$, with the weather type selected by a text prompt and the effect localized by a semantic map. (2) Temporal-View attention: during inference, a frame's query $Q^v_t$ attends to keys and values from the center view and from its adjacent frames $t-1$ and $t+1$, and the outputs are blended with self-attention via a weight $\lambda$, producing multi-view and multi-frame consistency without retraining. (3) A dynamic 4D Gaussian field: particles are 3D Gaussians with attributes $A_i = \{C_i, P_i, R_i, S_i, O_i\}$, position updates follow $P_j(t+\Delta t) = P_j(t) + D_j\Delta t$, particles reset at field bounds, and the whole field is shifted by the camera-relative transform $\Delta T = T_tT_0^{-1}$, so the weather persists over large scenes with a bounded particle budget.

What would settle it

Run WeatherEdit on a multi-view sequence recorded with a rig that has no center camera overlapping the side views, then measure warp error and the Bhattacharyya distance between views; a large degradation would confirm the center-view assumption, while no degradation would show the assumption is not load-bearing. A second check is to compare the optical flow of rendered rain or snow streaks over several seconds with footage captured under controlled rainfall and snowfall, since the paper models particle motion as a single constant direction and speed.

Watch

Extended reading notes

Core claim

The central claim is that weather can be edited into a 3D driving scene in two decoupled stages and still come out coherent and controllable. In the first stage, background weather tones are injected per image by one diffusion model adapted with style-specific low-rank weights and conditioned on semantic segmentation maps, while a parameter-free Temporal-View attention makes side cameras query the center camera and each frame query its neighbours so stochastic edits do not drift across views or time. In the second stage, weather is built as a 4D Gaussian field: 3D Gaussians carry color, position, rotation, scale, and opacity, positions are advanced by a constant velocity, and particles are recycled inside a camera-aligned local field, giving falling snow, rain, and drifting fog whose density is user-tunable. The paper presents this as a unified framework for the realism requirements R1-R3 and controllability requirements C1-C3, and supports it with qualitative comparisons, ablations, and downstream segmentation gains.

Load-bearing premise

The multi-view consistency claim depends on the cameras being arranged so a front or center camera overlaps both side cameras; on a rig without such a shared center view, the view-alignment mechanism has no reference to align to.

Editorial extensions

If this is right

One diffusion model with one all-in-one adapter can switch between snow, rain, and fog from a text prompt and a semantic map, so no separate editing model is needed per weather type.
The Temporal-View attention keeps edited frames coherent across adjacent timestamps and across left-front-right cameras, directly suppressing the drift that independent diffusion edits produce.
The 4D Gaussian field can render dynamic snow, rain, and fog with per-particle attributes, and adjusting quantity and attribute distributions translates into light-to-heavy severity control.
Separating background editing from particle injection lets users specify static weather states (a wet road without falling rain) and ongoing events (active snowfall) independently.
Segmentation models trained with WeatherEdit-simulated adverse frames improve mIoU on adverse-weather datasets, with gains up to 14.9 percentage points, indicating the simulated data carries useful weather degradation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Not tested in the paper: a constant-velocity update is used for all particles, so gusts, acceleration, and particle-surface interactions are left out; replacing $D_j$ with a stochastic or field-driven velocity would be a natural extension and could be checked by comparing rendered streak motion to real footage.
Implicit in the camera-aligned local field: because particles are recycled and the field follows the camera, the same particle budget should support arbitrarily long drives; a long-trajectory stress test is a direct next experiment.
Because weather type and severity are controlled by different stages, an untested interpolation route is blending the style matrices $L_i$ to produce intermediate weather (e.g., rain-to-snow transitions), which the current text-prompt selection does not directly offer.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: WeatherEdit is a competent systems paper with a clear pipeline and one strong result — the downstream segmentation gains — but the formal statement of its key TV-attention mechanism is sloppy enough that the multi-view consistency claim is not reproducible from the text. The stress-test note is correct, not a nitpick.

What is genuinely new: this is the first unified pipeline I know of that combines multi-style LoRA injection into a single diffusion model, a parameter-free temporal-view attention anchored on the front camera, and a camera-relative 4D Gaussian field for snow, rain, and fog particles. The particle-recycling and local field alignment are sensible engineering choices that make dynamic weather cheap to render. The paper also gives real credit by using an independent downstream semantic segmentation test: +12.4/+14.9 mIoU on ACDC and +9.0/+6.8 on MUSE are meaningful numbers, and that test is not circular. CLIP metrics support the qualitative story, though they are weak evidence by themselves.

Soft spots, in order of size. First, Eq. (2) as written is self-attention: Q, K, and V all carry the same view index v. The prose says key and value come from the center view, but the equation does not say that. Either the notation is wrong or the mechanism is something else. Since Table 4 shows removing view attention hurts Bhattacharyya distance, the implementation probably does what the prose claims, but a reader cannot verify it. This must be fixed before the paper is usable. Second, "physical-based modelling" is an overstatement: Eq. (5) is a constant-velocity update, and severity is controlled by hand-set particle counts and attributes. Fine as an approximation, but call it what it is. Third, evaluation is thin: no error bars, only eight scenes, no failure cases, no comparison on the 4D particle side, and no code or data release. The camera-rig assumption (front camera overlapping both side cameras) is load-bearing and not tested on other layouts.

Overall, the central claim — controllable weather editing via this two-stage pipeline — probably holds. It is an engineering contribution, not a conceptual one. I would send it to peer review, but only with the expectation that the authors clarify Eq. (2), tone down the physics claim, and release at least the editing code.

Referee Report

4 major / 6 minor

Summary. WeatherEdit proposes a two-stage pipeline for editing sunny driving sequences into snowy, rainy, or foggy scenes. In the first stage, an all-in-one LoRA adapter conditions a diffusion model on multiple weather styles and semantic maps, while a Temporal-View attention module is introduced to align edited images across frames and viewpoints. In the second stage, the edited images are used to reconstruct a 3D scene, and a 4D Gaussian field with attribute modeling, constant-velocity particle dynamics, particle recycling, and camera-aligned local field is composited into the scene. The paper reports CLIP-based improvements over ControlNet, InstructPix2Pix, TurboEdit, 3D Stylization, and ClimateNeRF, ablations of semantic conditioning and TV-attention, and downstream semantic segmentation gains up to +14.9 mIoU when the simulated weather frames are used as training augmentation.

Significance. If the framework works as described, it is a useful unification of several capabilities that are usually treated separately: multi-type weather control, multi-frame and multi-view consistency, static versus ongoing weather events, and severity control in a single 2D-to-4D pipeline. The paper has clear practical value for autonomous-driving simulation, and I credit the authors for an evaluation that goes beyond image-quality metrics: the downstream segmentation experiment (Table 3) is independent of the training objective and demonstrates a real use case. The design of the 4D Gaussian field with particle recycling and camera alignment is also a sensible efficiency contribution. However, the novelty is partly incremental over prior diffusion editing, 3D Gaussian reconstruction, and the authors' previous WeatherGS, and the quantitative evidence is concentrated on 2D background editing; the dynamic 4D weather realism and the severity control are demonstrated mostly qualitatively. With the TV-attention specification corrected and the severity-control mechanism clarified, the contribution would be a solid systems paper for a computer vision venue.

major comments (4)

[This concerns Equation (2) in the 'Temporal-View Consistency Alignment' section.] Equation (2) is written with the same view index v on the query, key, and value, so as printed it is self-attention within a single view and cannot implement the described cross-view mechanism in which left and right cameras query the center view. The accompanying sentence says that K_v^t and V_v^t are 'derived from the center view v,' which makes the notation inconsistent: v cannot simultaneously denote the side view and the center view. Because this attention block is one of the three pillars of the central claim, the multi-view consistency mechanism is not reproducible from the manuscript as it stands. Please rewrite Eq. (2) with distinct indices (for example, Q_v^t, K_c^t, V_c^t for a center view c), and specify the order in which cameras and frames are aggregated.
[This concerns 'Dynamic Simulation' (Eqs. 5-9) and the 'Controllable Weather Simulation' section.] The paper's 'physical-based modelling and simulation' overstates what Eq. (5) actually provides: particle motion is a constant directional velocity D_j rather than an integration of gravity or wind, and the severity knob is implemented by hand-set attributes (color, opacity, scale, velocity) and quantity q. There is no defined severity variable linking these parameters to meteorological or perceptual quantities, and the light/moderate/heavy distinction in Figure 8 is not evaluated quantitatively. Since controllable severity (C3) is a central contribution, please define the severity parameter explicitly and provide at least one quantitative or perceptual evaluation of severity adjustment.
[This concerns the assumed camera rig in 'Temporal-View Consistency Alignment' and the datasets used in the experiments.] The TV-attention design assumes a structured left-front-right camera arrangement in which a front camera overlaps both side cameras, but the experiments are run on Pandaset, Waymo, nuScenes, and KITTI-360, whose camera rigs differ in number and layout. The paper does not report which camera subset was used in the eight scenes or whether the chosen views satisfy the overlap assumption. To support the multi-view consistency claim outside the assumed rig geometry, please specify the camera configuration and, ideally, include an experiment on a rig without a central overlapping view, where Eq. (2) would otherwise have no shared reference.
[This concerns the ablation results in Table 4 and the text describing them.] The ablation table does not fully support the narrative that TV-attention improves temporal consistency: the full model's warp error (0.041) is higher than self-attention plus temporal attention alone (0.039), and the term 'balanced reduction' is not defined. The claim of global coherence rests mainly on the best Bhattacharyya distance (0.245). Please clarify the aggregation criterion or acknowledge the trade-off between view and temporal consistency, and ideally report multiple runs or significance information, since the differences are small.

minor comments (6)

[This concerns Equation (9) in the 'Local Field Alignment' subsection.] Equation (9) uses Δt for the translation vector, which collides with the time increment Δt in Eq. (5); please rename the translation, for example Δτ, to avoid ambiguity.
[This concerns the all-in-one adapter description in the 'All-in-one Adapter' paragraph.] The all-in-one LoRA adapter is described only at the level h = W0 x + L_i x; please report how the multiple LoRA branches are trained (jointly or separately), how the style index i is selected at inference, and what rank is used for L_i.
[This concerns Table 3 in the 'Benefits for Downstream Tasks' section.] Table 3 reports only point estimates; please add the number of training samples, the evaluation protocol, and error bars or significance testing before interpreting the +14.9 mIoU as a general gain.
[This concerns terminology in the 'Temporal-View Attention' ablation subsection.] The ablation text refers to 'spatial attention' when discussing the view-attention component, while the method section calls it 'view attention'; please unify the terminology to avoid confusion.
[This concerns the reference list entry for Schmidt et al.] The bibliography contains an incomplete entry 'Schmidt, V.; ... et al. ????' with no year or venue; please complete the citation.
[This concerns typos throughout the manuscript.] Please fix the typos 'physical-based' (should be 'physics-based') and 'Simluation' (should be 'Simulation'), and elsewhere proofread the text for similar issues.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: WeatherEdit's claims are empirically validated against external benchmarks rather than derived from its own inputs.

full rationale

WeatherEdit is an empirical construction; it does not claim a first-principles derivation, and no quantitative 'prediction' is shown to reduce to a fitted parameter or to the authors' own definitions. The all-in-one LoRA adapter is trained on paired weather images and evaluated with CLIP scores against external baselines (ControlNet, InstructPix2Pix, TurboEdit, ClimateNeRF). The 4D Gaussian field is parameterized by user-controlled attributes and approximated dynamics; severity control is a manual parameter adjustment, not a fitted quantity presented as a prediction. The downstream semantic segmentation experiment (Table 3) uses an independent benchmark task and external segmentation models, so the reported mIoU gains are not forced by construction. The only self-citation (Qian et al. 2025, WeatherGS) appears in related work as an example of 3D Gaussian Splatting artifact removal and is not load-bearing. The notation in Eq. (2) is ambiguous and could be read as self-attention, but this is a specification/reproducibility concern rather than circularity: the ablation in Table 4 shows that removing view attention changes the Bhattacharyya distance, so the implemented module is not trivially identical to its input. No circular step satisfying the quoted-evidence requirement was found.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim depends on hand-set simulation parameters (velocities, attribute distributions, quantity, field size, recycle bounds, attention weight) and on domain assumptions: a left-front-right camera rig with shared center view, the adequacy of constant-velocity particle motion, the availability of reliable sunny-condition semantic maps, and the validity of CLIP metrics as realism proxies. The 4D Gaussian field is a rendering representation rather than a new physical entity, so invented_entities is left empty.

free parameters (6)

Constant particle velocity D_j = not disclosed
Eq. 5 updates each particle position as P(t+dt)=P(t)+D*dt; the three velocity components are chosen by hand per weather type and are not reported, yet they determine the visual speed and direction of snow and rain.
Particle attribute distributions G(A_i) = not disclosed
Section 'Attribute Modeling' models color, position, rotation, scale, and opacity as Gaussian distributions; means and variances are not given, so severity and appearance tuning cannot be reproduced.
Particle quantity q and field size B_i = user-controlled, not calibrated
These control particle density and severity; the paper shows qualitative light/moderate/heavy examples but gives no mapping from q or B_i to measured intensity.
Attention balance lambda = not disclosed
Eq. 4 combines self-, view-, and temporal-attention using lambda; no value or sensitivity analysis is provided, though this weighting controls consistency behavior.
Recycle bounds pmin, pmax and offset delta_p = not disclosed
Eq. 7 resets particles that leave the simulation field; the bounds and offset are required to reproduce the field coverage but are not specified.
All-in-one LoRA adapter weights = trained on 1,237 pairs; not released
The adapted diffusion model is central to style editing; it is fitted, not derived, and the learned weights are not available.

assumptions (5)

domain assumption Multi-view driving cameras are arranged in a left-front-right configuration with the front camera overlapping both side cameras.
Used to justify Eq. 2; if the rig lacks this layout, view attention has no shared anchor.
ad hoc to paper Constant directional velocity is an adequate model of rain, snow, and fog dynamics for visual realism.
Eq. 5 deliberately replaces gravity, drag, and turbulence with a constant speed; the realism claim rests on this approximation.
domain assumption Sunny-condition semantic segmentation maps are accurate enough to condition weather editing.
The adapter is trained and applied with segmentation maps; noisy or out-of-domain maps could degrade editing, and no failure analysis is given.
domain assumption CLIP-S and CLIP-DS scores are valid proxies for content preservation and weather alignment.
Tables 1 and 2 use these metrics; no human perceptual study or calibrated alternatives are provided.
domain assumption Reconstructed 3D scenes from edited 2D images remain geometrically reliable enough to host 4D particles.
The pipeline reconstructs scenes from edited images, so editing distortions become 3D geometry errors; the paper does not quantify this.

how reviews work

0 comments

Cite this review

Pith. "Pith review of WeatherEdit: Controllable Weather Editing with 4D Gaussian Field." pith.science (2026). https://pith.science/paper/NI3BDSFW

@misc{pith2026250520471,
  author       = {Pith},
  title        = {Pith review of: WeatherEdit: Controllable Weather Editing with 4D Gaussian Field},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NI3BDSFW}},
  note         = {Machine review of arXiv:2505.20471}
}

read the original abstract

In this work, we present WeatherEdit, a novel weather editing pipeline for generating realistic weather effects with controllable types and severity in 3D scenes. Our approach is structured into two key components: weather background editing and weather particle construction. For weather background editing, we introduce an all-in-one adapter that integrates multiple weather styles into a single pretrained diffusion model, enabling the generation of diverse weather effects in 2D image backgrounds. During inference, we design a Temporal-View (TV-) attention mechanism that follows a specific order to aggregate temporal and spatial information, ensuring consistent editing across multi-frame and multi-view images. To construct the weather particles, we first reconstruct a 3D scene using the edited images and then introduce a dynamic 4D Gaussian field to generate snowflakes, raindrops and fog in the scene. The attributes and dynamics of these particles are precisely controlled through physical-based modelling and simulation, ensuring realistic weather representation and flexible severity adjustments. Finally, we integrate the 4D Gaussian field with the 3D scene to render consistent and highly realistic weather effects. Experiments on multiple driving datasets demonstrate that WeatherEdit can generate diverse weather effects with controllable condition severity, highlighting its potential for autonomous driving simulation in adverse weather. See project page: https://jumponthemoon.github.io/w-edit

Figures

Figures reproduced from arXiv: 2505.20471 by the authors.

**Figure 1.** Realistic, Controllable Weather Editing. Given input sunny weather images from multi-view cameras across multiple timesteps, WeatherEdit can generate high-fidelity scenarios under diverse weather conditions (e.g., foggy, snowy, and rainy) with various severity levels (e.g., light, moderate, heavy) based on user instructions. images (Rothmeier and Huber 2021). More recently, ClimateNeRF (Li et al. 2023) utilized Neu… view at source ↗

**Figure 2.** Overview of the WeatherEdit Framework. (a) Weather Background Editing: An all-in-one adapter enables the synthesis of diverse weather effects (e.g., snow, rain, fog) from text prompts and segmentation maps. Inference takes multi-frame and multi-view images as input, using temporal-view (TV) attention to ensure consistency. (b) Weather Particle Construction: A 4D Gaussian field, initialized from 3D Gaussians, under… view at source ↗

**Figure 3.** Single-style Adapters vs. All-in-One Adapter. (a) Utilizes separate adapters for each weather style (e.g., snow, rain, fog), while (b) employs a unified all-in-one adapter for efficient multi-style adaptation. Blue box highlights adapters to be trained. mentation map M, we encode the latent representation as z0 = E(x, M). The model fine-tuning objective is then formulated as: L(∆θ) = Eϵ,t ∥ϵ − ϵθ+∆θ( √ atz0 + √ 1 … view at source ↗

Figures from the paper (6 more)

**Figure 5.** Figure 5: Weather Particle Construction Process. The attribute modeling stage defines weather properties to construct a 4D Gaussian field, which then undergoes dynamic simulation to model fog, rain, and snow. Local field alignment ensures seamless integration with the 3D scen…

**Figure 6.** Figure 6: Comparison with 2D-based Editing Methods. We show (a) original images with edited results from (b) ControlNet, (c) InstructPix2Pix, (d) TurboEdit, and (e) Ours. Existing methods often suffer from over-stylization, content removal, and inconsistencies in spatial and tem…

**Figure 7.** Figure 7: Comparison with 3D-based weather synthesis methods in editing snowy. 3D Stylization (b) produces only a subtle snowy effect on the sidewalk, while (c) ClimateNeRF present unnatural snow cover. In contrast, our method (d) synthesizes a realistic snowy effect while ensur…

**Figure 8.** Figure 8: Controllable Weather Effect and Severity in Driving Scenes. WeatherEdit can modify (a) original driving scenes from Pandaset, Nuscenes, and Waymo datasets by generating snowy, foggy, and rainy effects while preserving scene structure in (b). Additionally, it enables (…

**Figure 10.** Figure 10 [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]

**Figure 11.** Figure 11: Ablation on semantic conditioned input. Left (a) shows the input image paired with a segmentation map. With semantic conditioning (b), snowfall presents more prominent feature (highlighted in red box) and aligns better with scene structures compared to (c) without it…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

muSync-GS: Physics-Synchronized Driving Video Synthesis for Weather and Geometric Road Hazards
cs.CV 2026-08 conditional novelty 6.0 of 10

muSync-GS couples weather and road-shape edits in driving videos to a calibrated vehicle-dynamics model, so the synthesized ego motion and telemetry change with the same controls that drive the visual edits.
Multi-Sensor Alignment for Weather Simulations
cs.AI 2026-07 conditional novelty 4.0 of 10

A new cross-sensor weather-simulation alignment method makes fog severity and snow/rain particle positions consistent between LiDAR and camera, yielding less optimistic but more robust 3D detection evaluations.

Reference graph

Works this paper leans on

45 extracted references · 33 canonical work pages · cited by 2 Pith papers

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
[2]

Deep learning-based robust positioning for all-weather autonomous driving

Yasin Almalioglu, Mehmet Turan, Niki Trigoni, and Andrew Markham. Deep learning-based robust positioning for all-weather autonomous driving. Nature Machine Intelligence, 4 0 (9): 0 749–760, 2022

work page 2022
[3]

On a measure of divergence between two statistical populations defined by their probability distributions

Anil Kumar Bhattacharyya. On a measure of divergence between two statistical populations defined by their probability distributions. Bulletin of the Calcutta Mathematical Society, 35: 0 99--109, 1943

work page 1943
[4]

Muses: The multi-sensor semantic perception dataset for driving under uncertainty

Tim Br \"o dermann, David Bruggemann, Christos Sakaridis, Kevin Ta, Odysseas Liagouris, Jason Corkill, and Luc Van Gool. Muses: The multi-sensor semantic perception dataset for driving under uncertainty. In European Conference on Computer Vision (ECCV), 2024

work page 2024
[5]

Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392--18402, 2023

work page 2023
[6]

nuscenes: A multimodal dataset for autonomous driving

Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621--11631, 2020

2020
[7]

Omnire: Omni urban scene reconstruction

Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, Li Song, and Yue Wang. Omnire: Omni urban scene reconstruction. In The Thirteenth International Conference on Learning Representations, 2025

2025
[8]

Turboedit: Text-based image editing using few-step diffusion models, 2024

Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models, 2024

work page 2024

Show all 45 references

[9]

Modeling the accumulation of wind-driven snow

Bryan E Feldman and James F O'Brien. Modeling the accumulation of wind-driven snow. In ACM SIGGRAPH 2002 conference abstracts and applications, pages 218--218, 2002

2002
[10]

Stylegan-nada: Clip-guided domain adaptation of image generators

Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41 0 (4): 0 1--13, 2022

2022
[11]

A history of the unity game engine

John K Haas. A history of the unity game engine. 2014

2014
[12]

Semantic understanding of foggy scenes with purely synthetic data

Martin Hahner, Dengxin Dai, Christos Sakaridis, Jan-Nico Zaech, and Luc Van Gool. Semantic understanding of foggy scenes with purely synthetic data. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC), pages 3675--3681. IEEE, 2019

2019
[13]

Denoising diffusion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

2020
[14]

Lora: Low-rank adaptation of large language models

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1 0 (2): 0 3, 2022

2022
[15]

Weathergan: Unsupervised multi-weather image-to-image translation via single content-preserving uresnet generator

Sunhee Hwang, Seogkyu Jeon, Yu-Seung Ma, and Hyeran Byun. Weathergan: Unsupervised multi-weather image-to-image translation via single content-preserving uresnet generator. Multimedia Tools and Applications, 81 0 (28): 0 40269--40288, 2022

2022
[16]

3d gaussian splatting for real-time radiance field rendering

Bernhard Kerbl, Georgios Kopanas, Thomas Leimk \"u hler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 0 (4), 2023

2023
[17]

A closed-form solution to photorealistic image stylization

Yijun Li, Ming-Yu Liu, Xueting Li, Ming-Hsuan Yang, and Jan Kautz. A closed-form solution to photorealistic image stylization. In Proceedings of the European conference on computer vision (ECCV), pages 453--468, 2018

2018
[18]

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. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

2023
[19]

Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d

Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (3): 0 3292--3310, 2022

2022
[20]

Magic3d: High-resolution text-to-3d content creation

Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognitio...

2023
[21]

Humangaussian: Text-driven 3d human generation with gaussian splatting

Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6646--6657, 2024

2024
[22]

Sdedit: Guided image synthesis and editing with stochastic differential equations

Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021

2021 arXiv
[23]

Srinivasan, Matthew Tancik, Jonathan T

Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceeding...

2020
[24]

T2i-adapter: learning adapters to dig out more controllable ability for text-to-image diffusion models

Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Th...

2024
[25]

One-step image translation with text-to-image models

Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models. arXiv preprint arXiv:2403.12036, 2024

2024 arXiv
[26]

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. In International Conference on Robotics and Automation(ICRA), 2025 a

2025
[27]

Weatherdg: Llm-assisted procedural weather generation for domain-generalized semantic segmentation

Chenghao Qian, Yuhu Guo, Yuhong Mo, and Wenjing Li. Weatherdg: Llm-assisted procedural weather generation for domain-generalized semantic segmentation. IEEE Robotics and Automation Letters, 2025 b

2025
[28]

Allweather-net: Unified image enhancement for autonomous driving under adverse weather and low-light conditions

Chenghao Qian, Mahdi Rezaei, Saeed Anwar, Wenjing Li, Tanveer Hussain, Mohsen Azarmi, and Wei Wang. Allweather-net: Unified image enhancement for autonomous driving under adverse weather and low-light conditions. In International Conference on Pattern Recognition, pages 151--1...

2025
[29]

Learning transferable visual models from natural language supervision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pa...

2021
[30]

High-resolution image synthesis with latent diffusion models, 2021

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021

2021
[31]

Let it snow: On the synthesis of adverse weather image data

Thomas Rothmeier and Werner Huber. Let it snow: On the synthesis of adverse weather image data. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3300--3306, 2021

2021
[32]

Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding

Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10765--10775, 2021

2021
[33]

Climategan: Raising climate change awareness by generating images of floods

Victor Schmidt, Alexandra Luccioni, M \'e lisande Teng, Tianyu Zhang, Alexia Reynaud, Sunand Raghupathi, Gautier Cosne, Adrien Juraver, Vahe Vardanyan, Alex Hern \'a ndez-Garc \' a, et al. Climategan: Raising climate change awareness by generating images of floods. In Internat...
[34]

Visualizing the consequences of climate change using cycle-consistent adversarial networks

Victor Schmidt, Alexandra Luccioni, S Karthik Mukkavilli, Narmada Balasooriya, Kris Sankaran, Jennifer Chayes, and Yoshua Bengio. Visualizing the consequences of climate change using cycle-consistent adversarial networks. arXiv preprint arXiv:1905.03709, 2019

1905 arXiv
[35]

A material point method for snow simulation

Alexey Stomakhin, Craig Schroeder, Lawrence Chai, Joseph Teran, and Andrew Selle. A material point method for snow simulation. ACM Transactions on Graphics (TOG), 32 0 (4): 0 1--10, 2013

2013
[36]

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, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang,...

2020
[37]

Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M. Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2353--2363, 2022

2022
[38]

Pandaset: Advanced sensor suite dataset for autonomous driving

Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In 2021 IEEE international intelligent transportation systems conference (ITSC), page...

2021
[39]

Nerf-ds: Neural radiance fields for dynamic specular objects

Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8285--8295, 2023

2023
[40]

Aviles-Rivero, Yulun Zhang, Jing Qin, and Lei Zhu

Yijun Yang, Hongtao Wu, Angelica I. Aviles-Rivero, Yulun Zhang, Jing Qin, and Lei Zhu. Genuine knowledge from practice: Diffusion test-time adaptation for video adverse weather removal. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25606-...

2024
[41]

Bdd100k: A diverse driving dataset for heterogeneous multitask learning

Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pag...

2020
[42]

Weatherstream: Light transport automation of single image deweathering

Howard Zhang, Yunhao Ba, Ethan Yang, Varan Mehra, Blake Gella, Akira Suzuki, Arnold Pfahnl, Chethan Chinder Chandrappa, Alex Wong, and Achuta Kadambi. Weatherstream: Light transport automation of single image deweathering. In Proceedings of the IEEE/CVF Conference on Computer ...

2023
[43]

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 (ICCV), pages 3836--3847, 2023 b

2023
[44]

Unpaired image-to-image translation using cycle-consistent adversarial networks

Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223--2232, 2017

2017
[45]

The flow from simulation to reality

K \'a roly Zsolnai-Feh \'e r. The flow from simulation to reality. nature physics, 18 0 (11): 0 1260--1261, 2022

2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

[2] [2]

Deep learning-based robust positioning for all-weather autonomous driving

Yasin Almalioglu, Mehmet Turan, Niki Trigoni, and Andrew Markham. Deep learning-based robust positioning for all-weather autonomous driving. Nature Machine Intelligence, 4 0 (9): 0 749–760, 2022

work page 2022

[3] [3]

On a measure of divergence between two statistical populations defined by their probability distributions

Anil Kumar Bhattacharyya. On a measure of divergence between two statistical populations defined by their probability distributions. Bulletin of the Calcutta Mathematical Society, 35: 0 99--109, 1943

work page 1943

[4] [4]

Muses: The multi-sensor semantic perception dataset for driving under uncertainty

Tim Br \"o dermann, David Bruggemann, Christos Sakaridis, Kevin Ta, Odysseas Liagouris, Jason Corkill, and Luc Van Gool. Muses: The multi-sensor semantic perception dataset for driving under uncertainty. In European Conference on Computer Vision (ECCV), 2024

work page 2024

[5] [5]

Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392--18402, 2023

work page 2023

[6] [6]

nuscenes: A multimodal dataset for autonomous driving

Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621--11631, 2020

2020

[7] [7]

Omnire: Omni urban scene reconstruction

Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, Li Song, and Yue Wang. Omnire: Omni urban scene reconstruction. In The Thirteenth International Conference on Learning Representations, 2025

2025

[8] [8]

Turboedit: Text-based image editing using few-step diffusion models, 2024

Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models, 2024

work page 2024