acceptodds
Under review as a conference paper at ICLR 2027

G-WADE: Geometry-Preserving Weather-Stylized Urban Assets Diffusion Enhancement

Abstract

Extending 3D Gaussian Splatting (3DGS) urban assets under adverse weather (e.g., rain, snow) directly from real-world captures is prohibitively expensive, motivating the alternative of editing an existing, geometrically accurate 3DGS scene to synthesize new weather conditions. However, existing weather editing methods are limited in asset augmentation and prone to geometry collapse, multi-view texture inconsistency, and long-range drift. We observe that video diffusion models with depth/edge control can preserve scene geometry while re-stylizing appearance, but often produce substantial appearance discrepancies when revisiting previously seen regions, undermining multi-view consistency and degrading the quality of the resulting 3DGS reconstructions. In this paper, we propose G-WADE: Geometry-preserving Weather-stylized urban Assets Diffusion Enhancement, which is a geometry-fixed framework for the augmentation of weather-stylized urban assets. G-WADE fine-tunes two domain-specific variants of Cosmos-Transfer2.5 (Base-Model) by leveraging VAE temporal downsampling property: a first-frame-conditioned model (F-Model) and a first-last-frame-conditioned model (FL-Model) that extracts a tail latent slice as an additional anchor. Additionally, G-WADE designs trajectory-aware generation strategies with coordinated model scheduling, separately tailored for high-precision single-building capture (orbit-and-descend trajectories) and large-scale urban scenes (center-out grid trajectories). The generated multi-view-consistent, geometry-preserving frames are then used to optimize the attributes of the original 3DGS. Experiments on rain and snow weather stylization over both single-building and large-scale urban 3DGS scenes show that our method achieves effective urban asset augmentation, with better multi-view consistency, perceptual realism, and geometry preservation than existing video and 3D editing baselines.

Primary area:
generative models
Keywords:
3D Gaussian Splatting, Video Diffusion, Weather Editing, Multi-view Consistency, Geometry Preservation, UAV Capture

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_B4LtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

G-WADE edits the weather appearance of urban 3DGS assets using tail-latent conditioning and trajectory-aware anchor reuse. Generated views supervise appearance-only optimization while the original geometry remains fixed.

Critical strengths

Endpoint anchoring addresses appearance inconsistency when revisiting scene regions, while freezing geometric parameters explicitly preserves the original asset geometry. Table 1 reports the best PSNR and SSIM in both scene settings.

Critical weaknesses

(1) The video baselines use non-fine-tuned models, whereas G-WADE receives domain-specific training. A same-data fine-tuned first-frame baseline, together with separate ablations of tail conditioning and trajectory scheduling under identical geometry constraints, is needed to establish the source of improvement. (2) The paper does not specify the numbers of scenes and training clips, the evaluation-set size, or whether fine-tuning and evaluation use disjoint scenes. Consequently, the breadth of evaluation and generalization to unseen scenes cannot be assessed. (3) PSNR and SSIM compare reconstructed views against independently generated stylized images, but it is unclear whether these targets are shared across methods or generated separately by each method. These scores measure agreement with synthetic targets, rather than independently establishing weather realism; the target-generation protocol and complementary realism evaluation should be provided.

Questions for authors

Please refer to the weaknesses. Section 5.2 describes the long-range baseline as independently generated clips, whereas Figure 4 states that both methods autoregressively condition on the previous chunk’s last frame. Which protocol was actually used?

1 comment

Response to Reviewer B4Lt

We thank the reviewer for the careful methodological critique. The points about baseline fairness and the protocol description are well taken and we have addressed them directly. The new experiments and protocols are in the supplementary.

W1: Baseline fairness and component attribution

We agree the previous video baselines used the non-fine-tuned Cosmos model, which confounds fine-tuning gain with FL/trajectory gain. We have added:

  1. A same-data fine-tuned first-frame-only baseline (F-Model + 3DGS), trained on the identical fine-tuning set under the identical geometry-frozen constraint.
  2. A component-wise ablation under identical geometry constraints (Supplementary Section C):
    • (a) raw Cosmos + 3DGS
    • (b) F-Model + 3DGS (fine-tuned first-frame-only)
    • (c) F-Model + FL (tail-latent) + 3DGS, without trajectory scheduling
    • (d) full G-WADE (F/FL + trajectory scheduling)

This decomposes the improvement into domain fine-tuning (a to b), tail-latent anchoring (b to c), and trajectory-aware scheduling (c to d), directly answering the source of each improvement.

W2: Dataset and evaluation scale

We report the full statistics in Supplementary Section A:

  1. Scenes/clips: 4268 single-building + 5984 multi-building scenes. From 24772 rendered clips to 19852 retained (retention 80.14%).
  2. Evaluation-set size: For each test scene, we split the views into training and test sets at a 4:1 ratio (train:test).
  3. Disjointness: the test scenes are completely disjoint from the fine-tuning scenes and are never seen during fine-tuning.

W3: Reference-image protocol and complementary realism evaluation

We clarify the protocol. In Table 1, PSNR/SSIM reference images are each method's own stylized views at held-out viewpoints (intra-method multi-view consistency). To address the "blurry is easier to fit" concern, we add a shared-reference setting (all methods measured against the same fixed base-Cosmos targets, Supplementary Section I), and we provide an independent perceptual user study on unseen scenes (2AFC on realism, consistency, geometry. See Supplementary Section F). The target-generation protocol (base-Cosmos stylized targets, curation criteria) is fully documented in Supplementary Section A.

Q1: The Sec. 5.2 vs. Fig. 4 protocol contradiction

We thank the reviewer for pointing out this inconsistency. We acknowledge that the description in Section 5.2 was an inadvertent writing error. We have corrected the text in the revised manuscript to make it consistent with Fig. 4. Specifically, both the baseline and our method generate the long sequence autoregressively, conditioning each subsequent generation on the last frame of the previous chunk. This autoregressive protocol is used to evaluate the accumulation of errors during long-range inference.

We believe these additions resolve the reviewer's concerns.

Reviewer_hEQzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper takes Gaussian splat reconstructions of outdoor scenes and changes their weather conditions, while assuming fixed geometry (so, only the appearance of the Gaussians is updated).

The solution is based on finetuned video models based on Cosmos 2.5 to edit the appearance in "video space", then propagating it back to the Gaussians. The method works by generating camera loops through the scene, and editing the appearance first along sparse anchors, and then filling in the in-between segments by using the F-Model and FL-Model (conditioned on first, and first-last frames of the segment).

Critical strengths

Overall, I feel this is a reasonable and powerful approach. The paper is clearly written and the results look great.

Critical weaknesses

I would not call this critical, but FL-models are not new; for example, Minimax H3 has a first-las frame conditioned variant that might be capable of some of these edits in a zero-shot manner, and this would be nice to compare to.

Overall, I don't believe the FL finetuning of Cosmos is much of a new contribution; the key contribution is the application to radiance field editing, and the method could be generalized well beyond outdoor weather edits. There is quite a bit of competition in the weather editing space, so generalizing beyond it to general editing would be the biggest improvement to the existing paper, in my opinion.

Questions for authors

How far beyond weather edits and urban scenes could the method be generalized? What about virtual floor / wall remodels in indoor scenes, or other applications?

1 comment

Response to Reviewer hEQz

We thank the reviewer for the positive assessment ("reasonable and powerful approach", "results look great") and for the constructive suggestion to broaden the scope. We address the weakness and question below.

W1: FL models are not new; the key contribution is the application

We agree, and we have reframed the paper accordingly: we no longer claim first-last-frame conditioning as the central novelty. We now present the contribution as an application-driven, end-to-end system for radiance-field appearance editing composed of two mechanisms that are new in combination:

  1. Tail-latent endpoint encoding: exploiting the 1+4n1{+}4n Wan VAE temporal structure to derive a compact endpoint anchor (co-trained with dropout), rather than generic two-frame conditioning.
  2. Trajectory-aware scheduling: orbit-and-descend / center-out grid anchor reuse for drift-bounded, loop-closing generation.

The reviewer specifically suggested comparing with existing first-last-frame conditioning variants. We appreciate this suggestion and have added comparisons with both Wan2.1-FLF2V and the MiniMax H3 first-last variant in Section E of the Supplementary Material. Since MiniMax H3 adopts a substantially different conditioning paradigm, including a dedicated video VAE and sequence-level conditioning for DiT denoising, we believe that the comparison with Wan2.1-FLF2V, which follows a more closely related first-last-frame conditioning formulation, provides a more direct and informative analysis of the improvements brought by our approach.

Q1: How far beyond weather and urban scenes can the method generalize?

The framework is not weather-specific and not urban-specific: it accepts depth/edge control plus a text prompt and transfers appearance under fixed geometry. The weather prompts and the urban dataset are one instantiation, not an architectural assumption. Generalizing beyond weather, e.g., indoor floor and wall re-coloring, is a natural extension of the same geometry-fixed protocol, but we have not run a non-weather evaluation in this revision. We agree this generalization is the most valuable direction and highlight it as our primary future work (Supplementary Section G).

We hope the reframed contribution and the zero-shot MiniMax H3 / Wan2.1-FLF2V comparison address the reviewer's points.

Reviewer_kpTFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

G-WADE edits the weather appearance of existing urban 3D Gaussian Splatting scenes while keeping their geometry fixed. It fine-tunes the depth and edge control branches of Cosmos-Transfer2.5 to support first-frame and first-last-frame conditioning, and combines these modes with trajectory-aware generation. The resulting views supervise appearance-only optimization of the original 3DGS.

Critical strengths

Improved consistency when revisiting viewpoints. Adapting the video model to first-last-frame conditioning is a reasonable way to reduce appearance drift when a camera trajectory returns to its starting point. The loop-closure visualizations support the usefulness of this mechanism.

Critical weaknesses
  1. Limited realism of the weather effects. The visualizations appear closer to globally consistent stylization than realistic weather editing. Rainy scenes often exhibit a broadly uniform dark-blue tint across buildings and streets, while snowy scenes are largely whitened, with limited evidence of convincing local weather details. Although the method improves consistency, the demonstrated effects do not sufficiently establish that it exploits the video backbone's capabilities to produce realistic weather appearance, limiting its practical value for urban asset augmentation.
  2. The evaluation does not adequately establish weather realism. The reference images used for PSNR and SSIM are themselves model-generated rather than real aerial captures under rain or snow. These metrics primarily assess agreement between generated views and the optimized 3DGS, rather than the realism of the weather transformation. CLIP-based semantic alignment also provides limited evidence of detailed visual realism. Independent perceptual evaluation would strengthen the claims.
  3. Insufficient dataset documentation. Training targets are generated by the original Cosmos model and manually filtered, but the paper does not clearly report the number of scenes and videos, the filtering retention rate, or the separation between training and test scenes. This makes it difficult to assess whether the method learns generalizable consistency or adapts to a limited set of urban scenes and weather appearances.
Questions for authors
  1. Please clarify the dataset size, filtering retention rate, and scene-level training/test split, including whether test scenes are unseen during fine-tuning.
  2. Could the authors provide additional evidence or discussion on how well the method captures local weather details and their visual realism?
2 comments

Response to Reviewer kpTF

We thank the reviewer for recognizing the loop-closure mechanism and for the pointed critique of weather realism, which we take seriously. We address each weakness and question below. The protocols are described in the supplementary.

W1 and Q2: Limited realism and lack of local weather details

We agree the current visualizations can read as a globally consistent stylization, and we address this with independent perceptual evidence plus local-detail visualization:

  1. User study. We conducted a 2AFC user study on unseen test scenes over three dimensions: realism (looks like genuine rain/snow rather than a global filter), multi-view consistency (turntable video), and geometry preservation. G-WADE is preferred over the strongest baseline in 62% of realism and 59% of consistency comparisons (Supplementary Section F). This is independent of the generated-reference metrics.
  2. Local-detail zoom-in. We provide zoom-in crops of snow accumulation on rooftop edges, and shadow/illumination variation, so the local weather details (rather than a global tint) are directly visible (Supplementary Section H).

We also acknowledge the fair point that our SH-only, geometry-frozen representation captures surface weather effects (wetness, snow cover, overcast illumination), and we believe the model can be readily extended to other surface-level weather conditions such as fog and haze with appropriate training data. Dynamic particles are orthogonal to our static surface-appearance objective and are left to future work (Supplementary Section G).

W2: Generated references and the realism of the evaluation

We agree that PSNR/SSIM against model-generated references measures agreement, not realism. We now make the evaluation's role explicit and complement it:

  1. What PSNR/SSIM measure. They measure intra-method multi-view consistency (how well a single 3D representation explains a method's own held-out views, also see Supplementary Section H for visualization).
  2. Independent realism. The user study is the independent perceptual evaluation the reviewer requested, and it is performed on held-out scenes. (Supplementary Section F)

W3 and Q1: Dataset documentation

We report the full statistics in Supplementary Section A:

  1. Size: 4268 single-building + 5984 multi-building scenes. 24772 rendered clips to 19852 retained (retention 80.14%).
  2. Filtering criteria: remove dynamic objects and exposure or color artifacts.
  3. Train/test split: the two sets of test scenes are completely disjoint from the fine-tuning scenes and are never seen during fine-tuning. For each test scene, we split the views into training and test sets at a 4:1 ratio (train:test).

We believe the added user study, local-detail visualizations, clarified/shared-reference evaluation, and full dataset documentation substantively address the realism and documentation concerns.

I appreciate the authors’ clarifications and additional experiments. Unfortunately, my main concern remains unresolved, as the authors have yet to address it in sufficient depth. A geometry-preserving assumption should not be used to simplify the problem, since weather conditions such as snow can require geometric changes to the scene. Overall, the paper addresses a relatively simple problem under a limited set of conditions, so I have decided to keep my original score.

Reviewer_LcnFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes G-WADE, which edits a 3D Gaussian Splatting scene into appearances under different weather condition. Using the synthesized weather videos, the authors fine-tune Cosmos-Transfer2.5's depth and edge control branches with a frame-latent conditioning. The generated consistent views are then distilled into the spherical-harmonic parameters of 3D GS using optimization.

Critical strengths
  1. The problem is well-motivated: Existing video generation methods need not generate consistent multi-view. This paper first fine-tunes the style-transfer video model to generate consistent views and then distill them back into 3D GS.
  2. The improvements, especially the Multi-building setting in Table 1, are promising and consistent across various metrics, with +1.40 dB PSNR and +0.085 SSIM for multi-building scenes, and +0.93 dB and +0.026 SSIM for single buildings over the baseline.
Critical weaknesses
  1. Lack of training and evaluation details: I didn't see the supplementary materials in the submitted manuscript. Many details including the number of scenes, training clips, evaluation scenes, or held-out views are not reported.
  2. Evaluation setting requires further clarification: In L.322, the authors mentioned that the reference images are “independently generated stylized images.” For comparison against other baselines, is every method using its own targets or shares a reference set?
Questions for authors
  1. The experiments cover only rain and snow. Have the authors evaluated the method under other weather conditions? Additional results would help clarify its generalizability.
  2. For Figure 3, the authors claim improved multi-view consistency. However, this is difficult to assess from the individual views shown. It would be clearer if you can highlight the inconsistency parts.
  3. In the snowy examples in Figure 3, WeatherCity produces snow-covered roads, whereas G-WADE leaves roads largely clear despite snow coverage on surrounding buildings. Visually, WeatherCity’s results appear more plausible to me in these examples.
1 comment

Thank you for your comments. We have added supplementary material and respond to each weakness and question below.

W1: Missing training/evaluation details and supplementary material

We have added supplementary material (Supplementary Section A) with details of the training data and dataset split. We clarify that the test data are completely separate from the fine-tuning data and were not used during fine-tuning. The supplementary material also includes more detailed visual comparisons of multi-view consistency across methods.

W2 and Q2: Reference images and multi-view consistency

Reference images. Each method uses its own generated reference images. For each test viewpoint, we first render the original scene. Each method then uses its own generation process to produce a weather-edited reference image from this input. We compare the corresponding optimized 3DGS rendering with this reference image.

To make the comparison fair, we use the same original input images, test viewpoints, and evaluation procedure for all methods. PSNR and SSIM therefore measure how well each method’s 3DGS results match its own generated targets. These metrics do not directly measure weather realism.

Multi-view consistency. We agree that individual views in Figure 3 do not clearly show the differences in consistency. We have added more detailed visual comparisons in the supplementary material (Supplementary Section H). These comparisons show more clearly that our method maintains better consistency across viewpoints than the other methods.

Q1: Generalization to other weather conditions

We agree that the current fine-tuned model only supports rain and snow. This is a limitation of our current work, and we have not yet evaluated the model under other weather conditions (Supplementary Section G). A similar generation pipeline could be used for other weather types. In future work, we will extend the model to more weather conditions and improve its generalization ability.

Q3: Snow coverage in the examples in Figure 3

We understand the concern about the snowy examples. WeatherCity tends to make most objects in these images white, including building sides and roads. However, a larger white area does not necessarily mean better weather effects.

In our results, snow is clearly visible on buildings, trees, and sidewalks, while there is less snow in the middle of the roads. This preserves more colors and scene details, and the lane markings remain clearly visible. In comparison, WeatherCity produces a more uniform white appearance.

For urban scene generation, we believe that preserving colors and useful scene details is also an important part of visual quality. We will clarify this point in the discussion of Figure 3.

Reviewer_qEgXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies rain and snow appearance editing for urban 3DGS scenes while keeping geometry fixed. It fine-tunes the depth and edge control branches of Cosmos-Transfer2.5 with first-frame and first-last-frame conditioning. It also designs camera trajectories that reuse generated views as anchors. The generated images are then used to update only the color coefficients of the original 3DGS. Experiments on single-building and multi-building scenes show better view consistency than the compared baselines.

Critical strengths

[S1] This paper works on a useful and clearly defined task. Editing weather appearance while keeping an existing 3D scene intact can reduce repeated data collection. The paper addresses a practical need for urban asset generation.

[S2] The proposed solution is a reasonable way to connect video generation and 3D editing. The trajectory design lets later video segments reuse existing views as endpoint anchors. The gains over the same-backbone Transfer2.5 + 3DGS baseline provide some evidence that the full pipeline helps.

Critical weaknesses

[W1] The technical novelty is limited. First-last-frame conditioning is already supported by models such as Wan2.1-FLF2V, Veo, Kling, and MiniMax H3. The main contribution in Sec. 4.3 therefore appears to be the tail-latent encoding strategy. However, it gives only a 0.8 dB gain over single-frame endpoint conditioning in the reported ablation, and its benefit for the final 3DGS results is not shown.

[W2] The results in Fig. 3 still show visual artifacts, which weaken the claim of realistic weather editing. The method only updates SH color coefficients while keeping geometry and opacity fixed. This may limit its ability to represent more complex weather effects, such as puddles with view-dependent reflections or snow accumulation that changes surface shape. The paper should discuss these limitations.

[W3] The data and evaluation setup lack important details. The paper did not report the number of scenes and videos, how training and test data are split, or the criteria used for manual filtering.

Questions for authors

[Q1] How many scenes and videos are included before and after filtering? What are the manual filtering criteria, and are training and test scenes separate?

[Q2] Can the current representation handle puddles with consistent reflections across views or snow accumulation that changes surface shape? Are dynamic weather effects supported, such as raindrops and snowflakes?

1 comment

Response to Reviewer qEgX

We thank the reviewer for the constructive feedback and for recognizing the task's practical value ([S1]) and the pipeline's soundness ([S2]). New experiments and protocols are in the supplementary.

W1: Novelty and the downstream benefit of tail-latent conditioning

We agree FL conditioning itself is not new (Wan2.1-FLF2V, Veo, Kling, MiniMax H3). Our contribution is not the conditioning concept but a VAE-aware tail-latent endpoint encoding combined with trajectory-aware generation scheduling for geometry-fixed appearance editing.

Tail-latent endpoint encoding. Wan2.1-FLF2V encodes one frame padded with three zeros. We encode five consecutive real frames and retain the second latent, exploiting the 1+4n temporal structure of the Wan VAE. This yields richer boundary temporal context and a more informative endpoint, and is co-trained with conditioning dropout so one model serves both F- and FL-modes. Following your suggestion, we add comparisons with Wan2.1-FLF2V and MiniMax H3 (Supplementary Section E). Wan2.1-FLF2V is the direct controlled comparison, since H3 uses a different video VAE and sequence-level conditioning. To further validate our endpoint encoding strategy, we will include a fine-tuned Wan2.1-FLF2V comparison under the same fine-tuning data and evaluation protocol in the final version.

Trajectory-aware scheduling. Orbit-and-descend and center-out grid scheduling reuse generated views as anchors, reducing drift and improving loop-closure consistency.

Regarding the 0.8 dB figure you cited: that is video-domain single-frame PSNR, not asset quality. We added the downstream 3DGS ablation (F-Model vs. FL-Model under identical geometry-fixed optimization), showing the tail-latent anchor removes the loop seam and improves final 3DGS PSNR by 1.4 dB (Supplementary Section D).

W2: Visual artifacts and the SH-only representation

We address realism with both evidence and an honest limitation discussion: User study (2AFC on realism, turntable-video consistency, geometry preservation) on unseen scenes, where G-WADE is preferred (Supplementary Section F). Limitations. Our SH-only, geometry-frozen model captures wetness, snow cover, and overcast illumination but by construction cannot synthesize view-dependent puddle reflections beyond the SH degree or shape-changing snow accumulation; we state these explicitly and show a wet-specularity proof-of-concept (higher SH order / explicit specular term), while shape-changing snow is out of scope (Supplementary Section G).

W3 and Q1: Data and evaluation details

We apologize for omitting these, full statistics and protocol are in Supplementary Section A: Scenes/clips: 4268 single-building and 5984 multi-building scenes; 24772 rendered clips, 19852 retained (80.14%). Filtering: retain clips with no dynamic objects (detector + manual check) and no exposure/color artifacts or large inter-frame flicker. Split: test scenes are fully disjoint from fine-tuning scenes and never seen during training; views split 4:1 (train:test).

Q2: Puddles / shape-changing snow / dynamic weather

Puddles/reflections: view-dependent reflections beyond the SH degree and shape-changing snow are out of scope under the geometry-fixed constraint (future work, Supplementary Section G). Dynamic weather: G-WADE targets static surface appearance; dynamic particles are orthogonal and addressed by WeatherEdit / Let-it-Snow, and combining their particle layer with our surface is future work (Supplementary Section G).

We believe these additions: reframed contributions, downstream tail-latent ablation, FLF2V/H3 comparisons, user study, and full dataset documentation can address the reviewer's concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.