Realistic rain synthesis is a persistent challenge for vision systems operating in adverse weather. Existing methods primarily focus on generating rain streaks, often neglecting consequent surface effects such as reflections and puddles, which are critical for perceptual realism and robustness of downstream vision tasks. We introduce CascadeRS, a rain synthesis framework that leverages depth and segmentation priors to render scene-consistent rain and surface effects. In addition to rain streaks, CascadeRS produces reflections and puddles aligned with scene semantics through a physics-inspired zero-shot, controllable generative pipeline. It produces photorealistic rainy urban scenes, with high perceptual realism, which we validate qualitatively with an extensive user study and quantitatively. We produce the Cityscapes-CascadeRS dataset, to study downstream task performance under real rainy conditions. Our results establish CascadeRS as a scalable and realistic rain-generation framework for evaluating and enhancing the robustness of vision in adverse rainy weather. Code and dataset are available via https://github.com/nis-research/CascadeRainSynthesis
Figures & tables
Figure 1: Rain rendering and its impact on perception. In the first two columns we show the output of different rain rendering methods. Our CascadeRS approach produces the most realistic puddles and reflections, which are absent in prior synthetic datasets. The third and fourth columns illustrate how realistically rendered rain affects downstream vision tasks, such as semantic segmentation and object detection, where rain causes visible performance degradation. Background images are from [ 5 ] .
Figure 2: CascadeRS overview. Overall pipeline (top row): depth and semantic cues guide sequential surface effects and rain synthesis. The colored boxes below illustrate the detail of each stage: reflection, puddle, fog, and streak generation.
Method
Rain
Fog
Pudd.
Refl.
TF
Control
Dataset agnostic
Synth.
PBR [ 11 ]
✓
✗
✗
✗
✓
✓
✓
Physics
CS-Rain [ 13 ]
✓
✓
✗
✗
✓
✓
✗
Physics
TPSeNCE [ 62 ]
✓
✓
✓
✓
✗
✗
✓
GAN
CascadeRS
✓
✓
✓
✓
✓
✓
✓
Physics
Table 1: Comparison of rain synthesis. CascadeRS combines multiple weather effects with a training-free (TF) approach.
Figure 3: Realism ratings per method , comparing user study (plain bars) and LLM-as-judge (striped bars). The stacked bars indicate score percentage (1: not realistic, to 5: very realistic). Detailed rating percentages are in Appendix M .
Table 5
Figure 4: mIoU across datasets. Boxplots show per-image mIoU aggregated over all images and, where applicable, rain intensities. ACDC [ 34 ] contains real rainy images.
ACDC
CS-val
Model
Metric
CS
CS-R
CRS
CS
CS-R
CRS
SegFormer-B2
mIoU ↑
0.4259
0.3882
0.4265
0.6813
0.6886
0.6679
FWIoU ↑
0.7824
0.7753
0.8059
0.9034
0.9015
0.9016
GCNet
mIoU ↑
0.3546
0.4047
0.4019
0.7156
0.7142
0.7086
FWIoU ↑
0.7241
0.7760
0.7958
0.9106
0.9093
0.9077
Table 4: Segmentation results of SegFormer-B2 [ 46 ] and GCNet [ 48 ] on ACDC and Cityscapes validation sets. Models are retrained on original Cityscapes (CS), Cityscapes mixed with Cityscapes-Rain (CS-R), and mixed with CascadeRS (CRS).
ACDC
CS-val
Metric
CS
CS-R
CRS
CS+CRS
CS
CS-R
CRS
CS+CRS
AP ↑
18.03
16.70
17.68
18.06
32.18
32.76
31.99
32.17
AP@50 ↑
43.41
39.41
41.79
40.46
61.16
62.11
60.64
61.68
Table 5: Object detection results of DINO Teacher on ACDC and Cityscapes, trained on the original Cityscapes (CS), Cityscapes mixed – by a 91/9% ratio – with Cityscapes-Rain (CS-R) and with CascadeRS (CRS), and Cityscapes training set plus its complete augmented version with CascadeRS.
Figure 5: FID scores of various depth and segmentation model compositions. M = MiDaS [ 32 ] , DA = DepthAnything [ 49 ] , SegF = SegFormer [ 46 ] , and FS = FastSeg [ 53 ] . Replacing the depth model has a negative effect on the FID score.
Figure 6: Segmentation results for the ablation of CascadeRS components under varying rain intensities.
Figure 7: Parameter sensitivity analysis. Effect of five CascadeRS parameters on FID, LPIPS, mIoU, and AP. Fog = fog density, Refl. = puddle reflection, Pers. = puddle persistence, Res. = puddle resolution, and Rain = rain intensity.
Figure 8: Generalizability of CascadeRS: rain effects synthesized on images from datasets with different road-view perspective. Additional examples are in Appendix H .
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: Rain-intensity sweep on a (cropped) Cityscapes sample [ 5 ] . Subfigures (a)–(k) correspond to intensities 0–100. Increasing intensity leads to denser streaks, thicker fog, and stronger atmospheric degradation in CascadeRS.
Figure 10: Progressive illustration of the CascadeRS components across multiple Cityscapes [ 5 ] scenes. Each row corresponds to a different image, showing the sequential addition of (1) rain streaks, (2) wet-surface reflections, and (3) puddles guided by scene geometry. Together, these effects yield realistic and coherent rainy appearances.
Figure 11: Zoomed ablation comparisons across two scenes. Each block shows the original image, rain-only rendering, wet-surface reflections, and puddles, together with a magnified region highlighting differences in surface interaction and consistency.
Figure 12: Qualitative comparison of rain synthesis methods. Each column shows a different rendering method, while each row presents a distinct Cityscapes scene. PBR [ 11 ] and Cityscapes-Rain [ 13 ] provide streak overlays but no surface interactions. TPSeNCE [ 62 ] introduces wetness effects but lacks clear streaks and fog. CascadeRS produces depth-aligned streaks, realistic reflections, puddles, and fog.
Figure 13: Rain synthesis with GenAI methods. The original image (first column) is processed with Stable Diffusion XL (second column) and three times with Gemini (Nano Banana) to synthesize rain effects. For all generations, the same text prompt was used.
Method
Intensity
Distribution similarity
Content preservation
FID ↓
FID ↓
KID ↓
LPIPS ↓
SSIM ↑
PSNR ↑
ACDC
RID
ACDC
No Rain
–
117.80
131.96
0.0708
0.000
1.000
∞
PBR [ 11 ]
20 mm
115.58
102.76
0.0725
0.428
0.827
21.71
40 mm
146.61
117.35
0.1219
0.550
0.695
19.01
60 mm
194.80
154.65
0.1962
0.621
0.586
17.61
Appendix
Table 6: Comparison of synthetic rain methods using distributional and paired similarity metrics. Distribution similarity to real rain is evaluated using FID against ACDC [ 34 ] and RID [ 23 ] , and KID against ACDC. Content preservation is evaluated relative to the corresponding original Cityscapes images using LPIPS, SSIM, and PSNR. Lower FID and KID indicate greater distributional similarity to the real-rain reference dataset, while lower LPIPS and higher SSIM and PSNR indicate greater perceptual and structural similarity to the corresponding original scene.
Effect
Intensity
FID (ACDC) ↓
FID (RID) ↓
Streaks
10
112.36
121.73
20
110.29
113.64
30
108.71
107.41
40
108.66
102.72
50
111.25
100.15
60
115.27
99.94
Appendix
Table 7: FID comparison of the individual CascadeRS rain effects. FID is evaluated against real rain from ACDC [ 34 ] and RID [ 23 ] . Lower values indicate greater distributional similarity to real rain.
Metric / Class
ACDC (Train)
Cityscapes (Validation)
CS Train
Occ. [%]
CS
CascadeRS
Δ
Occ. [%]
CS
CascadeRS
Δ
Occ. [%]
Mean
Std.
Mean
Std.
(a) SegFormer-B2
mIoU ↑
–
0.4259
0.4215
0.0056
− 0.0044
–
0.6813
0.6705
0.0036
− 0.0108
–
FWIoU ↑
–
0.7824
0.8003
0.0036
+0.0179
–
0.9034
0.9004
0.0007
− 0.0030
–
Sky
28.35
0.9713
0.9550
0.0065
− 0.0163
3.35
0.9432
0.9415
0.0008
− 0.0017
4.019
Appendix
Table 8: Cross-dataset segmentation comparison. Per-class IoU on ACDC train and Cityscapes validation sets for SegFormer-B2 [ 46 ] and GCNet [ 48 ] , trained on Cityscapes (CS) with and without CascadeRS augmentation. Mean and standard deviation over five random seeds are reported separately for SegFormer-B2; GCNet results are from a single training run. Classes are ordered by their occurrence in ACDC; the last column lists their relative frequency in the Cityscapes training set. The three largest improvements on ACDC are shown in bold , and the two largest reductions are underlined .
Method
Intensity
AP
AP 50
AP 75
AP s
AP m
AP l
ACDC (real rain)
Reference
19.0782
41.1327
15.1855
3.7629
21.4794
41.7014
Cityscapes
Clean
32.1823
61.2716
30.4232
5.8718
30.9993
57.3393
CascadeRS
10
30.8019
59.6321
27.9950
5.3069
28.9772
55.4775
20
29.7713
56.8627
28.9478
4.6032
26.9596
55.5098
30
28.6653
54.8198
27.3550
4.0520
25.6022
54.0948
40
27.3669
52.9954
24.2306
3.3321
24.1872
52.7352
Appendix
Table 9: Object detection performance under rainy conditions. Performance of the pretrained DINO Teacher model [ 19 ] on real and synthetic rainy images. Results are reported using COCO AP metrics.
Figure 14: Qualitative results of rain synthesis across different datasets. Each row shows four example renderings from our CascadeRS pipeline for the given dataset.
Implementation
No rain
10
20
30
40
50
60
70
80
90
100
FID-ACDC (500 reference images)
CascadeRS original (MiDaS + SegFormer)
117.80
99.08
99.60
101.72
106.74
113.42
122.43
132.34
143.39
155.80
164.76
CascadeRS (DepthAnything + SegFormer)
117.80
97.58
99.00
103.52
112.81
124.04
139.24
154.97
170.45
184.76
198.45
CascadeRS (MiDaS + FastSeg)
117.80
97.09
97.59
100.44
105.25
111.70
120.07
131.43
142.64
153.50
164.54
CascadeRS (DepthAnything + FastSeg)
117.80
95.79
96.85
102.04
111.22
123.59
139.01
154.52
171.26
187.08
201.27
FID-RID (2495 reference images)
Appendix
Table 10: FID comparison for CascadeRS module variants. Fréchet Inception Distance (FID) is reported across rain intensities, using ACDC and RID as reference datasets. Lower values indicate greater distributional similarity to the reference dataset.
Figure 15: Qualitative comparison of different pretrained depth and semantic segmentation model compositions. M = MiDaS, DA = DepthAnything, SegF = SegFormer, and FS = FastSeg. All images are rendered on rain intensity level 50.
Figure 16: UMAP visualization of DINOv3 embeddings for four CascadeRS implementations using different pretrained depth and semantic segmentation models. Each implementation contains 500 randomly sampled images across all rain intensities. The substantial overlap between the embedding distributions indicates that replacing MiDaS with DepthAnything and/or SegFormer with FastSeg has little effect on the overall feature distribution of the synthesized images.
Figure 17: Effect of synthesis parameters on image characteristics and downstream task performance. We vary rain intensity, fog density β , puddle resolution f0 , puddle persistence a , and puddle reflection λp,0 while keeping the remaining synthesis parameters at their default settings. Distribution similarity is evaluated using FID and KID with ACDC [ 34 ] as the reference distribution, while image fidelity with respect to the original Cityscapes images [ 5 ] is measured using LPIPS, SSIM, and PSNR. Downstream performance is evaluated using semantic segmentation mIoU with a pretrained SegFormer model [ 46 ] and object detection AP and AP@50 with a pretrained DINO-Teacher model [ 19 ] . Shaded regions indicate standard deviation where applicable; for mIoU, the standard deviation is estimated by bootstrapping.
Figure 18: Semantic segmentation performance across rain intensity levels. Full-dataset mIoU for SegFormer [ 46 ] , FastSeg [ 53 ] , and GCNet [ 48 ] at each available rain intensity. Dashed lines denote the corresponding performance on real rainy ACDC images; for TPSeNCE, which does not provide multiple intensity levels, ACDC performance is shown as reference points.
Method
Evaluator
1
2
3
4
5
PBR
Survey
28.4
16.7
21.3
21.6
12.0
Gemini
0.0
50.0
50.0
0.0
0.0
CS-Rain
Survey
10.4
16.6
23.1
35.4
14.6
Gemini
0.0
12.5
25.0
62.5
0.0
TPSeNCE
Survey
4.2
10.9
24.5
39.4
20.9
Gemini
0.0
12.5
0.0
75.0
12.5
Appendix
Table 11: Distribution of realism ratings. Percentage of ratings for each realism score (1: not realistic, to 5: very realistic), comparing the user survey and Gemini evaluation.
Figure 19: Example question of part 1 of the survey.
Figure 20: Example question of part 2 of the survey.
Figure 21: Example question of part 3 of the survey.
School of Information Engineering, Guangdong University of Technology, Guangzhou, 510006, China · School of Computer Science and Engineering, Nanjing University of Science and Technology