Weather-Conditioned Depth Anything

Zhaoming Xu; Chan-Wei Hu; Kuan-Ru Huang; Zihao Zhu; Renjie Li; Yang Zhou; Zhengzhong Tu

cs.CVSep 4, 2026

Weather-Conditioned Depth Anything

Authors: Zhaoming Xu, Chan-Wei Hu, Kuan-Ru Huang, Zihao Zhu, Renjie Li, Yang Zhou, Zhengzhong Tu

Organizations: Texas A&M University, College Station, TX 77843, USA

Abstract

Monocular depth estimation foundation models, such as the Depth Anything series, have achieved remarkable performance across diverse domains. However, they still suffer from critical failures under adverse weather conditions, such as fog, rain, snow, or at night. To address this, we present Weather-Conditioned Depth Anything (DA-W), a framework that explicitly disentangles style from content for weather-robust depth estimation. Specifically, we introduce a Style Filter trained on a curated mix of real and synthetic degradation datasets to extract content-independent, degradation-aware weather embeddings. This style embedding is then injected into the Depth Anything backbone using a parameter-efficient, zero-initialized adapter. Such a lightweight modulation allows a single unified model to robustly adapt to diverse conditions, including fog, rain, snow, and low-light, while avoiding catastrophic forgetting of its core generalization abilities in normal conditions. We train the adapter using a pseudo-label distillation and alignment strategy. Our comprehensive experiments demonstrate that our proposed DA-W achieves state-of-the-art robust depth estimation, improving AbsRel by an average of 3.7% on our curated weather benchmarks, while matching or slightly outperforming performance on standard clean benchmarks. Our project page is available at https://zhaoming-tamu.github.io/WCDA/.

Figures & tables

Figure 2: Performance

\delta_{1}\uparrow

comparison across three datasets : real adverse weather, synthetic adverse weather, and clean weather. Values are normalized per dataset (Baseline = 0.6).

Panel of Figure 3 — Figure 3: Feature space comparison for style characterization. Depth Anything v2 [ 61 ] CLS tokens fail to separate weather conditions, while our style filter features form distinct clusters.

Figure 4: Overview of our proposed DepthAnything-W. (a) The training pipeline: clean and degraded images pass through the Depth Anything backbone, with weather embeddings from the Style Filter injected via AdaLN-Zero at multiple decoder scales to produce weather-conditioned depth maps supervised by $\mathcal{L}_{dis}$ and $\mathcal{L}_{align}$ . (b) The Style Filter extracts weather embeddings from images via multi-scale Gram matrices, using contrastive learning to cluster by weather type while maintaining scene invariance and real-synthetic alignment.

Method	Encoder	NuScenes-night		RobotCar-night		DS-rain		DS-cloud		DS-fog
Method	Encoder	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$
DynaDepth [ 63 ]	ResNet	0.381	0.394	0.512	0.294	0.239	0.606	0.172	0.608	0.144	0.901
EC-Depth [ 47 ]	ViT-S	0.243	0.623	0.228	0.552	0.155	0.766	0.158	0.767	0.109	0.861
STEPS [ 64 ]	ResNet	0.252	0.588	0.350	0.367	0.301	0.480	0.252	0.588	0.216	0.641
robustdepth [ 42 ]	MonoViT	0.260	0.597	0.311	0.521	0.167	0.755	0.168	0.775	0.105	0.882
weather-depth [ 53 ]	ViT-S	–	–	–	–	0.158	0.764	0.160	0.767	0.105	0.879

Table 1: Zero-shot relative depth estimation on real adverse-weather benchmarks. “–” denotes unavailable results (not reported or not supported by the released model, e.g., MWFormer on NuScenes-night and RobotCar-night).

Method	Encoder	Dark		Snow		Fog		Motion
Method	Encoder	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$
DepthPro [ 2 ]	ViT–S	0.145	0.793	0.197	0.685	0.087	0.926	0.170	0.746
DepthAnything v1 [ 60 ]	ViT–S	0.129	0.826	0.113	0.870	0.092	0.912	0.133	0.822
DepthAnything v2 [ 61 ]	ViT–S	0.130	0.832	0.115	0.872	0.097	0.905	0.127	0.840
DepthAnything v3 [ 31 ]	ViT–S	0.180	0.737	0.182	0.715	0.119	0.858	0.196	0.698
DepthAnything-AC [ 48 ]	ViT–S	0.130	0.834	0.114	0.873	0.096	0.906	0.126	0.841

Table 2: Zero-shot relative depth estimation on KITTI-C (synthetic corruptions). “–” denotes unavailable results.

Panel of Figure 5 — Figure 5: Qualitative comparison on various synthetic weather degradations. Our method (DA-W) produces more accurate and stable depth predictions under challenging synthetic weather conditions. The red boxes highlight regions where DA-W maintains clearer structural details and smoother depth transitions, while other methods suffer from noisy artifacts and depth discontinuities.

Method	Encoder	KITTI [ 17 ]		NYU-D [ 45 ]		Sintel [ 5 ]		ETH3D [ 43 ]		DIODE [ 52 ]
Method	Encoder	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$
DynaDepth [ 63 ]	ResNet	0.085	0.916	0.212	0.630	0.458	0.391	0.192	0.703	0.180	0.737
EC-Depth [ 47 ]	ViT-S	0.075	0.931	0.220	0.606	0.447	0.405	0.181	0.779	0.172	0.760
STEPS [ 64 ]	ResNet	0.204	0.683	0.220	0.606	0.458	0.375	0.197	0.699	0.214	0.652
robustdepth [ 42 ]	MonoViT	0.076	0.929	0.174	0.718	0.411	0.436	0.185	0.776	0.182	0.741
weather-depth [ 53 ]	ViT-S	0.075	0.932	0.167	0.733	0.419	0.443	0.181	0.771	0.170	0.760

Table 3: Relative depth estimation results on general benchmarks. Our method preserves DepthAnything’s performance on clean scenes. Best/second best are bold / underlined .

Corruption	DA v1 [ 60 ]		DA v2 [ 61 ]		DA v3 [ 31 ]		DA-AC [ 48 ]		DA-W (Ours)
Corruption	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$	$\delta_{1}\uparrow$	AbsRel $\downarrow$
Clean	0.937	0.082	0.934	0.083	0.912	0.094	0.934	0.083	0.937	0.080
Fog+Rain	0.708	0.192	0.726	0.185	0.636	0.225	0.731	0.181	0.750	0.173
Fog+Snow	0.681	0.205	0.712	0.192	0.481	0.303	0.731	0.179	0.734	0.177
Fog+Night	0.715	0.185	0.738	0.178	0.640	0.233	0.742	0.175	0.747	0.171
Rain+Snow	0.671	0.209	0.754	0.171	0.627	0.228	0.754	0.167	0.771	0.158

Table 4: Mix-weather benchmark on KITTI Eigen split with composed corruptions (RoboDepth [ 26 ] ). We report

\delta_{1}\uparrow

and AbsRel

\downarrow

. Best/second best are bold / underlined .

Config (✓ = enabled)				Zero-shot (AbsRel $\downarrow$ / $\delta_{1}\uparrow$ )							Overall
Encoder FT	Decoder FT	Real Deg.	Weather Inject.	NS-N	RC-N	DS-Rain	DS-Fog	NYU-D	KITTI	KITTI-Snow	Avg. rank $\downarrow$
✓	✓	–	–	0.194 /0.727	0.245/0.493	0.122 / 0.842	0.098 /0.898	0.056/0.968	0.084/0.932	0.109 / 0.888	2.89
–	✓	–	–	0.190 / 0.741	0.248/0.495	0.123/0.840	0.102/ 0.902	0.050 / 0.974	0.081 / 0.936	0.113/0.877	2.68
–	✓	✓	–	0.195/0.731	0.243 / 0.507	0.121 / 0.842	0.100 / 0.901	0.050 / 0.974	0.081 / 0.936	0.112/0.878	2.29
–	✓	✓	✓	0.194 / 0.737	0.239 / 0.513	0.123/ 0.842	0.101/0.896	0.051 / 0.973	0.080 / 0.937	0.107 / 0.884	2.14

Table 5: Ablation across training and weather-conditioning choices. Metrics are AbsRel

\downarrow

\delta_{1}\uparrow

. Avg. rank

\downarrow

is computed over all 14 reported metrics with ties averaged. Bold and underlined denote the best and second-best values for each metric, respectively.

Panel of Figure 6 — Figure 6: Qualitative comparison on mixed degradations. Each row shows real-world images of snow at night.

Variant	Real night	DrivingStereo	KITTI-C	Global
Full objective	0.217	0.108	0.110	0.117
w/o $L_{\rm dis}$	0.535	0.409	0.337	0.352
w/o $L_{\rm pair}$	0.222	0.109	0.114	0.120
w/o $L_{\rm aug}$	0.215	0.109	0.109	0.116
$L_{\rm dis}$ only	0.223	0.110	0.111	0.118
RGB restore ( $\lambda_{\rm rgb}=0.1$ )	0.214	0.106	0.111	0.118

Table 6: Controlled loss ablations. Results are AbsRel

\downarrow

. All rows use the same 20% clean/synthetic subset, full real-weather data, seed, and evaluation protocol unless noted. Real night averages RC-N and NS-N; DrivingStereo averages DS-Rain and DS-Fog; KITTI-C averages Fog, Snow, and Dark. Global is the pooled per-image mean over evaluated weather sets.

Panel of Figure 7 — Figure 7: Visualization of decoder feature modulation by weather-aware conditioning. From left to right: input image, decoder features before modulation, decoder features after modulation, and the absolute difference map.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: t-SNE Visualization of real and synthetic dataset distributions.

Panel of Figure 9 — Figure 9: Qualitative Depth Map Comparison Under Real-World Weather Degradations.

Panel of Figure 10 — Figure 10: Qualitative Depth Map Comparison Under Synthetic Degradations.

Panel of Figure 11 — Figure 11: Visualization of decoder feature modulation by weather-aware conditioning. The same scene is shown under clean (top), foggy (middle), and snowy (bottom) conditions, where degradations are synthetically applied to enable controlled comparison. From left to right: input image, decoder features before conditioning (zero condition baseline), decoder features after conditioning (with learned weather condition), and absolute difference (hot colormap: darker $\rightarrow$ smaller, brighter $\rightarrow$ larger).

Panel of Figure 12 — Figure 12: Feature space visualization for CLIP-style alignment variants. Left : Variant (i) with prompt-driven image-text alignment shows severe overlap between text-adapted rain (blue) and clean weather (pink) embeddings, while real rain (green) forms separate clusters. Right : Variant (ii) with direct image-image alignment shows that synthetic rain (blue) and clean weather (pink) form relatively distinct clusters, but real rain (green) occupies entirely different regions in the embedding space.

Depth Estimation

Depth estimation infers scene distance from images and sensors, enabling 3D reconstruction, navigation, and spatial understanding.

See all 198 papers

OpenDepth: Repurposing Image Editing Priors for Robust Open-World Monocular Depth Estimation
Oct 13, 2026Mengke Zhang, Tianyi Yang, Xinzhuo Li +1Monocular Depth Estimation Diffusion Models
International School, Beijing University of Posts and Telecommunications, Beijing, China · International School, Queen Mary University of London, Beijing, China · School of Computer Science, Beijing University of Posts and Telecommunications, Beijing, China
AURORA-KITTI: Any-Weather Depth Completion and Denoising
Mar 16, 2026Yiting Wang, Tim Brödermann, Hamed Haghighi +4Depth Completion Multimodal Robustness
Warwick Manufacturing Group, University of Warwick, Coventry, CV4 7AL, United Kingdom · School of Engineering and Materials Science, Queen Mary University of London, London, E1 4NS, United Kingdom · Computer Vision Laboratory, ETH Zürich, Zurich, 8057, Switzerland +1
Boosting Robustness for All-Weather Self-Supervised Depth Estimation in Autonomous Driving
Jul 23, 2026Mengshi Qi, Xiaoyang Bi, Xianlin Zhang +1Camera-Radar Fusion Depth Estimation
State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, China

Oct 13, 2026J/K move · Enter open · S save

OpenDepth: Repurposing Image Editing Priors for Robust Open-World Monocular Depth Estimation

Mengke Zhang, Tianyi Yang, Xinzhuo Li, Anlong Ming

International School, Beijing University of Posts and Telecommunications, Beijing, China · International School, Queen Mary University of London, Beijing, China · School of Computer Science, Beijing University of Posts and Telecommunications, Beijing, China

Monocular Depth Estimation Diffusion Models

Open paper AskPDF arXiv

Recently, the integration of generative priors has revolutionized Monocular Depth Estimation (MDE), achieving state-of-the-art (SOTA) performance with remarkably high data efficiency. However, although methods based on Text-to-Image (T2I) models excel in general scenes, their efficacy degrades sharply when confronted with open-world environmental perturbations, such as illumination variations and adverse weather. Since T2I priors are optimized for semantic synthesis rather than preserving pixel-level structure, they struggle to decouple geometry from appearance under severe visual shifts, easily triggering structural hallucinations. Therefore, our core insight is to model open-world depth estimation as consistent geometric inference across multi-style images, fully leveraging the powerful structure-preserving editing priors of Image-to-Image (I2I) models. Based on this, we propose OpenDepth, a robust open-world MDE framework that introduces "Intra-Domain Awareness" -- the capability to map diverse stylistic inputs sharing identical geometry into a unified latent space. To achieve this, we first align shallow features within the Diffusion Transformer (DiT) to anchor degraded observations to unified geometric representations. Additionally, we introduce a joint RGB latent reconstruction task, requiring the model to recover clean RGB information from degraded inputs. This multi-stage strategy effectively decouples environmental style interference, compelling the model to focus strictly on underlying geometric recovery. Extensive evaluations show that OpenDepth sets new SOTA records for robust zero-shot depth estimation across standard and extreme weather benchmarks, demonstrating exceptional structural integrity and adaptability.

Abstract

Depth Estimation

Explore similar work

OpenDepth: Repurposing Image Editing Priors for Robust Open-World Monocular Depth Estimation

AURORA-KITTI: Any-Weather Depth Completion and Denoising

Boosting Robustness for All-Weather Self-Supervised Depth Estimation in Autonomous Driving