OmniRestore

OmniRestore: A Parameter-Efficient Framework for Universal Adverse-Weather Image Restoration

Judith N. Njoku Diksha Shukla

Dept. of Electrical Engineering & Computer Science, University of Wyoming, USA

CVPR Workshop 2026

Paper Supplemental GitHub Poster LinkedIn BibTeX

29.86 dB

PSNR on CDD-11 , #1 SoTA

2.6M

Inference parameters

+1.14 dB

Over OneRestore (ECCV'24)

6.3 ms

Latency per image

41%

LPIPS improvement

OmniRestore overview: adverse weather degradations and efficiency comparison

Visual Restoration Results

Qualitative comparison of restoration results across CDD-11, Snow100K, WeatherBench (rain and snow), and LOL benchmarks. From left to right: Input, SRResNet (SRR), SwinIR (Swin), Restormer (Resto.), PromptIR (Prompt), OneRestore (OneRe), M2Restore (MSrest), AdaIR, OmniRestore (OmniR), and Ground Truth (GT). OmniRestore consistently recovers clearer structural details, sharper edges, improved illumination, and more realistic textures under diverse adverse-weather degradations while maintaining lightweight real-time efficiency.

Qualitative comparison across CDD-11, Snow100K, WeatherBench, and LOL

Abstract

Reliable perception under adverse weather remains a significant challenge for vision systems. While state-of-the-art multi-weather image restoration models achieve strong performance, they often rely on large architectures with high parameter counts, limiting their utility for real-time or edge deployment. To address this efficiency gap, we propose OmniRestore, a lightweight multimodal framework specifically designed for adverse-weather image restoration across rain, fog, snow, and low-light conditions, including challenging composite degradations.

OmniRestore consists of two key components: (i) a multimodal weather-scene embedder that aligns CLIP-derived text prototypes with image features via a ResNet encoder, refined offline by a Kolmogorov-Arnold Network (KAN) adapter to produce compact, weather-discriminative condition embeddings; and (ii) a parameter-efficient restoration backbone that injects these embeddings through hybrid attention blocks integrating embedding-conditioned cross-attention, shifted window self-attention, and channel recalibration. Evaluated across CDD-11, Snow100K, LOL, and WeatherBench, OmniRestore achieves state-of-the-art restoration fidelity and perceptual quality, outperforming the leading unified model OneRestore by +1.14 dB PSNR and 41% in LPIPS on CDD-11. Notably, with only 2.6M inference-time parameters , a 56.5% reduction compared to OneRestore , OmniRestore offers a significantly more practical solution for resource-constrained and edge-deployment environments.

Contributions

CLIP-KAN Weather-Scene Embedder

Aligns CLIP text prototypes with visual features via a Kolmogorov-Arnold Network (KAN) adapter. Operates fully offline at inference , zero VLM, zero text encoder at test time. Achieves 98.46% weather classification accuracy.

Lightweight Hybrid Attention Backbone

U-Net with embedding-conditioned cross-attention, Swin shifted windows (W=8), and channel recalibration. FFN width pruning (factor 2.0) delivers ~70% parameter reduction, compensated by 5×5 DWConv kernels.

Adaptive Refinement Stage (ARS)

StyleModulation derives per-instance gain (γ) and bias (β) from condition embedding e. Enables degradation-specific color recalibration with negligible overhead. +41% LPIPS improvement over OneRestore on CDD-11.

Perceptual Composite Loss

Five-component objective: Smooth L1 + MS-SSIM + VGG-Contrastive + VGG-Perceptual + Deep Supervision. VGG contrastive explicitly pushes restored output away from degraded input and weather-degraded negatives.

Method

OmniRestore architecture: CLIP-KAN embedder and lightweight U-Net backbone

Fig. 2. Overview of OmniRestore. (1) Multimodal Embedder (left): ResNet + KAN adapter aligns visual features with CLIP text prompts offline, generating weather condition embedding e. (2) Lightweight Backbone (right): U-Net-style transformer injects e via cross-attention, concluding with ARS for final artifact removal.

Quantitative Results

Comparison on CDD-11 (11 degradation types: 4 single + 7 composite). Red = best, blue = second best. All One-to-Composite models retrained on CDD-11.

Category	Method	Venue & Year	PSNR ↑	SSIM ↑	#Params
One-to-One (Single Degradation)
	MIRNet	ECCV'20	25.97	0.8474	31.79M
	MPRNet	CVPR'21	25.47	0.8555	15.74M
	Restormer	CVPR'22	26.99	0.8646	26.13M
	NAFNet	ECCV'22	24.13	0.7964	17.11M
	SRUDC	ICCV'23	27.64	0.8600	6.80M
One-to-Many (Multi-degradation)
	AirNet	CVPR'22	23.75	0.8140	8.93M
	PromptIR	NeurIPS'23	25.90	0.8499	38.45M
	WGWNSNet	CVPR'23	26.96	0.8626	25.76M
One-to-Composite (Unified Restoration)
	OneRestore	ECCV'24	28.72	0.8821	5.98M
	AdaIR	ICLR'25	28.54	0.8779	28.77M
	M2Restore	TIP'25	27.38	0.8614	47.5M
	OmniRestore (Ours)	CVPRW'26	29.86	0.9244	2.60M

Ablation Studies

Step-by-step validation of each design choice on CDD-11. The KAN adapter outperforms an equivalent MLP by +0.93 dB PSNR and -0.0371 LPIPS.

Backbone Component Ablation

Baseline	FFN	5×5 DWConv	ARS	Params	PSNR	LPIPS
✓				5.10M	28.50	0.1821
✓	✓			2.25M	28.15	0.1956
✓	✓	✓		2.45M	29.10	0.1134
✓	✓	✓	✓	2.60M	29.86	0.0941

Cross-Module Embedding Comparison

Embedding Module	Accuracy	PSNR	LPIPS
OneRestore Embedder	0.8950	27.85	0.1603
CLIP-ResNet + MLP	0.8510	28.93	0.1312
CLIP-ResNet + KAN (Ours)	0.9846	29.86	0.0941

Poster

BibTeX

@InProceedings{Njoku_2026_CVPR,
  author    = {Njoku, Judith and Shukla, Diksha},
  title     = {OmniRestore: A Parameter-Efficient Framework for Universal Adverse-Weather Image Restoration},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  month     = {June},
  year      = {2026},
  pages     = {2236-2246}
}

Contact

If you have any questions about OmniRestore, feel free to reach out: jnjoku@uwyo.edu

Judith N. Njoku · University of Wyoming · LinkedIn