[CVPRW 2026] OmniRestore: A Parameter-Efficient Framework for Universal Adverse-Weather Image Restoration

Paper | Supplementary | Project Page | Poster | LinkedIn

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

Judith N. Njoku, Diksha Shukla Department of Electrical Engineering & Computer Science, University of Wyoming, USA

Accepted to CVPR 2026 Workshops, NTIRE.

---

Figure 1. OmniRestore addresses adverse-weather image restoration under rain, snow, fog, low-light, and composite degradations. Compared with single-task restoration models and existing unified restoration frameworks, OmniRestore achieves superior restoration quality while requiring only 2.6M parameters.

---

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 and 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, low-light conditions, and 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 on CDD-11, Snow100K, LOL, and WeatherBench, OmniRestore achieves state-of-the-art restoration fidelity and perceptual quality, outperforming the leading unified restoration framework OneRestore by +1.14 dB PSNR and 41% in LPIPS on CDD-11. Notably, OmniRestore requires only 2.6M inference-time parameters, representing a 56.5% reduction compared to OneRestore (5.98M), while simultaneously improving restoration quality.

These results demonstrate that accurate universal adverse-weather restoration can be achieved without sacrificing efficiency, making OmniRestore a practical solution for resource-constrained and edge-deployment environments.

---

News

• 2026.05.29: Code and pre-trained weights released.
• 2026.05.27: Project page released.
• 2026.05.15: OmniRestore published in the CVPR 2026 Workshop Proceedings.
• 2026.04.12: OmniRestore accepted to CVPR 2026 Workshops.

---

Highlights

• Universal restoration for rain, fog, snow, low-light, and composite weather degradations.
• Only 2.60M inference-time parameters.
• +1.14 dB PSNR improvement over OneRestore on CDD-11.
• 6.3 ms per-image inference latency on 256 × 256 inputs.
• CLIP-guided weather semantic conditioning with KAN refinement.
• No text encoder or VLM executed per image during inference.

---

Network Architecture

OmniRestore consists of:

1. Multimodal Weather-Scene Embedder Aligns degraded image features with CLIP-derived weather text prototypes using a ResNet image stream and KAN adapter.

2. Lightweight Restoration Backbone Uses embedding-conditioned cross-attention, shifted-window self-attention, compact FFN pruning, and channel recalibration.

3. Adaptive Refinement Stage Applies style modulation using condition-dependent gain and bias parameters to improve perceptual restoration quality.

---

Qualitative Results

Qualitative comparison across CDD-11, Snow100K, WeatherBench, and LOL. From left to right: Input, SRResNet, SwinIR, Restormer, PromptIR, OneRestore, M2Restore, AdaIR, OmniRestore, and Ground Truth.

---

Quick Start

Installation

git clone https://github.com/Judith989/omnirestore.git
cd omnirestore

conda create -n omnirestore python=3.10
conda activate omnirestore

pip install torch torchvision torchaudio
pip install timm einops pykan open_clip_torch lpips
pip install pillow numpy matplotlib opencv-python scikit-image pandas scikit-learn tqdm

---

Pre-trained Models

Please place the pre-trained weights in the corresponding folders.

Model	Path	Description
Restoration model	ckpts/best.ckpt	OmniRestore restoration backbone
Embedder model	logs/embedder_resnet18_bs64_optadamw_lr1.00e-05_cw1.0_conw0.1_temp0.07_20260101-190539.pt	ResNet18 CLIP-KAN weather-scene embedder

---

Directory Structure

OmniRestore/
├── ckpts/
│   └── best.ckpt
├── data/
│   └── cdd11/
│       └── splits/
├── logs/
│   └── embedder_resnet18_bs64_optadamw_lr1.00e-05_cw1.0_conw0.1_temp0.07_20260101-190539.pt
├── model/
│   ├── embedder.py
│   ├── wadt_net_shift.py
│   └── wadt_net_no_shift.py
├── utils/
│   ├── dataset_loader.py
│   ├── losses.py
│   └── utils.py
├── eval_embedder.py
├── test_only_images.py
└── test_with_text.py

---

Dataset Preparation

Please prepare the CDD-11 dataset as follows:

data/
└── cdd11/
    ├── train/
    ├── val/
    ├── test/
    └── splits/

Additional datasets used for evaluation include:

• Snow100K
• LOL
• WeatherBench

---

Inference

1. Multimodal Restoration, Image + Text

python test_with_text.py \
    --embedder-model-path logs/embedder_resnet18_bs64_optadamw_lr1.00e-05_cw1.0_conw0.1_temp0.07_20260101-190539.pt \
    --best-ckpt ckpts/best.ckpt \
    --cdd-root path/to/your/cdd11 \
    --output ./results_full_test_with_text \
    --bs 4 \
    --num-works 4 \
    --image-size-h 224 \
    --image-size-w 224

2. Blind Restoration, Image Only

python test_only_images.py \
    --embedder-model-path logs/embedder_resnet18_bs64_optadamw_lr1.00e-05_cw1.0_conw0.1_temp0.07_20260101-190539.pt \
    --best-ckpt ckpts/best.ckpt \
    --cdd-root path/to/your/cdd11 \
    --output ./results_full_test_only_images \
    --bs 4 \
    --num-works 4 \
    --image-size-h 224 \
    --image-size-w 224

---

Embedder Evaluation

To evaluate semantic clustering and weather classification accuracy:

python eval_embedder.py \
    --cdd_test_root path/to/your/cdd11 \
    --checkpoint logs/embedder_resnet18_bs64_optadamw_lr1.00e-05_cw1.0_conw0.1_temp0.07_20260101-190539.pt \
    --emb_out_dir embedder_eval_out \
    --backbone resnet18 \
    --stage 1

Depending on your local configuration, run --stage 1 to extract embeddings and --stage 2 to compute metrics and generate t-SNE/confusion matrix plots.

---

Performance

CDD-11

Method	Venue	PSNR ↑	SSIM ↑	Params
OneRestore	ECCV 2024	28.72	0.8821	5.98M
M2Restore	TIP 2025	27.38	0.8614	47.5M
AdaIR	ICLR 2025	28.54	0.8779	28.77M
OmniRestore	CVPRW 2026	29.86	0.9244	2.60M

Inference Latency

Method	Params	Latency
M2Restore	47.5M	95.4 ms
AdaIR	28.77M	42.1 ms
OneRestore	5.98M	14.2 ms
OmniRestore	2.60M	6.3 ms

---

Citation

If you find this work useful, please cite:

@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

For questions, please contact:

Judith N. Njoku Website: https://www.judithnnjoku.me LinkedIn: https://www.linkedin.com/in/judith989/