DehazeSNN

U-Net-Like Spiking Neural Networks for Single Image Dehazing

[πŸ“„ Paper] [πŸ’» GitHub]

DehazeSNN integrates a U-Net-like encoder-decoder architecture with Spiking Neural Networks (SNNs), using an Orthogonal Leaky-Integrate-and-Fire Block (OLIFBlock) for efficient cross-channel communication. This yields competitive dehazing quality with fewer parameters and MACs compared to Transformer-based methods.

Available Checkpoints

Revision Size Dataset Load Command
main L RESIDE-ITS DehazeSNNHub.from_pretrained("FengShaner/DehazeSNN")
l-reside-its L RESIDE-ITS DehazeSNNHub.from_pretrained("FengShaner/DehazeSNN", revision="l-reside-its")
l-reside-ots L RESIDE-OTS DehazeSNNHub.from_pretrained("FengShaner/DehazeSNN", revision="l-reside-ots")
l-reside-6k L RESIDE-6k DehazeSNNHub.from_pretrained("FengShaner/DehazeSNN", revision="l-reside-6k")
l-rs-haze L RS-Haze DehazeSNNHub.from_pretrained("FengShaner/DehazeSNN", revision="l-rs-haze")
m-reside-its M RESIDE-ITS DehazeSNNHub.from_pretrained("FengShaner/DehazeSNN", revision="m-reside-its")
m-reside-6k M RESIDE-6k DehazeSNNHub.from_pretrained("FengShaner/DehazeSNN", revision="m-reside-6k")
m-rs-haze M RS-Haze DehazeSNNHub.from_pretrained("FengShaner/DehazeSNN", revision="m-rs-haze")

Quick Start

Installation

# Clone the DehazeSNN repository (for model code + custom CUDA kernels)
git clone https://github.com/HaoranLiu507/DehazeSNN.git
cd DehazeSNN

# Create environment (requires CUDA 12.x)
conda create -n DehazeSNN python=3.11 -y
conda activate DehazeSNN

# Install PyTorch with CUDA 12.1
conda install pytorch=2.1.2 torchvision pytorch-cuda=12.1 -c pytorch -c nvidia -y

# Install dependencies
pip install huggingface_hub safetensors cupy-cuda12x timm

Load and Run Inference

import torch
from PIL import Image
import numpy as np

# Import from the cloned repository
from models.hub import DehazeSNNHub

# Load model (default: DehazeSNN-L trained on RESIDE-ITS)
model = DehazeSNNHub.from_pretrained("FengShaner/DehazeSNN")
model.cuda().eval()

# Load a different variant
# model = DehazeSNNHub.from_pretrained("FengShaner/DehazeSNN", revision="m-reside-6k")

# Prepare input image (RGB, normalized to [-1, 1])
img = Image.open("hazy_image.jpg").convert("RGB")
img_np = np.array(img).astype(np.float32) / 255.0
img_tensor = torch.from_numpy(img_np).permute(2, 0, 1).unsqueeze(0)  # [1, 3, H, W]
img_tensor = (img_tensor - 0.5) / 0.5  # normalize to [-1, 1]
img_tensor = img_tensor.cuda()

# Inference
with torch.no_grad():
    output = model(img_tensor).clamp(-1, 1)

# Convert output back to image
output = (output * 0.5 + 0.5).squeeze(0).permute(1, 2, 0).cpu().numpy()
output = (output * 255).astype(np.uint8)
Image.fromarray(output).save("dehazed_image.jpg")

Requirements

  • CUDA GPU required: The custom LIF CUDA kernels require an NVIDIA GPU with CUDA support.
  • CuPy: cupy-cuda12x (for CUDA 12.x) - CPU-only inference is not supported.
  • PyTorch >= 2.1 with CUDA 12.1.
  • Python 3.11 recommended.

Model Sizes

Variant Depths Parameters
M (Medium) [8, 12, 16, 12, 8] 2.70M
L (Large) [8, 16, 32, 16, 8] 4.75M

Citation

If you find this work useful, please cite our paper:

@INPROCEEDINGS{11228727,
  author={Li, Huibin and Liu, Haoran and Liu, Mingzhe and Xiao, Yulong and Li, Peng and Zan, Guibin},
  booktitle={2025 International Joint Conference on Neural Networks (IJCNN)}, 
  title={U-Net-Like Spiking Neural Networks for Single Image Dehazing}, 
  year={2025},
  pages={1-9},
  doi={10.1109/IJCNN64981.2025.11228727}
}

License

This project is released under the MIT License.

Downloads last month
86
Safetensors
Model size
4.75M params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Paper for FengShaner/DehazeSNN