NAFNet: Image restoration without nonlinear activations
Efficient PyTorch architecture for image restoration tasks.
Learn more about NAFNet
NAFNet is a convolutional neural network architecture designed for low-level vision tasks including image deblurring, denoising, and stereo super-resolution. The model is implemented in PyTorch and operates by replacing traditional nonlinear activation functions with multiplication operations or removing them entirely from the network. The architecture achieves competitive results on standard benchmarks such as GoPro for deblurring and SIDD for denoising while requiring substantially fewer computational operations than comparable methods. The framework is applicable to various image restoration scenarios where computational efficiency and inference speed are relevant considerations.
Activation-free design
The architecture removes or replaces nonlinear activation functions (ReLU, GELU, Sigmoid, Softmax) with simpler operations like multiplication, simplifying the model structure while maintaining restoration quality across multiple tasks.
Computational efficiency
Achieves comparable or improved performance on standard benchmarks while requiring significantly fewer multiply-accumulate operations than prior methods, as demonstrated on GoPro and SIDD datasets.
Multi-task capability
Provides implementations for multiple image restoration tasks including image deblurring, image denoising, and stereo image super-resolution through a unified architectural approach with task-specific variants.
pip install basicsrRelated Repositories
Discover similar tools and frameworks used by developers
evo2
Foundation model for DNA sequence generation and scoring.
dinov2
PyTorch vision transformers pretrained on 142M unlabeled images.
AI-Trader
LLM agent benchmarking framework for autonomous market trading.
streamlit
Python framework for reactive data web applications.
higgsfield
Cluster manager for multi-node PyTorch model training.