ZhendongWang6/Uformer

Combines hierarchical encoder-decoder architecture with local-enhanced window Transformer blocks using non-overlapping self-attention and depth-wise convolution to reduce computational cost while maintaining long-range dependency modeling. Implements flexible skip-connection schemes for effective information flow across encoder-decoder layers, enabling a unified framework for diverse low-level vision tasks including denoising, deblurring, deraining, and demoireing. Built on PyTorch with pre-trained weights available for multiple datasets (SIDD, GoPro, DND, HIDE, RealBlur variants).

932 stars. No commits in the last 6 months.