Swap Inception for MobileNet and deblur 100× faster

What it does\nDeblurGAN-v2 is an end-to-end GAN that removes motion blur from a single image. Its generator wraps a Feature Pyramid Network around an interchangeable backbone—pick Inception-ResNet-v2 for peak quality or a lightweight MobileNet variant for speed. The authors also pitch the architecture as a general-purpose image restoration toolkit.\n\nThe interesting bit\nInstead of locking the deblurring logic to one feature extractor, the authors made the backbone plug-and-play. Five FPN feature maps are upsampled, concatenated, and refined with a direct skip connection so the model learns the residual blur; swap the backbone and you get a different point on the speed-accuracy curve without rewriting the pipeline.\n\nKey highlights\n- Claims 10–100× speedup over competitors when using MobileNet backbones, while keeping PSNR within about 1.5 points of the heavy Inception model on the GoPro benchmark.\n- Ships pretrained weights for GoPro: Inception-ResNet-v2 (29.55 PSNR / 0.934 SSIM), MobileNet (28.17 / 0.925), and MobileNet-DSC (28.03 / 0.922).\n- Uses a relativistic conditional GAN with a double-scale discriminator.\n- Skip connection from input to output constrains the generator to learn only the blur residue.\n- README notes the same pipeline extends to general image restoration (noise, compression, etc.).\n\nCaveats\n- The parent repository hyperlink is broken (href=\"\"), so finding the upstream RestoreGAN code takes extra digging.\n- The MobileNet-DSC pretrained model link in the table is empty.\n- Most architectural detail lives in commented-out HTML blocks, leaving the rendered README thinner than the raw source.\n\nVerdict\nWorth a look if you need a modular deblurring baseline with ready-made speed-quality dials. Skip it if you want a maintained, production-ready pipeline with active community support—the repo is a 2019 paper artifact with the usual research-code rough edges.