The 2016 super-resolution paper, rebuilt for a multi-backend world (almost)

What it does This is a faithful implementation of the 2016 SRGAN paper, which uses a generative adversarial network to upscale a single low-resolution image by 4× while trying to keep the output photo-realistic. It is built specifically on TensorLayerX and includes training and evaluation pipelines for the DIV2K dataset or your own high-resolution images.

The interesting bit The hook is a one-line backend switch—set TL_BACKEND to TensorFlow or PaddlePaddle—though PyTorch and MindSpore support are still marked “coming soon.” That makes it less a generic repo and more a flagship demo for the TensorLayerX ecosystem.

Key highlights

• Reference implementation of Ledig et al.’s SRGAN for 4× single-image super-resolution.
• Pretrained generator and discriminator weights are provided for TensorFlow and PaddlePaddle.
• Supports DIV2K, MirFlickr25k, or custom image folders via config.py.
• One-line backend switch via TL_BACKEND environment variable.

Caveats

• PyTorch and MindSpore backends and pretrained weights are explicitly listed as “Coming soon” in the README.
• Requires manually downloaded VGG19 weights and a prepared high-resolution training set.
• Licensed for academic and non-commercial use only; commercial use requires contacting the authors.

Verdict Worth a look if you need a working SRGAN baseline in TensorFlow or PaddlePaddle and do not mind the TensorLayerX dependency. Avoid if you require PyTorch support today or a permissive open-source license.