Resurrecting EDSR in PyTorch, minus the Torch7 bloat

What it does This is the official PyTorch port of the CVPRW 2017 paper Enhanced Deep Residual Networks for Single Image Super-Resolution. It provides training and evaluation scripts to reproduce the paper’s results, plus pre-trained EDSR and MDSR checkpoints for 2×, 3×, and 4× upscaling. Over time the repo has also absorbed implementations of related architectures—D-DBPN, RDN, RCAN—and optional perceptual losses borrowed from SRGAN.

The interesting bit The authors bill this as a deliberate slim-down: models are about half the size of the original Torch7 versions, the codebase is more compact, and reported PSNR is marginally higher. For hardware-constrained inference, a --chop_forward mode splits oversized images into patches, super-resolves them, and stitches the result back together—tested up to a 16,000×8,000 output on a 12 GB card.

Key highlights

• Reproduces paper numbers on DIV2K and standard benchmarks (Set5, Set14, B100, Urban100)
• Baseline and full EDSR/MDSR models with published parameter counts and PSNR
• Half-precision test-time inference (--precision half) without image degradation
• Video input/output support for inference-only upscaling
• Includes SRGAN-style adversarial and VGG perceptual losses, plus untested WGAN/gradient-penalty code

Caveats

• MDSR training and evaluation is temporarily disabled on the master branch for PyTorch 1.2.0 due to torch.utils.data.dataloader issues; legacy branches are required to use it
• Some newer loss additions (WGAN, gradient penalty) are implemented but explicitly not tested
• EDSR x3 and x4 training expect a pre-trained x2 model unless you manually remove the --pre_train constraint

Verdict A solid reference implementation for anyone benchmarking classic super-resolution baselines or training variants from scratch. Look elsewhere if you want a polished, maintenance-free product; the repo carries years of PyTorch version migration patches.