Fast-SRGAN

Why we included this project

Teams dealing with old or low-resolution video often need a fast way to bring footage up to modern display standards, and this repo offers a concrete option. It implements an SRGAN-style generator that uses pixel shuffle to keep upsampling quick, with published benchmarks showing roughly 720p output at about 30fps on an M1 Pro. A pretrained generator trained on DIV2K is included, so you can run inference on your own images right away with a single command instead of training from scratch. Retraining is also straightforward: Hydra-based config lets you change batch size, residual block count, and dataset paths from the command line, and TensorBoard logs training progress. It's a focused, single-purpose codebase that gives you a realistic baseline for evaluating super-resolution in video pipelines or image enhancement work.