A Keras magnifying glass that’s about to retire

What it does ISR is a Keras-based library for single-image super-resolution built by idealo. It packages implementations of Residual Dense Networks and ESRGAN-style architectures, letting you either download pre-trained weights or train your own generator-discriminator pipeline with perceptual VGG loss. The goal is straightforward: reconstruct a convincing high-resolution image from a low-resolution input.

The interesting bit The project bundles both PSNR-driven models and a GAN-flavoured variant that chases perceptual sharpness, plus a niche artefact-cancelling weight set for compressed images. That tension—mathematical fidelity versus believable fake detail—is usually left to the reader by research repos; here the weights are baked in and selectable by a string argument.

Key highlights

• Ships with four pre-trained weight sets: psnr-large, psnr-small, noise-cancel, and gans.
• Includes a full training stack: generator, discriminator, cut VGG19 feature extractor, and a Trainer class with configurable loss weights.
• Supports patch-based inference via by_patch_of_size to keep memory in check on large images.
• Provides Docker, Colab, and AWS/nvidia-docker scripts for remote training.
• Licensed under Apache 2.0.

Caveats

• The repository is no longer maintained and is scheduled for archival on 2025-01-03.
• Targets Python 3.6 and relies on older Keras/TensorFlow patterns; the README notes an h5py version incompatibility when loading weights.
• Training requires careful tuning of loss weights—starting with GAN loss too early yields patchy results.

Verdict Worth a look if you need a ready-made Keras implementation of RDN or ESRGAN for research or batch upscaling. Skip it if you need active maintenance or modern PyTorch-based tooling.