A living museum of super-resolution architectures

What it does

The VSR package is a training and inference framework that reimplements a sprawling catalog of single-image and video super-resolution architectures—SRCNN, EDSR, RCAN, VESPCN, FRVSR, and many others—under one roof. It is built primarily on TensorFlow and deliberately avoids heavy abstraction, aiming for a straightforward codebase. The framework also ingests raw NV12/YUV video directly, which saves you from preprocessing pipelines when working with real footage.

The interesting bit

Rather than scattering reference implementations across incompatible repos, the author has normalized them into a single API with swappable backends (TensorFlow and, for some models, PyTorch). That makes it unusually easy to run controlled comparisons between a 2014 classic and a 2019 champion using the same data loader and training loop.

Key highlights

• Covers both image SR (e.g., EDSR, DBPN, RDN) and video SR (e.g., VESPCN, FRVSR, RBPN) in one framework.
• Many models ship with pretrained weights available via the prepare_data helper or linked releases.
• Handles raw NV12/YUV sequences natively, not just stacked PNG frames.
• Several newer models include dual TensorFlow and PyTorch implementations.
• Published as a PyPI package named VSR.

Caveats

• The documented install path targets TensorFlow 1.15, which is years out of date; you will likely need to wrestle with modern CUDA compatibility.
• A handful of listed models are marked “under-testing” rather than fully integrated.
• The README notes that pretrained weights are still being uploaded, so availability may vary by model.

Verdict

Researchers who need to compare historical and modern SR architectures on standard benchmarks should find this a convenient launchpad. Production engineers looking for a maintained, dependency-fresh toolkit should probably look elsewhere.