thera-rdn-plus

Overview

This is a model from the paper Thera: Aliasing-Free Arbitrary-Scale Super-Resolution with Neural Heat Fields. It enables SOTA arbitrary-scale super-resolution, leveraging a built-in analytically correct observation model for anti-aliasing when moving across scales.

• Project Page: https://therasr.github.io
• Code Repository: https://github.com/prs-eth/thera
• Demo: https://huggingface.co/spaces/prs-eth/thera

Model Details

• Description: This model can be used to enable super-resolution of single images at arbitrary, non-integer scaling factors.
• Backbone: RDN
• Variant: Plus
• Training Dataset: DIV2K

Usage

To use this model, first clone the official repository and set up the environment. You will need a Python 3.10 environment and an NVIDIA GPU.

git clone https://github.com/prs-eth/thera.git
cd thera
pip install --upgrade pip
pip install -r requirements.txt

After setting up the environment and downloading the thera-rdn-plus.pkl checkpoint (available in the "Files and versions" tab of this repository), you can super-resolve any image with the following command:

./super_resolve.py IN_FILE OUT_FILE --scale 3.14 --checkpoint thera-rdn-plus.pkl

License

Apache-2.0