Thera: Aliasing-Free Arbitrary-Scale Super-Resolution with Neural Heat Fields

Official implementation of the paper by Alexander Becker*, Rodrigo Daudt*, Dominik Narnhofer, Torben Peters, Nando Metzger, Jan Dirk Wegner and Konrad Schindler (* equal contribution)

Thera is the first arbitrary-scale super-resolution method with a built-in physical observation model.

News

2025-03-12: Pre-trained checkpoints are released

Setup environment

You need a Python 3.10 environment (e.g., installed via conda) on Linux as well as an NVIDIA GPU. Then install packages via pip:

> pip install --upgrade pip
> pip install -r requirements.txt

Use with pre-trained models

Download checkpoints:

Backbone	Variant	Download
EDSR-base	Air	Google Drive
	Plus	Google Drive
	Pro	Google Drive
RDN	Air	Google Drive
	Plus	Google Drive
	Pro	Google Drive

Super-resolve any image with:

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

You can evaluate the models on datasets using the run_eval.py script, e.g.:

> python run_eval.py --checkpoint thera-rdn-pro.pkl --data-dir path_to_data_parent_folder --eval-sets data_folder_1 data_folder_2 ...

Check the arguments in args.py (bottom of file) for all testing options.

Training

Training code will be released soon.

Useful XLA flags

• Disable pre-allocation of entire VRAM: XLA_PYTHON_CLIENT_PREALLOCATE=false
• Disable jitting for debugging: JAX_DISABLE_JIT=1

Citation

Citation coming soon.