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

Alexander Becker❄️🔥, Rodrigo Daudt❄️🔥, Dominik Narnhofer🔥, Torben Peters🔥, Nando Metzger🔥, Jan Dirk Wegner🌶️, Konrad Schindler🔥
 ❄️ Equal contribution
🔥 Photogrammetry and Remote Sensing, ETH Zurich
🌶️ Department of Mathematical Modeling and Machine Learning, University of Zurich

</div> <img src="assets/teaser_dark.svg#gh-dark-mode-only" alt="teaser" width=98%"/> <img src="assets/teaser_light.svg#gh-light-mode-only" alt="teaser" width=98%"/> <emph>Thera</emph> is the first arbitrary-scale super-resolution method with a built-in physical observation model.

News

2025-07-10: Training code is released 2025-03-15: We are #1 on Hacker News 🎉 2025-03-14: Interactive Hugging Face Space is online 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:

<table> <tr> <td>Backbone</td> <td>Variant</td> <td>Download</td> </tr> <tr> <td rowspan="3">EDSR-baseline</td> <td>Air</td> <td><a href="https://huggingface.co/prs-eth/thera-edsr-air">🤗 Hugging Face</a> | <a href="https://drive.google.com/file/d/18_XYU65ZYQuQOrfnlYAoee2KjjUhXuay/view?usp=sharing">Google Drive</a></td> </tr> <tr> <td>Plus</td> <td><a href="https://huggingface.co/prs-eth/thera-edsr-plus">🤗 Hugging Face</a> | <a href="https://drive.google.com/file/d/1ydYspibEQUskn67-CTc5IsTysEWsdO8Q/view?usp=sharing">Google Drive</a></td> </tr> <tr> <td>Pro</td> <td><a href="https://huggingface.co/prs-eth/thera-edsr-pro">🤗 Hugging Face</a> | <a href="https://drive.google.com/file/d/18slBa-dv-Z6SCTzL65MtmIryksfzdcnE/view?usp=sharing">Google Drive</a></td> </tr> <tr> <td rowspan="3">RDN</td> <td>Air</td> <td><a href="https://huggingface.co/prs-eth/thera-rdn-air">🤗 Hugging Face</a> | <a href="https://drive.google.com/file/d/1EzJaexc_OoxinaLZYXs6BRJMQ1XgZRwO/view?usp=sharing">Google Drive</a></td> </tr> <tr> <td>Plus</td> <td><a href="https://huggingface.co/prs-eth/thera-rdn-plus">🤗 Hugging Face</a> | <a href="https://drive.google.com/file/d/1mnn3XUSeWs-yBEpWcXSV7_jep_4nOdSo/view?usp=sharing">Google Drive</a></td> </tr> <tr> <td>Pro</td> <td><a href="https://huggingface.co/prs-eth/thera-rdn-pro">🤗 Hugging Face</a> | <a href="https://drive.google.com/file/d/1h6MPs6HSx5kVx3m703gZNbE-d5EMV5CD/view?usp=sharing">Google Drive</a></td> </tr> </table>

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 ...

You can run python run_eval.py -h to display all testing options.

Use with interactive Gradio app

You can also host a local version of our Hugging Face demo. To do so, clone the dedicated demo repo:

> git clone https://huggingface.co/spaces/prs-eth/thera thera-demo

Instructions for running the demo locally can be found in the repo's README.md.

Training

A simple command for training looks like this:

> python run_train.py --data-dir data_dir --size pro

Check args/eval.py for all training options.

Useful XLA flags

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

Citation

If you found our work helpful, consider citing our paper 😊:

@article{becker2025thera,
  title={Thera: Aliasing-Free Arbitrary-Scale Super-Resolution with Neural Heat Fields},
  author={Becker, Alexander and Daudt, Rodrigo Caye and Narnhofer, Dominik and Peters, Torben and Metzger, Nando and Wegner, Jan Dirk and Schindler, Konrad},
  journal={arXiv preprint arXiv:2311.17643},
  year={2025}
}

Thera

Install / Use

README