✨ thera-mlx

Aliasing-Free Arbitrary-Scale Super-Resolution for Apple Silicon

An MLX port of Thera — running natively on M1/M2/M3/M4

128×128 → 512×512 in ~1.5 seconds on M1 · Arbitrary scale factors · Web UI + CLI · Video support

---

Model Description

thera-mlx contains MLX-format weights for the Thera super-resolution model, converted from the original JAX/Flax checkpoints released by PRS-ETH (ETH Zurich).

Thera uses neural heat fields to perform aliasing-free image super-resolution at any scale factor — not just 2× or 4×. This repo provides two variants:

File	Model	Speed	Quality
weights-air.safetensors	Air	~1.5s for 128→512px on M1	Great
weights-pro.safetensors	Pro (+ SwinIR refinement)	~5s for 128→512px on M1	Best

Usage

git clone https://github.com/madeleinelmuller/thera-mlx
cd thera-mlx
pip install -r requirements.txt

# Download weights (pulls from this repo automatically)
python run.py convert --model air
python run.py convert --model pro

# Launch web UI
python run.py

# Or use CLI
python run.py run input.png output.png --scale 4 --model air

See the full GitHub repository for complete documentation.

Original Weights

The weights in this repository are converted from:

• prs-eth/thera-rdn-air
• prs-eth/thera-rdn-pro

Citation

If you use this in your work, please cite the original Thera paper:

@article{becker2023thera,
  title   = {Thera: Aliasing-Free Arbitrary-Scale Super-Resolution with Neural Heat Fields},
  author  = {Becker, Alexander and Caye Daudt, Rodrigo and Narnhofer, Dominik and
             Peters, Torben and Metzger, Nando and Wegner, Jan Dirk and Schindler, Konrad},
  journal = {arXiv preprint arXiv:2311.17643},
  year    = {2023}
}

Original resources:

• Paper: arxiv.org/abs/2311.17643
• Project page: therasr.github.io
• Original HF weights: huggingface.co/prs-eth