Thera
First super-resolution with built-in physical observation — alias-free at any scale
The first super-resolution method with a built-in physical observation model — alias-free at any magnification.
AI video generation, digital humans, and video restoration.
Key features
- Arbitrary-scale super-resolution with neural heat fields
- Built-in physical observation model; aliasing-free upscaling
- Pre-trained Air, Plus, Pro checkpoints on EDSR and RDN backbones
- Single command for any scale factor; training code released
Good for
- Upscaling images to arbitrary scales
- Researchers in neural image restoration
Stack & tags
PythonJAXsuper-resolutionimage-upscalingarbitrary-scalealiasing-freeopen-weightsneural-fieldsjax
Quick start
./super_resolve.py IN_FILE OUT_FILE --scale 3.14 --checkpoint thera-rdn-pro.pklFrom the project README on GitHub.
Similar open-source projects
Hyperframes
★ 55,468Write HTML. Render video. Built for agents — HeyGen's open-source video framework
Hypit
★ 18,649Clone any viral video with AI agents — swap the face, the words, the B-roll, ship 100 variants