yinboc/liif

Encodes images as continuous functions via local implicit representations, enabling arbitrary resolution upsampling without retraining. The architecture combines a feature extraction backbone (EDSR or RDN) with coordinate-based MLPs that predict pixel values at any spatial location, trained end-to-end on DIV2K and CelebA-HQ. Supports flexible super-resolution across multiple upsampling factors (2×, 3×, 4×) and arbitrary output resolutions through a single trained model.

1,368 stars. No commits in the last 6 months.