Latent Space Manifold Mapping for Super-Resolution

This research argues that current super-resolution methods fail by operating in reconstruction-oriented or pixel spaces rather than semantic manifolds. It proposes mapping degraded embeddings back to the high-resolution manifold to better preserve both semantics and fine pixel details.