English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

The primary aim of single-image super-resolution is to constructhigh-resolution (HR) images from corresponding low-resolution (LR) inputs. Inprevious approaches, which have generally been supervised, the trainingobjective typically measures a pixel-wise average distance between thesuper-resolved (SR) and HR images. Optimizing such metrics often leads toblurring, especially in high variance (detailed) regions. We propose analternative formulation of the super-resolution problem based on creatingrealistic SR images that downscale correctly. We present an algorithmaddressing this problem, PULSE (Photo Upsampling via Latent Space Exploration),which generates high-resolution, realistic images at resolutions previouslyunseen in the literature. It accomplishes this in an entirely self-supervisedfashion and is not confined to a specific degradation operator used duringtraining, unlike previous methods (which require supervised training ondatabases of LR-HR image pairs). Instead of starting with the LR image andslowly adding detail, PULSE traverses the high-resolution natural imagemanifold, searching for images that downscale to the original LR image. This isformalized through the "downscaling loss," which guides exploration through thelatent space of a generative model. By leveraging properties ofhigh-dimensional Gaussians, we restrict the search space to guarantee realisticoutputs. PULSE thereby generates super-resolved images that both are realisticand downscale correctly. We show proof of concept of our approach in the domainof face super-resolution (i.e., face hallucination). We also present adiscussion of the limitations and biases of the method as currently implementedwith an accompanying model card with relevant metrics. Our method outperformsstate-of-the-art methods in perceptual quality at higher resolutions and scalefactors than previously possible.

PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models