Super-Resolution Via Mixture-of-Experts

The significant growth of screen content sharing, driven by remote collaboration and online education, has increased the demand for high-quality screen content image (SCI) upscaling, particularly under bandwidth-constrained scenarios where high-resolution content is downsampled or degraded. Existing super-resolution methods (SR) designed for natural images fail to generalize well to SCIs, which require restoring fine graphical structures and sharp textual details at arbitrary upscaling and degradation levels. This work proposes a novel mixture-of-experts framework addressing screen-sharing content. First, we apply the Implicit Transformer Super-Resolution Network (ITSRN) to reconstruct the general image structure, leveraging implicit transformers and continuous position encoding for high-quality continuous SR. To further enhance the readability of independent text paragraphs within the image, we introduce a second expert: we localize text regions, apply a diffusion-based text region synthesis model based on TextSSR, and refine the generated high-resolution text with the help of a Large Language Model (LLM), ensuring semantic correctness and visual fidelity.