HiTokSR: A Coarse-to-Fine Tokenizer with Hierarchical Codebooks for High-Fidelity Real-World Image Super-Resolution
Researchers have introduced HiTokSR, a novel framework for image super-resolution that utilizes a hierarchical approach to codebooks. This method separates global structures from fine details, improving representational capacity and stability compared to existing monolithic latent space models. The framework also incorporates priors from vision foundation models and an index-level perturbation strategy to enhance semantic consistency and bridge the train-test discrepancy, achieving state-of-the-art results on real-world benchmarks. AI
IMPACT Introduces a novel approach to image super-resolution that could lead to more detailed and accurate image enhancements.