Reproduce “Local Implicit Wavelet Transformer for Arbitrary-Scale Super-Resolution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Minghong Duan, Linhao Qu, Shaolei Liu, Manning Wang
The Local Implicit Wavelet Transformer (LIWT) enhances arbitrary-scale super-resolution by integrating Discrete Wavelet Transform-derived high-frequency information with an implicit neural representation and a wavelet-aware attention mechanism. This approach consistently achieves superior PSNR, SSIM, and LPIPS scores across various integer and non-integer scaling factors, improving the restoration of intricate texture details in generated images.
Key insights
- Explicitly incorporating high-frequency prior information, derived from DWT, is crucial for significantly improving the reconstruction of intricate texture details in arbitrary-scale super-resolution.
- Utilizing feature-aware attention, rather than solely positional weighting, enhances the model's ability to effectively leverage local information and focus on critical high-frequency components during implicit function decoding.
- A curriculum learning strategy that gradually expands the range of scaling factors during training improves model robustness and overall performance across a wide spectrum of arbitrary upscaling factors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.