Reproduce “Implicit neural representation of textures” 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
Albert Kwok, Zheyuan Hu, Dounia Hammou
Researchers from the University of Cambridge developed and evaluated Implicit Neural Representations (INRs) for textures, demonstrating that Fourier-encoded MLPs achieve high perceptual quality and efficiency over traditional compression methods like ASTC. The work integrates INRs into a real-time renderer and explores their utility for mipmap fitting and INR-space texture generation.
Key insights
- Conventional MLPs exhibit spectral bias, leading to blurry texture reconstructions, while Fourier-encoded MLPs and SIRENs effectively learn high-frequency details for superior perceptual quality.
- Fourier-encoded MLPs consistently achieved higher perceptual quality compared to SIRENs and traditional ASTC compression at similar bitrates.
- The Adam optimizer provided more stable and superior training results for INR architectures compared to Rprop.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.