Reproduce “Spectral Image Tokenizer” 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
Carlos Esteves, Mohammed Suhail, Ameesh Makadia
Researchers at Google Research developed a Spectral Image Tokenizer (SIT) that leverages Discrete Wavelet Transforms to represent images as coarse-to-fine spectral token sequences. This approach improves image reconstruction quality and enables autoregressive models to achieve competitive generation performance while supporting novel multiscale applications like text-guided upsampling and image editing.
Key insights
- Leveraging the inherent spectral properties of natural images, such as higher compressibility of high frequencies, allows for more efficient and perceptually aligned image tokenization.
- A coarse-to-fine spectral token ordering provides superior conditioning for autoregressive models, improving generation quality and efficiency compared to traditional raster-scan approaches.
- Integrating DWT into a learnable tokenizer enables truly multiscale image representations, facilitating resolution-agnostic processing and novel applications within a unified AR framework.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.