<ul data-eligibleForWebStory="true"><li>Image tokenizers are essential for autoregressive transformer-based image generation, mapping images to sequences of discrete tokens.</li><li>The proposed spectral image tokenizer in this paper tokenizes the image spectrum obtained from a discrete wavelet transform.</li><li>Advantages of the spectral image tokenizer include leveraging the compressibility of natural images at high frequencies and enabling image reconstruction at different resolutions without retraining.</li><li>The tokenizer improves conditioning for next-token prediction compared to traditional approaches and enables partial decoding for coarse image reconstruction.</li><li>It also allows autoregressive models to be utilized for image upsampling, providing versatility in image manipulation tasks.</li><li>Evaluation of the tokenizer includes reconstruction metrics, multiscale image generation, text-guided image upsampling, and editing.</li></ul>

Spectral Image Tokenizer

Discover more