Multi-Scale Generative Modeling in Wavelet Domain
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: wavelet transform; score-based generative model; diffusion model; wavelet decomposition
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: While working within the spatial domain can pose problems associated with ill-conditioned scores, recent advancements in diffusion-based generative models have shown that transitioning to the wavelet domain offers a promising alternative. However, within the wavelet domain, we encounter unique complexities, especially the sparse representation of high-frequency coefficients, which deviates significantly from the Gaussian assumptions in the diffusion process. To this end, we propose developing a multi-scale generative model directly within the wavelet domain using Generative Adversarial Networks. This Multi-Scale Generative Model in the Wavelet Domain (i.e., Wavelet Multi-Scale Generative Model (WMGM)) leverages the benefits of wavelet coefficients, with a specific emphasis on using low-frequency coefficients as conditioning variables. Based on theoretical analysis and experimental results, our model provides a pioneering framework for implementing generative models in the wavelet domain, showcasing remarkable performance improvements and significant reduction in trainable parameters, sampling steps and time. This innovative approach represents a promising step forward in the field of diffusion modeling techniques.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
Supplementary Material: pdf
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8926
Loading