DCTdiff: Intriguing Properties of Image Generative Modeling in the DCT Space

Published in arXiv preprint, 2024

Abstract

This paper investigates the intriguing properties of image generative modeling in the Discrete Cosine Transform (DCT) space. Through deep understanding of DCT transform, we design more refined generation algorithms that better recover image details and features across different frequency bands.

Key Contributions

  • DCT Space Analysis: Comprehensive analysis of generative modeling in frequency domain
  • Novel Architecture: Frequency-aware generation algorithms
  • Enhanced Detail Recovery: Improved image detail preservation across frequency bands
  • Theoretical Insights: New understanding of frequency domain properties in image generation

Technical Innovation

Our approach leverages the inherent properties of DCT transformation to design more effective image generation models, opening new avenues for frequency-domain image synthesis.

Download paper here

Recommended citation: Jia, H. et al. (2024). "DCTdiff: Intriguing Properties of Image Generative Modeling in the DCT Space." arXiv preprint arXiv:2412.15032. https://arxiv.org/abs/2412.15032