FouriScale explores high-resolution image generation from pre-trained diffusion models from a frequency analysis perspective. Through an innovative, no-training method, it replaces the original convolutional layers in pre-trained diffusion models with a combination of dilation techniques and low-pass operations, further enhanced by a fill-and-crop strategy. This allows for flexible handling of various aspect ratios in text-to-image generation. Guided by FouriScale, this method successfully balances the structural integrity and fidelity of generated images, achieving remarkable capabilities for arbitrary-sized, high-resolution, and high-quality generation. With its simplicity and compatibility, this method provides valuable insights for future explorations in ultra-high-resolution image synthesis.