Generating images faster by filling in details last

Diffusion models generate images by denoising over many steps, with low frequencies appearing early and fine details emerging late. Spectral Progressive Diffusion exploits this structure by progressively increasing image resolution during generation, avoiding redundant computation on noise-heavy, high-resolution stages. The method uses a spectral noise expansion mechanism and derives an optimal resolution schedule from the model's power spectrum. It works training-free on existing pretrained models and offers an optional fine-tuning approach to further improve speed and quality. Experiments on state-of-the-art image and video models show substantial acceleration while maintaining visual fidelity.