Spectrally-Guided Diffusion Noise Schedules
Carlos Esteves, Ameesh Makadia
This paper presents a better way to control noise in diffusion models (AI systems that generate images by gradually removing noise). Instead of using the same hand-tuned noise schedule for all images, the researchers propose customizing noise schedules based on each image's spectral properties (its frequency content), which theoretically works better and requires fewer steps to generate high-quality images.
diffusion modelsimage generationnoise schedulinggenerative models