We can make hybrid images from existing images by extracting the low or high pass component from an image and generating the missing component. This special case of our method effectively allows us to solve inverse problems with diffusion models, and has been studied in much more detail in many, many, many, many, many, many prior works. In contrast to these papers, we show it is possible to sample illusions, which are quite out of distribution for these pre-trained diffusion models.
