The newly proposed framework, called Ambient Diffusion, gets around this problem by training diffusion models through access only to corrupted image-based data. Early efforts suggest the framework is ...