This paper presents an extension of Diffusion Probabilistic Models using Fields to model data. The basic idea is to model data as functions (fields) e.g. images are functions from to , a map between pixel coordinates and color pixel value.
This approach is motivated by the need to unify the generative model architecture to cover different modality, thus avoiding the need to construct specific architecture for score functions which are the denoisers used in the generation process. Fields are parameterized using pairs of coordinates in source and target domains and evaluated in a similar manner using pairs of query pairs of coordinates.
