Manifold Diffusion Fields – Apple Machine Learning Research

This paper was accepted on the Diffusion Fashions workshop at NeurIPS 2023.

Rating-based fashions have rapidly change into the de facto selection for generative modeling of photographs, textual content and extra not too long ago molecules. Nevertheless, to adapt a score-based generative modeling to those domains the rating community must be fastidiously designed, hampering its applicability to arbitrary knowledge domains. On this paper we sort out this downside by taking a textit{useful} view of knowledge. This useful view permits to forged seemingly totally different domains to a standard shared illustration. We then re-formulate the rating operate to take care of useful knowledge and present: i) this unified structure might be successfully utilized to totally different modalities: photographs, geometry, video, and ii) we will study generative fashions of alerts outlined on non-euclidean geometry.