Hierarchically branched diffusion models leverage dataset structure for class-conditional generation
This addresses the need for more structured class-conditional generation in scientific domains, representing a novel method for a known bottleneck.
The authors tackled the problem that class-conditional diffusion models ignore hierarchical relationships in labeled datasets by proposing hierarchically branched diffusion models, which learn separate reverse diffusion processes for each branch of a hierarchy and demonstrated advantages like extension to novel classes and improved interpretability on benchmark and scientific datasets.
Class-labeled datasets, particularly those common in scientific domains, are rife with internal structure, yet current class-conditional diffusion models ignore these relationships and implicitly diffuse on all classes in a flat fashion. To leverage this structure, we propose hierarchically branched diffusion models as a novel framework for class-conditional generation. Branched diffusion models rely on the same diffusion process as traditional models, but learn reverse diffusion separately for each branch of a hierarchy. We highlight several advantages of branched diffusion models over the current state-of-the-art methods for class-conditional diffusion, including extension to novel classes in a continual-learning setting, a more sophisticated form of analogy-based conditional generation (i.e. transmutation), and a novel interpretability into the generation process. We extensively evaluate branched diffusion models on several benchmark and large real-world scientific datasets spanning many data modalities.