ETD

• generative classification
• normalizing flows

This thesis studies how to align classification and generation in architectures designed for image conditional generation and interventions. In disjoint classifier-generator architectures, conditioning a generator on desired class values for a set of concepts may produce an image whose predicted classes do not match the requested values. To formalize this issue, we introduce consistency metrics for conditional generation and counterfactual interventions.
We propose Partitioned Mixture Flows (PMF), a multi-concept extension of single-concept normalizing-flow classifiers based on mixture priors. PMF partitions the latent space into concept-specific subspaces, enabling multi-concept classification, conditional generation, and targeted interventions within a single unified model. Experiments show that PMFs achieve classification accuracy comparable to discriminative baselines while ensuring consistency between the requested class values and the classifier predictions for both conditional generation and targeted interventions.