PAPT++: Risk-Aware Adversarial Tuning and Generation for Single Domain Generalization

Single-domain generalization (SDG) aims to use a single-labeled source domain learning model to generalize to unseen target domains. For this task, the PAPT++ method was proposed, which is a risk-aware adversarial generation and training framework. This method first learns diverse semantic reference images for each class through image-text alignment and intra-class diversity regularization, using these images as denoising targets to guide diffusion synthesis, thereby reducing semantic drift and focusing on challenging variations. Subsequently, the generated samples are combined with the source data to update the classifier, and the updated classifier is used to guide the next round of synthesis, gradually exposing the model to challenging but semantically consistent sample variations. Extensive experiments on standard single-domain generalization benchmarks confirmed that the PAPT++ method and its main components demonstrate superiority and effectiveness.