Learning Adaptive and Visually Aligned Conditions for Generative Zero-Shot Learning
The ADiVA framework addresses challenges in generative zero-shot learning by jointly modeling attribute distributions and aligning semantic and visual spaces. It introduces an Attribute Distribution Encoder to sample instance-level attributes, tackling the class–instance gap. Simultaneously, a Visual-Guided Alignment module maps these attributes to visual priors, bridging the semantic–visual domain gap for better unseen class feature generation.