Zero-Shot Learning (ZSL) aims to recognize unseen categories by transferring semantic knowledge from seen to unseen classes. However, achieving precise visual–semantic alignment and robust generalization to unseen classes remains challenging in ZSL. This stems from the fact that semantic attributes exhibit varying preferences for visual granularities, while prediction reliability across semantic representations varies notably across samples. Conventional supervised visual encoders are often biased toward seen classes, which aggravates these challenges. To tackle these challenges, we propose ZeroDINO, leveraging Self-Distillation with No Labels (DINO) as the backbone to provide semantically meaningful and label-agnostic visual representations. The framework comprises three components: 1) Granularity-Aware Visual Encoder (GAVE) performs cross-granularity interaction and distillation to provide discriminative features; 2) Dynamic Semantic Decoder (DSD) models attribute-wise granularity preferences and adaptively fuses fine- and coarse-grained semantic representations; 3) Entropy-Guided Prediction Fusion (EGPF) adaptively integrates predictions from multiple semantic representations based on entropy-derived reliability, improving robustness and generalization to unseen classes. On three standard ZSL benchmarks, ZeroDINO achieves consistent state-of-the-art performance, showing that granularity-aware modeling and entropy-guided fusion effectively enhance visual-semantic alignment and generalization.