复制研究：面向视觉语言模型的联邦文本驱动提示生成

Vision-language models such as CLIP have shown remarkable zero-shot capabilities; however, their adaptation to federated learning scenarios poses significant challenges, particularly in generalizing to unseen classes. The original FedTPG paper addresses this limitation by introducing a text-driven prompt generation network that dynamically creates prompts conditioned on class names, enhancing cross-class generalization in federated settings. A faithful replication study of FedTPG was conducted, evaluating the pre-trained model across six diverse vision datasets: Caltech101, Oxford Flowers, FGVC Aircraft, Oxford Pets, Food-101, and DTD. The evaluation achieved results within 0.2% of the original paper's findings.