Synth^2: Boosting Visual-Language Models with Synthetic Captions and Image Embeddings