explore the refinement of text-image cross-model alignment and enhance the model's adaptability to text diversity and noise through token-level feature alignment