For diagnosis of diabetic retinopathy based on domain adaptation methods in deep learning, the diffusion-enhanced domain-attention transfer learning model proposed in this paper consists of two main modules. Firstly, the denoising diffusion probabilistic diabetic retinopathy generation module generates abundant and diverse target domain samples, enabling the model to learn more comprehensive target domain features. Secondly, our model designs a multi-source-free attention ensemble module, which achieves weighted attention integration of multiple source domain pre-trained models, without the need to access source domain data. Therefore, this model obtains a good balance between instance-specific features and domain-consistent features. Experimental results demonstrate that the model achieves an accuracy of 90.66%, a precision of 87.47%, a sensitivity of 85.41%, a specificity of 91.63%, and an F1 score of 86.42% in the referable diabetic retinopathy diagnosis task. Meanwhile, in the normal/abnormal retinopathy recognition task, the model reaches an accuracy of 96.75%, a precision of 99.23%, a sensitivity of 90.47%, a specificity of 99.27%, and an F1 score of 94.65%. The model proposed in this paper can conduct effective retinopathy diagnosis without accessing source domain data and without target domain labels.