This paper presents an interpretable image restoration framework based on primal-dual variational learning. Different from conventional total variation models with fixed regularization structures, the proposed method introduces learnable anisotropic priors constructed through trainable filtering operators and nonlinear influence functions. A primal-dual optimization strategy is employed to transform the variational model into an efficient trainable architecture, enabling adaptive parameter learning while preserving the interpretability of model-based restoration. To facilitate stable optimization, a differentiable projection mechanism is further incorporated into the learning process. Experimental results on image denoising and image deblurring benchmarks demonstrate that the proposed framework achieves superior restoration performance compared with several representative variational and learning-based methods. The learned regularization priors further provide meaningful interpretations of image structures, illustrating the effectiveness of combining primal-dual optimization with adaptive prior learning.