High-resolution pathological images are crucial for improving the accuracy of disease diagnosis. However, acquiring such images in real-time is often challenging due to limitations in medical imaging equipment resolution and scanning duration. Consequently, software-based super-resolution (SR) techniques for pathological images hold significant practical value. Classical image SR algorithms often face difficulties in model parameter estimation, while most deep learning-based methods do not adequately model the conditional relationships between low-frequency and high-frequency image information, leading to blurred and unrealistic reconstructed details unsuitable for pathological analysis. To address these issues, this paper proposes a Sparse-coding Non-local Attention Flow Model (SNAFM). This model achieves pathological image super-resolution through sparse-coding non-local attention modules, unconditionally invertible transformation modules, conditionally invertible transformation modules, Gaussian constraints, parameter sharing strategies, and Bayesian optimization within a dual-branch architecture. The reconstructed pathological images achieved a Peak Signal-to-Noise Ratio (PSNR) of 30.92 dB and a Structural Similarity (SSIM) index of 0.917. Experimental results demonstrate that the proposed method not only accurately reconstructs high-frequency details but also enhances modeling efficiency via a lightweight sparse-coding non-local attention mechanism.