<p>No-reference image quality assessment aims to predict the visual quality of images without reference images. While deep learning has greatly advanced NR-IQA, many methods show limited generalization from synthetic to realistic distortions. To address this challenge, we propose TFFNRIQA, a two-branch feature fusion framework that explicitly decouples and jointly models non-local semantic dependencies and local distortion-sensitive features. Specifically, multi-level features extracted by a pretrained ResNet50 backbone are enhanced through two complementary branches: a non-local feature branch based on Transformer self-attention for capturing global dependencies, and a local feature branch based on spatial attention for emphasizing local distortion regions. An adaptive attention-based fusion module dynamically integrates features from both branches to generate more comprehensive quality representations. Experiments on both synthetic and realistic datasets demonstrate the superiority of the proposed method. On realistic distortion datasets, TFFNRIQA achieves strong correlation with human perception, with SROCC/PLCC values of 0.896/0.901 on LIVEC and 0.927/0.939 on KonIQ-10k, outperforming state-of-the-art approaches. The experiments also demonstrate the superiority of TFFNRIQA in terms of computational complexity. The code is available at <a href="https://github.com/With-Sky/TFFNRIQA.">https://github.com/With-Sky/TFFNRIQA.</a></p>