<p>In-the-wild facial expression datasets often encounter the coupled challenges of noisy labels and long-tailed distributions, which significantly diminish the performance of facial expression recognition (FER) models in real-world applications. However, previous noisy FER methods tend to overlook the significance of fine-grained structural information, such as the textures and edges of facial muscle movements, in distinguishing the inherent inter-class similarity features of facial expressions. To tackle these challenges, we propose a novel consistent prompt learning (CPL) approach, which primarily consists of an augmented fine-grained prompts (AFP) module, a random semantic consistency (RSC) module, and an adaptive binary cross-entropy (ABC) loss, each contributing uniquely to the robustness of CPL. The AFP module innovatively integrates random prompts with augmented prior knowledge derived from facial image textures and edges, extracting discriminative fine-grained features that indirectly enhance inter-class separation under noisy labels. Despite the strengths of the AFP module, complex scenarios such as pose variations and occlusions may obscure critical features. In response, the RSC module ensures a consistent learning environment by maintaining discriminative feature extraction through the imposition of consistency between the original and randomly cropped images, thereby preventing the model from overfitting to noisy labels in complex facial scenarios. Furthermore, to mitigate the negative impact of noisy labels in tail classes, we employ the ABC loss, which introduces an adaptive factor to refine the resampling process, balancing fine-grained feature extraction across different classes. Qualitative and quantitative experiments conducted on FER benchmark datasets validate the effectiveness and utility of our CPL approach.</p>