Semi-supervised semantic segmentation leverages both labeled and unlabeled images to accomplish pixel-wise classification task. Within this field, the weak-to-strong consistency regularization has been widely popularized and has become a standard approach. However, unidirectional regularization often leads to the ignorance of correct but filtered predictions and brings the noise of wrong but confident predictions. To address these inherent flaws, we fully leverage Cross-Augmentation Consensus and Conflict (CACC), including Augmentation Feedback Mechanism (AFM) and Category Threshold Controller (CTC). AFM aims to mitigate the influence of incorrect predictions with high-confidence and mine unconfident but accurate predictions by re-weighting the pixel-wise pseudo supervision and applying supplementary regularization. Concurrently, CTC adopts category-specific thresholds by considering the model’s overall performance and the varying category-specific learning difficulty. Experimental results on benchmark datasets demonstrate the superior performance of our method, showcasing its effectiveness in improving semi-supervised semantic segmentation.