Figure 2. Overview of our framework. (a) The supervision enhancement framework produces enhanced ground truth images by fusing frequency components from the original ground truth I_0^GT and its super-resolved variants I_i^GT using adaptive frequency masks M_i. (b) The conditional frequency mask generator constructs M_i by combining predefined masks R_b weighted with predicted coefficients c_{i,b}, followed by a softmax function. (c) The output refinement network (ORNet) is trained with this enhanced supervision to improve perceptual quality.