Knowledge Distillation for Learned Image Compression

Figure 1. State-of-the-Art Performance of KDIC. We propose a knowledge distillation framework for learned image compression and develop KDIC, a compact and efficient student model based on S2CFormer [12]. As shown in Figure (a), KDIC outperforms recent advanced methods in rate-distortion (RD) metrics and model complexity measures. Figure (b) highlights the effectiveness of knowledge distillation: it achieves a 2.5% reduction in BDrate while reducing parameters by 40% and FLOPs by 57% compared to the teacher model. The circle radius corresponds to the parameter count of each model.