Distributed Image Compression (DIC) offers substantial improvements in image transmission under bandwidth-constrained conditions, demonstrating its significant promise for image processing tasks in various applications. Recent state-of-the-art (SOTA) methods (e.g., LDMIC, CVCDIC) leverage cross-attention mechanisms to capture correlations between target and reference images. However, current cross-attention maps exhibit sparsity. This issue is further intensified in low-bitrate or complex scenes, where fine-grained feature alignment becomes more challenging. This phenomenon impedes the model’s ability to acquire spatially aligned local auxiliary information for different regions of the target image, consequently restricting the performance of target image reconstruction. To address this problem, we propose an Adaptive Multi-Granularity Joint Context Transfer for Distributed Image Coding (AMDIC) framework, which incorporates independent encoders and two Adaptive Multi-Granularity Joint Context Transfer (AMJCT) module on the decoder side. This module simultaneously extracts global and dual-scale local auxiliary features from reference images, addressing the sparsity issue of cross-attention maps through more comprehensive local context modeling. Furthermore, we introduce learnable weighting networks to enable the adaptive fusion of multi-scale auxiliary information. Extensive experiments conducted on benchmark datasets such as InStereo2K and Cityscapes demonstrate that AMDIC consistently achieves SOTA performance in terms of PSNR and MS-SSIM, significantly outperforming existing DIC methods.