Subject-driven image generation aims to synthesize images that preserve the identity of a specific subject while allowing diverse variations guided by text descriptions. Recent methods have made notable progress in generating realistic and semantically faithful images, and further improvements can be made by better disentangling subject identity from background cues, particularly in low-reference settings. To enhance representation quality, we propose Disentangled Subject Customization via Information Bottleneck (DiscoIB), a novel framework that introduces the information bottleneck (IB) theory to promote cleaner subject representations. Specifically, DiscoIB includes a Disentangled Information Bottleneck Encoder (DIBE), which extracts background-specific features from CLIP image embeddings and injects them into the text embedding stream. DIBE is trained to minimize mutual information with the original image while maximizing it with background-only content, thereby encouraging effective subject-background separation. Additionally, we introduce a cross-attention mechanism specialized for background supervision. Experiments demonstrate that DiscoIB achieves superior identity preservation, background control, and generation editability compared to prior methods, offering a simple yet effective solution for personalized image generation with enhanced disentanglement. Our code is available at https://github.com/ycfang-lab/DiscoIB .