Diffusion models generate highly realistic images from text, enabling wide applications. However, they often struggle with ambiguous object positioning for multiple objects in a prompt. While recent methods like Complementary Regional Diffusion define positions via segmentation, constraining diffusion regions may reduce object recognizability and fail to solve the positioning problem, primarily due to insufficient pixel-level control over object regions during noise prediction. To address these issues, we propose Spatial Positioning via Attention-guided Diffusion without Extra Training (SPADE). First, we introduce Split Noise Prediction Method, which partitions prediction regions for individual objects prior to noise prediction, significantly enhancing positional accuracy and object recognizability. Second, we propose Split Attention Prediction Method, integrating Region-Specific Weighted and Split Parameters attention mechanisms. This enhances positional accuracy while countering overall image integrity degradation introduced by the Split Noise Prediction Method. Finally, quantitative evaluation on T2I-CompBench shows SPADE surpasses Complementary Regional Diffusion by 27.6% in spatial relationship. Qualitative comparisons further confirm superior prompt alignment in generated images. Consequently, SPADE significantly enhances diffusion models’ practicality in interior layout design. Our code is available at https://github.com/yuanming0415/SPADE .