To address the challenges of limited data availability, high annotation costs, and privacy issues in medical image analysis, generative models are a potential solution for medical imaging. However, existing methods often produce uncontrollable results, particularly in tumor generation, resulting in inaccurate shape or low-fidelity images. In this paper, we propose a dual-guided conditional control by text prompts and anatomical masks for generating 3D liver CT images. Our framework introduces a dual-guidance approach—integrating text prompts and anatomical masks—within a three-stage architecture. First, a variational autoencoder compresses high-resolution CT images into a latent space to reduce computational demands. Next, a latent diffusion model utilizes semantic information from radiology report features. Finally, a ControlNet module incorporates segmentation masks to precisely control tumor location. Experimental results demonstrate that the generated abdominal CT images exhibit superior visual realism and anatomical consistency compared to existing models. Moreover, the synthesized data significantly enhances downstream task performance.