Molecular generation constitutes a pivotal technology in computational chemistry with profound implications for drug discovery and materials de-sign. While recent advances in generative models have accelerated molecular exploration, existing approaches exhibit limitations in synthesizing molecules under precise property constraints while preserving structural integrity. In this paper, we propose the Conditional Geometric Latent Diffusion Model (CGLDM), a novel framework that improves the guidance of conditional molecular properties while maintaining the stability of molecular structure. Our approach comprises two innovations: (1) A pretraining paradigm utilizing Conditional Variational Autoencoder (CVAE) to enforce latent space alignment with molecular conditions; (2) A denoising architecture integrating classifier-free guidance with geometric-aware diffusion steps, enabling precise control over both chemical properties and 3D conformations. Experimental results demonstrate CGLDM’s superior performance in both single- and multi-condition generation scenarios. CVAE pretraining paradigm substantially improves conditional distribution matching, while the proposed guidance mechanism effectively reduces property prediction errors and achieves high molecular structural stability, effectiveness and novelty.