<p>Generative image steganography has become a research hotspot due to its advantages in quality and anti-detection capabilities. Object-based semantic embedding methods help improve the security of the steganographic process by mapping secret information to image semantic categories and then using a generator network to synthesize stego images. However, most existing methods tend to adopt a single-object embedding approach, which usually allows only a limited number of bits to be embedded in a single image; moreover, when the scale of the semantic category library is insufficient, their embedding capacity often decreases quite significantly. This paper proposes a generative steganography method based on multi-object semantic joint encoding. First, the secret information is segmented and mapped to multiple semantic categories, which are then combined to generate composite textual prompts. Subsequently, based on a latent diffusion model, these prompts guide the generation of stego images, allowing a single image to carry multiple semantic information categories and effectively multiplying the embedding capacity. The receiver extracts the multi-object semantic information using a classifier to complete the decryption process. Furthermore, we design a joint training strategy for the generator and classifier to enhance the accuracy of secret information extraction. We train the model on a self-constructed dataset, and experimental results demonstrate the feasibility of this approach. Additionally, we analyze its security, embedding capacity, and robustness.</p>