Digital watermarking, a vital technique for image copyright protection, is now facing serious challenges from object-theft attacks. In this context, object watermarking, which aims to protect target objects with precision, has attracted growing attention. However, with the rapid advancement of deep learning, generative models have proliferated. Attacks based on these models may add or remove local content during the image generation process, thereby distorting embedded watermarks and posing a significant threat to copyright protection. Existing object watermarking methods are mainly designed to resist traditional distortions (e.g., JPEG compression or additive noise) and therefore exhibit poor robustness against images regenerated or locally edited by generative models, resulting in unstable watermark extraction. To address this issue, we propose a novel object watermarking method that embeds watermarks using a generative approach. Specifically, we adopt the Variational Autoencoder (VAE) framework (encoder–decoder) as our watermark-embedding module, embedding the watermark into the module’s deep features (i.e., low-frequency components) to enhance robustness against generative-model attacks. We then encode the watermark information as a dynamically trainable vector, enabling adaptive robustness to different attack types. Experimental results demonstrate that our method outperforms state-of-the-art digital-watermarking techniques across a variety of complex attack scenarios, achieving both excellent watermark-extraction stability and imperceptibility.