In recent times, there has been a significant surge in the need to create, distribute, and preserve vast quantities of digital images captured by smart devices and sensors. This gives rise to several problems, such as unauthorized access and deceptive utilization of these images, as well as other security considerations. Nevertheless, the inherent digital aspect of these images renders them readily accessible, modifiable, or susceptible to misuse. To mitigate these issues, watermarking solutions are introduced, basically the process of embedding a unique design or identifier into a digital cover and then removing it later. This is done to address conflicts related to ownership and copyright violations that may arise with media data. The exponential development of deep neural networks has facilitated the rapid growth of research on deep learning-based watermarking techniques. Deep-learning methods are highly advantageous in watermarking because of their precision, exceptional outcomes, and robust learning capability. This study presents a unified architecture for watermarking based on deep learning to achieve high invisibility and resilience. The cover image is subsampled during the preparation stage. To accomplish watermark embedding, the single encoder utilizes normalizing flow, which essentially merges the secret image with the subsampled cover image. A parallel decoder has been implemented to enhance watermark extraction’s imperceptibility and capability. The suggested architecture has been proven to achieve superior watermark robustness and imperceptibility after extensive experimentation. Our approach attains a pixel error rate below 0.1% in the face of several attacks, including Speckle noise, Sharpen noise, Gaussian noise, Alphadropout, and Blur attack. In addition, the suggested approach also demonstrates significant robustness as compared to existing works.