Data transformation pipelines are critical for modern machine-learning workflows, yet existing watermarking frameworks lack adaptability to hybrid attacks and global context modeling. We introduce a hybrid CNN–ViT architecture for blind image watermarking, implemented in PyTorch with a dynamically reconfigurable noise pipeline. Our encoder fuses local convolutional feature extraction with Vision Transformer self-attention to distribute watermark bits redundantly across spatial and semantic tokens. Extensive experiments on COCO, ImageNet, and CelebA show that, compared to the pure-CNN HiDDeN [5] baseline, our model reduces decoder MSE by up to 49 % under combined distortions and cuts bitwise error rate by as much as 68 %. An ablation study confirms that replacing the ViT block with additional convolutional layers doubles decoder MSE and increases bitwise error by two orders of magnitude. Without adversarial training, we achieve competitive encoder MSE (< 0.0083) and maintain imperceptibility under all noise conditions. Our open-source framework thus delivers a new state-of-the-art trade-off between robustness, imperceptibility, and efficiency, suitable for both cloud and edge deployments.