**GaussianStego: A Generalizable Stenography Pipeline for Generative 3D Gaussians Splatting** Recent advancements in large generative models and real-time neural rendering have led to the widespread distribution of visual data through synthesized 3D assets. However, methods for embedding proprietary or copyright information in these assets, such as Gaussian Splatting, remain underexplored. This paper introduces *GaussianStego*, a method for embedding steganographic information in the rendering of generated 3D assets. The approach uses an optimization framework to accurately extract hidden information from images rendered using Gaussian assets derived from large models, while maintaining their original visual quality. The paper discusses the challenges and opportunities in applying steganography to 3D content, particularly in the context of Gaussian Splatting. It highlights the need for a method that can preserve the visual quality of rendered content while embedding customizable, imperceptible, and recoverable information. *GaussianStego* addresses these challenges by integrating hidden information into the intermediate features of a 3D Gaussian generation baseline via cross-attention. A U-Net-based decoder is used to retrieve the concealed information from images rendered at specific viewpoints. The paper also introduces an adaptive gradient harmonization strategy to balance the rendering quality and information hiding objectives. Extensive experiments demonstrate the model's superiority in novel view synthesis rendering quality and high-fidelity transmission of hidden information. The method is evaluated across various deployment scenarios, including embedding image watermarks and multimodal watermarks such as text, QR codes, and videos. The paper concludes by highlighting the significance of *GaussianStego* in addressing the emerging challenge of seamlessly integrating customizable, almost imperceptible, and recoverable information into generated 3D Gaussians, providing valuable insights for future research.**GaussianStego: A Generalizable Stenography Pipeline for Generative 3D Gaussians Splatting** Recent advancements in large generative models and real-time neural rendering have led to the widespread distribution of visual data through synthesized 3D assets. However, methods for embedding proprietary or copyright information in these assets, such as Gaussian Splatting, remain underexplored. This paper introduces *GaussianStego*, a method for embedding steganographic information in the rendering of generated 3D assets. The approach uses an optimization framework to accurately extract hidden information from images rendered using Gaussian assets derived from large models, while maintaining their original visual quality. The paper discusses the challenges and opportunities in applying steganography to 3D content, particularly in the context of Gaussian Splatting. It highlights the need for a method that can preserve the visual quality of rendered content while embedding customizable, imperceptible, and recoverable information. *GaussianStego* addresses these challenges by integrating hidden information into the intermediate features of a 3D Gaussian generation baseline via cross-attention. A U-Net-based decoder is used to retrieve the concealed information from images rendered at specific viewpoints. The paper also introduces an adaptive gradient harmonization strategy to balance the rendering quality and information hiding objectives. Extensive experiments demonstrate the model's superiority in novel view synthesis rendering quality and high-fidelity transmission of hidden information. The method is evaluated across various deployment scenarios, including embedding image watermarks and multimodal watermarks such as text, QR codes, and videos. The paper concludes by highlighting the significance of *GaussianStego* in addressing the emerging challenge of seamlessly integrating customizable, almost imperceptible, and recoverable information into generated 3D Gaussians, providing valuable insights for future research.