3D Gaussian Splatting (3DGS) has emerged as a cornerstone technique for 3D asset acquisition. However, existing copyright protection methods for 3DGS predominantly rely on embedding watermarks directly into Gaussian primitives, which inevitably degrades rendering quality. To address this issue, this paper proposes a zero-watermarking framework. By directly mapping the inherent features of rendered images to copyright information without modifying Gaussian parameters, the framework achieves perfect visual fidelity. Conventional image zero-watermarking maps features of a single image to a dedicated watermark. In contrast, our method guarantees mapping consistency: features of rendered images from any unknown viewpoint can be mapped to the same copyright identifier. To address this cross-view consistency challenge, we introduce an uncertainty-guided strategy that scores individual pixels to guide the decoder to mine shared features across multiple perspectives. This strategy enables accurate watermark retrieval even from novel viewpoints. Extensive experiments on the Blender, LLFF, and MipNeRF-360 datasets demonstrate that our method achieves superior performance, characterized by high message capacity, strong adversarial robustness, and a low false positive rate (FPR), while fully maintaining the integrity of the original 3DGS model.
Zhu et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: