MultiNeRF: Multiple Watermark Embedding for Neural Radiance Fields

We present MultiNeRF, a novel 3D watermarking method that enables embedding multiple uniquely keyed watermarks within images rendered by a single Neural Radiance Field (NeRF) model while maintaining high visual quality. Our approach extends TensoRF by incorporating a dedicated watermark grid alongside geometry and appearance grids, ensuring higher watermark capacity without entanglement with scene content. We propose a FiLM-based conditional modulation mechanism that dynamically activates watermarks based on input identifiers, allowing multiple independent watermarks to be embedded and extracted without retraining. Experiments on NeRF-Synthetic and LLFF demonstrate improved robustness and watermark capacity while preserving rendering quality. MultiNeRF generalizes prior single-watermark methods, providing a scalable solution for securing ownership and attribution in 3D content.

MultiNeRF extends the TensoRF representation by adding a dedicated watermark grid alongside the geometry and appearance grids. Each watermark is indexed by a unique integer ID, which is processed through a lightweight embedding network to generate a 16-dimensional conditioning vector. A FiLM-based modulation layer converts this embedding into per-channel scaling and shifting parameters that selectively activate the corresponding watermark features during rendering.

The modulated watermark features are injected into the final layer of the TensoRF colour-decoding MLP, enabling ID-specific watermark patterns to be embedded into rendered views without retraining the model. For robust extraction, full-resolution renders undergo a 2-level Discrete Wavelet Transform (DWT) before being decoded using a HiDDeN-based watermark decoder. Training optimises perceptual, patch-based, and bit-accuracy losses, with optional differentiable noise augmentations improving robustness to transformations and regeneration attacks.