Title: Cross-Model Transferability of Multi-Bit Watermarks in Frozen Diffusion Architectures Using Latent Consistency Bridges

Abstract:

As generative artificial intelligence continues to evolve, international regulatory bodies are increasingly imposing requirements for verifiable content provenance. Yet, a significant chasm remains between these policy demands and current technological capabilities. Traditional watermarking strategies suffer from notable limitations: sampling-based techniques necessitate computationally expensive inversion processes, whereas fine-tuning methods are locked to specific model checkpoints, thereby obstructing standardized oversight across different models.

To resolve this disparity, we present DiffMark, a modular, multi-bit watermarking framework designed for immediate integration. DiffMark functions by embedding a persistent, learned perturbation during each denoising phase of a frozen diffusion model. This process accumulates a signal that can be recovered within the final latent space. To facilitate efficient training through the static network, we employ Latent Consistency Models (LCMs) as a differentiable bridge.

DiffMark demonstrates remarkable efficiency, capable of extracting 64 bits of information in a single forward pass lasting just 16.4 milliseconds. This performance represents a $45\times$ acceleration compared to inversion-based baselines. Furthermore, by supporting per-image key flexibility and cross-architecture transferability without the need for retraining, DiffMark delivers the practical, scalable infrastructure required to implement user accountability and comply with emerging AI governance standards.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC