FlagThis

Abstract

Low-Rank Adaptation (LoRA) has been widely adopted for customizing text-to-image diffusion models, enabling low-cost adaptation through lightweight and portable modules that are increasingly shared, reused, and commercialized on platforms such as Civitai and HuggingFace. This emerging LoRA assets ecosystem shifts the practical focus of copyright protection from large-scale foundation models to independently distributed LoRA modules, making LoRA copyright protection practical and urgent. However, existing methods either protect the foundation diffusion model, leaving LoRA assets vulnerable to unauthorized copying and redistribution, or depend on watermark retraining tailored to each specific LoRA, making them impractical for open community settings where diverse creators continuously publish new LoRA assets. To address this limitation, we propose LoRAKey, a user-centric LoRA watermarking method that treats copyright protection as a reusable ownership key instead of retraining a new watermark for each LoRA asset. LoRA-Key protects customized LoRA assets by attaching a user-specific Watermark LoRA, which carries a recoverable secret message for identifying the creator. Once trained, the Watermark LoRA can be combined with different target LoRAs through training-free linear superposition, without per-LoRA retraining or structural modification. To train such a reusable key, we first establish a latent watermark prior in the frozen VAE latent space for robust message embedding and recovery, and then optimize the Watermark LoRA with message-conditioned watermark supervision and semantic consistency constraints. We further introduce Gradient Orthogonal Projection (GOP) to suppress watermark updates that conflict with semantic-preserving directions, thereby reducing interference with generation fidelity and downstream style adaptation. Extensive experiments demonstrate that LoRAKey provides lightweight plug-and-play copyright protection while preserving generation quality and style fidelity, and maintains robust ownership verification under image-level distortions, downstream fine-tuning, and multi-LoRA composition.