A Watermark-Conditioned Diffusion Model for IP Protection This paper proposes a watermarking framework for content copyright protection within diffusion models, focusing on both practical scenarios and the identification of users responsible for generating outputs. The model provider grants public access to a diffusion model via an API, while users can only query the model API and generate images in a black-box manner. The goal is to embed hidden information into generated contents to facilitate detection and owner identification. The proposed Watermark-conditioned Diffusion model (WaDiff) manipulates the watermark as a conditioned input and incorporates fingerprinting into the generation process. All generative outputs from WaDiff carry user-specific information, which can be recovered by an image extractor and further facilitate forensic identification. Extensive experiments on two popular diffusion models demonstrate that the method is effective and robust in both detection and owner identification tasks. The watermarking framework has a negligible impact on the original generation and is more stealthy and efficient compared to existing strategies. The code is publicly available at https://github.com/rmin2000/WaDiff. The paper introduces a unified watermarking framework for diffusion models, which seamlessly integrates the fingerprinting process into the image generation process. This approach allows for the embedding of user-specific watermarks without customized fine-tuning. The framework is evaluated on two popular open-source diffusion models and shows accurate detection and identification in a large-scale system with numerous users. The generated images maintain exceptional generation quality and are visually indistinguishable across different watermarks. The contributions include a scalable watermarking strategy that efficiently integrates user-specific fingerprints into the diffusion generation process, a watermark-conditioned diffusion model (WaDiff) with a unified watermarking framework, and extensive experiments demonstrating precise and robust performance in detecting AI-generated content and identifying the source owner of generated images.A Watermark-Conditioned Diffusion Model for IP Protection This paper proposes a watermarking framework for content copyright protection within diffusion models, focusing on both practical scenarios and the identification of users responsible for generating outputs. The model provider grants public access to a diffusion model via an API, while users can only query the model API and generate images in a black-box manner. The goal is to embed hidden information into generated contents to facilitate detection and owner identification. The proposed Watermark-conditioned Diffusion model (WaDiff) manipulates the watermark as a conditioned input and incorporates fingerprinting into the generation process. All generative outputs from WaDiff carry user-specific information, which can be recovered by an image extractor and further facilitate forensic identification. Extensive experiments on two popular diffusion models demonstrate that the method is effective and robust in both detection and owner identification tasks. The watermarking framework has a negligible impact on the original generation and is more stealthy and efficient compared to existing strategies. The code is publicly available at https://github.com/rmin2000/WaDiff. The paper introduces a unified watermarking framework for diffusion models, which seamlessly integrates the fingerprinting process into the image generation process. This approach allows for the embedding of user-specific watermarks without customized fine-tuning. The framework is evaluated on two popular open-source diffusion models and shows accurate detection and identification in a large-scale system with numerous users. The generated images maintain exceptional generation quality and are visually indistinguishable across different watermarks. The contributions include a scalable watermarking strategy that efficiently integrates user-specific fingerprints into the diffusion generation process, a watermark-conditioned diffusion model (WaDiff) with a unified watermarking framework, and extensive experiments demonstrating precise and robust performance in detecting AI-generated content and identifying the source owner of generated images.