ProMark is a proactive diffusion watermarking technique for causal attribution, enabling the identification of training data concepts that influence the generation of synthetic images. The method embeds imperceptible watermarks into training images, which are then retained by diffusion models during image generation. ProMark can embed up to 2^16 unique watermarks, each corresponding to a training concept such as objects, scenes, or styles. The technique outperforms correlation-based attribution methods by providing a causal link between training data and generated images. ProMark is trained on both conditional and unconditional diffusion models and can be used for fine-tuning pre-trained models with minimal iterations. The method maintains image quality while achieving higher accuracy in concept attribution across diverse datasets, including Adobe Stock, ImageNet, LSUN, WikiArt, and BAM. ProMark's approach involves embedding multiple orthogonal watermarks into a single image, allowing for multi-concept attribution. The technique is robust to various image degradations and demonstrates high accuracy in identifying the correct training concept for generated images. ProMark's effectiveness is validated through extensive experiments, showing its potential for real-world applications in attributing creative contributions to generative AI. The method is efficient, with low computational cost during inference and training, and offers a balance between causality and image quality.ProMark is a proactive diffusion watermarking technique for causal attribution, enabling the identification of training data concepts that influence the generation of synthetic images. The method embeds imperceptible watermarks into training images, which are then retained by diffusion models during image generation. ProMark can embed up to 2^16 unique watermarks, each corresponding to a training concept such as objects, scenes, or styles. The technique outperforms correlation-based attribution methods by providing a causal link between training data and generated images. ProMark is trained on both conditional and unconditional diffusion models and can be used for fine-tuning pre-trained models with minimal iterations. The method maintains image quality while achieving higher accuracy in concept attribution across diverse datasets, including Adobe Stock, ImageNet, LSUN, WikiArt, and BAM. ProMark's approach involves embedding multiple orthogonal watermarks into a single image, allowing for multi-concept attribution. The technique is robust to various image degradations and demonstrates high accuracy in identifying the correct training concept for generated images. ProMark's effectiveness is validated through extensive experiments, showing its potential for real-world applications in attributing creative contributions to generative AI. The method is efficient, with low computational cost during inference and training, and offers a balance between causality and image quality.